ProbeShift reproducibility bundle: code + results + paper + figures

.gitattributes

@@ -58,3 +58,5 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 # Video files - compressed
 *.mp4 filter=lfs diff=lfs merge=lfs -text
 *.webm filter=lfs diff=lfs merge=lfs -text
+repro_bundle/probeshift_honest.pdf filter=lfs diff=lfs merge=lfs -text
+repro_bundle/probeshift_main.pdf filter=lfs diff=lfs merge=lfs -text

repro_bundle/CODE_SPEC.md

@@ -0,0 +1,428 @@
+这是一个工程架构设计任务,不是 ADR 决策。我直接输出完整的实验代码架构 spec。下面是为《Probe Stability Score》设计的、可在单卡 4090 上运行的完整工程化 spec。
+
+---
+
+# Probe Stability Score · 实验代码架构 Spec
+
+设计目标:一次性激活抽取 + 缓存复用,所有探针/指标计算只读缓存;严控显存与 GPU·h(≤30 GPU·h);单条命令复现最小可行实验(MVE)。
+
+---
+
+## 1. 目录 / 文件树
+
+```
+probe_stability/
+├── README.md
+├── pyproject.toml                 # 依赖 + 入口 console_scripts
+├── Makefile                       # make mve / make full / make figures
+├── configs/
+│   ├── base.yaml                  # 全局:路径、dtype、seed 列表、CI 参数
+│   ├── models.yaml                # 7 个模型的 HF id、层数、d_model、最佳 batch
+│   ├── datasets.yaml              # 8 个任务:HF id、字段映射、label 空间、采样数
+│   ├── shifts.yaml               # back-translation / domain / length 三类 shift 配置
+│   ├── probes.yaml                # logreg / mass_mean / mlp / control 超参
+│   ├── metrics.yaml               # Score 双分量权重、bootstrap、多重检验方法
+│   └── experiments/
+│       ├── mve.yaml               # 最小可行:pythia-70m + SST-2 + 1 shift + 2 seed
+│       └── full.yaml              # 全量矩阵
+├── env/
+│   └── download_models.py         # 预下载所有 HF 权重到本地缓存(离线复现)
+├── src/probe_stability/
+│   ├── __init__.py
+│   ├── cli.py                     # Typer/argparse 主入口,子命令路由
+│   ├── config.py                  # OmegaConf 加载 + dataclass schema 校验
+│   ├── registry.py                # MODEL/DATASET/PROBE/SHIFT 注册表(字符串→构造器)
+│   ├── utils/
+│   │   ├── seeding.py             # set_all_seeds(seed)
+│   │   ├── logging.py             # rich/structlog,JSONL run log
+│   │   ├── io.py                  # shard 路径生成、原子写、manifest 读写
+│   │   ├── gpu.py                 # 显存监控、autocast、空缓存、OOM 自动降 batch
+│   │   └── hashing.py             # config/input 指纹 → 缓存 key、幂等跳过
+│   ├── data/
+│   │   ├── loaders.py             # 8 个数据集统一加载为 Example schema
+│   │   ├── schema.py              # Example / DistributionSet dataclass
+│   │   ├── sampling.py            # 分层下采样、固定 split、length buckets 切分
+│   │   └── shifts/
+│   │       ├── base.py            # Shift 抽象基类
+│   │       ├── backtranslation.py # NLLB / opus-mt 本地往返翻译
+│   │       ├── domain.py          # Amazon multi-domain / IMDB↔SST-2 域迁移
+│   │       └── length.py          # 短/中/长 token 分桶
+│   ├── fidelity/
+│   │   └── mnli_filter.py         # 本地 DeBERTa-v3-MNLI 过滤 label flip
+│   ├── extract/
+│   │   ├── activations.py         # 核心:逐层 residual-stream 抽取 + 流式落盘
+│   │   ├── hooks.py               # 各架构 residual hook 适配(Pythia/GPT2/Qwen2)
+│   │   └── pooling.py             # last-token / mean pooling
+│   ├── cache/
+│   │   ├── store.py               # ShardStore:read/write float16 分片 + manifest
+│   │   └── manifest.py            # 全局缓存索引(parquet),完整性校验
+│   ├── probes/
+│   │   ├── base.py                # Probe 抽象:fit/predict/direction
+│   │   ├── logreg.py              # sklearn / torch LogReg
+│   │   ├── mass_mean.py           # difference-of-means 方向探针
+│   │   ├── mlp.py                 # 2-layer MLP (torch)
+│   │   └── control.py             # random-label control-task 包装器
+│   ├── eval/
+│   │   ├── train_eval.py          # 读缓存→训练→IID/OOD 评估 单元
+│   │   ├── directions.py          # 概念方向、跨分布 cosine 旋转
+│   │   └── selectivity.py         # selectivity = acc(real) - acc(control)
+│   ├── score/
+│   │   └── stability.py           # 双分量 Probe Stability Score 计算
+│   ├── stats/
+│   │   ├── bootstrap.py           # bootstrap CI
+│   │   ├── correlation.py         # Spearman(Score ↔ OOD drop)
+│   │   ├── effect_size.py         # Cliff's delta
+│   │   └── multiple_testing.py    # Holm / Benjamini-Hochberg
+│   └── pipeline/
+│       ├── stage_extract.py       # 阶段1:激活抽取(写缓存)
+│       ├── stage_probe.py         # 阶段2:探针训练+评估(读缓存)
+│       ├── stage_score.py         # 阶段3:Score 聚合
+│       └── stage_stats.py         # 阶段4:统计检验 + 表格
+├── scripts/
+│   ├── run_mve.sh                 # 一条命令复现 MVE
+│   ├── run_full.sh
+│   └── make_figures.py            # 论文图(cosine 旋转、Score↔drop 散点)
+├── tests/
+│   ├── test_cache_roundtrip.py
+│   ├── test_shapes.py
+│   ├── test_probes_smoke.py
+│   └── test_score_math.py
+└── artifacts/                     # 运行产物(git-ignored)
+    ├── cache/                     # 激活分片
+    ├── results/                   # parquet 结果表
+    ├── figures/
+    └── logs/
+```
+
+---
+
+## 2. 各模块职责 + 关键函数签名
+
+### 2.1 config / registry
+
+```python
+# config.py
+@dataclass
+class RunConfig:
+    models: list[str]; datasets: list[str]; shifts: list[str]
+    probes: list[str]; seeds: list[int]
+    cache_dir: Path; results_dir: Path
+    pooling: str = "last_token"        # last_token | mean
+    dtype: str = "float16"
+    max_examples_per_dist: int = 4000   # 控制规模/显存
+    extract_batch: dict[str, int] = ...  # 每模型 batch 覆盖
+
+def load_config(exp_path: str, overrides: list[str]) -> RunConfig: ...
+
+# registry.py
+def register(kind: str, name: str): ...            # 装饰器
+def build_model(name: str) -> "LMWrapper": ...
+def build_dataset(name: str) -> "DistributionSet": ...
+def build_probe(name: str, **kw) -> "Probe": ...
+def build_shift(name: str) -> "Shift": ...
+```
+
+### 2.2 data
+
+```python
+# schema.py
+@dataclass(frozen=True)
+class Example:
+    uid: str; text: str; label: int; meta: dict   # meta 含 domain/length_bucket
+
+@dataclass
+class DistributionSet:
+    name: str                 # e.g. "sst2"
+    distribution: str         # "iid" | "bt_de" | "domain_books" | "len_long"
+    examples: list[Example]
+    num_labels: int
+
+# loaders.py
+def load_dataset(name: str, split: str, max_n: int, seed: int) -> DistributionSet: ...
+# 支持: sst2 imdb amazon_multi ag_news dbpedia ud_pos truthfulqa_bin counterfact_bin
+
+# sampling.py
+def stratified_subsample(ds: DistributionSet, n: int, seed: int) -> DistributionSet: ...
+def length_buckets(ds: DistributionSet, tokenizer, edges=(32,128)) -> dict[str, DistributionSet]: ...
+
+# shifts/base.py
+class Shift(ABC):
+    name: str
+    @abstractmethod
+    def apply(self, ds: DistributionSet) -> DistributionSet: ...
+    # 产出 distribution 字段已改写、label 暂保留(待 fidelity 过滤)
+
+# shifts/backtranslation.py
+class BackTranslationShift(Shift):
+    def __init__(self, pivot="deu_Latn", model="facebook/nllb-200-distilled-600M",
+                 batch_size=32, device="cuda"): ...
+    def apply(self, ds) -> DistributionSet: ...     # text→pivot→en,流式 batch
+```
+
+### 2.3 fidelity
+
+```python
+# mnli_filter.py
+class MNLIFidelityFilter:
+    def __init__(self, model="MoritzLaurer/DeBERTa-v3-base-mnli", thresh=0.5): ...
+    def keep_mask(self, src_texts: list[str], shifted_texts: list[str]) -> np.ndarray:
+        """双向蕴含判定:保留语义未翻转(label-preserving)样本,返回 bool mask"""
+    def filter(self, src: DistributionSet, shifted: DistributionSet) -> DistributionSet: ...
+```
+
+### 2.4 extract(核心)
+
+```python
+# hooks.py
+def residual_hook_points(model, arch: str) -> list[tuple[int, nn.Module]]:
+    """返回 [(layer_idx, module)],module 输出即该层 residual-stream。
+    Pythia(GPTNeoX): gpt_neox.layers[i]; GPT2: transformer.h[i]; Qwen2: model.layers[i]"""
+
+# pooling.py
+def pool(hidden: Tensor, attn_mask: Tensor, mode: str) -> Tensor:
+    """(B,T,D)->(B,D);last_token 取最后非 pad 位,mean 做 mask 平均"""
+
+# activations.py
+class ActivationExtractor:
+    def __init__(self, model_name: str, dtype="float16", device="cuda",
+                 pooling="last_token", max_len=256): ...
+
+    @torch.inference_mode()
+    def extract(self, ds: DistributionSet, store: "ShardStore",
+                batch_size: int, layers: list[int] | None = None) -> ShardMeta:
+        """单次前向抓全层 residual,逐 batch pool→cast fp16→流式 append 落盘。
+        返回写入分片的元信息(n, layers, d_model)。"""
+```
+
+抽取核心逻辑:注册 forward hook 抓全部层 → 一次前向 → 每层 pool → fp16 → 立即追加写盘,**不在显存/内存累积全量激活**。
+
+### 2.5 cache
+
+```python
+# store.py  分片单位 = (model, dataset, distribution);层维进 .npz 同文件多 key 或单层一文件
+class ShardStore:
+    def __init__(self, root: Path, dtype="float16"): ...
+    def shard_path(self, model, dataset, distribution, layer, seed_split="all") -> Path: ...
+    def open_writer(self, model, dataset, distribution, n_layers, d_model, n: int): ...
+    def append(self, layer_arrays: dict[int, np.ndarray], labels, uids): ...  # 流式
+    def finalize(self): ...                               # 写 X.npy / y.npy / uids
+    def load(self, model, dataset, distribution, layer) -> tuple[np.ndarray, np.ndarray, list[str]]:
+        """mmap 读取 X(fp16), y(int8), uids"""
+    def exists(self, model, dataset, distribution, layer) -> bool: ...  # 幂等跳过
+
+# manifest.py
+def update_manifest(root: Path, meta: dict): ...         # 追加到 manifest.parquet
+def verify_cache(root: Path) -> list[str]: ...           # 返回缺失/损坏 shard
+```
+
+### 2.6 probes
+
+```python
+# base.py
+class Probe(ABC):
+    @abstractmethod
+    def fit(self, X: np.ndarray, y: np.ndarray) -> "Probe": ...
+    @abstractmethod
+    def predict(self, X: np.ndarray) -> np.ndarray: ...
+    @abstractmethod
+    def direction(self) -> np.ndarray | None:   # 概念方向(单位向量),MLP 返回 None 或一阶近似
+        ...
+
+# logreg.py  LogRegProbe(C=1.0, max_iter=1000)  —— 内部 X 升精到 float32 训练
+# mass_mean.py MassMeanProbe —— direction = (μ_pos - μ_neg)/||·||;阈值=中点投影
+# mlp.py  MLPProbe(hidden=256, epochs=30, lr=1e-3, device="cuda")  —— direction=输入层雅可比近似
+# control.py
+class ControlTaskProbe(Probe):
+    def __init__(self, inner: Probe, seed: int):
+        """random-label control:固定随机标签置换后训练同结构探针"""
+```
+
+### 2.7 eval / score / stats
+
+```python
+# train_eval.py
+def fit_and_eval(probe_name, store, model, dataset,
+                 train_dist="iid", eval_dists=("iid","bt_de",...),
+                 seed=0) -> ProbeResult:
+    """读 train_dist 缓存→fit→在各 eval_dist 上算 acc;返回 acc 字典 + 方向向量"""
+
+# directions.py
+def cosine_rotation(dir_iid: np.ndarray, dir_ood: np.ndarray) -> float: ...   # 1 - cos
+def aggregate_directions(results: list[ProbeResult], by="layer") -> pd.DataFrame: ...
+
+# selectivity.py
+def selectivity(real_acc: float, control_acc: float) -> float: ...
+
+# score/stability.py
+def probe_stability_score(acc_iid, acc_ood, dir_iid, dir_ood,
+                          w_perf=0.5, w_geom=0.5) -> dict:
+    """双分量:
+       perf_component = acc_ood / acc_iid              (性能保持度,∈[0,1])
+       geom_component = cos(dir_iid, dir_ood)          (方向稳定度,∈[-1,1]→clip[0,1])
+       score = w_perf*perf + w_geom*geom
+       返回 {score, perf_component, geom_component}"""
+
+# stats/
+def bootstrap_ci(values, fn=np.mean, n=1000, alpha=0.05, seed=0) -> tuple[float,float,float]: ...
+def spearman_with_ci(score, ood_drop, n_boot=1000) -> dict: ...   # rho + CI + p
+def cliffs_delta(a, b) -> float: ...
+def holm_bonferroni(pvals) -> np.ndarray: ...
+def benjamini_hochberg(pvals, q=0.05) -> np.ndarray: ...
+```
+
+### 2.8 pipeline
+
+```python
+# stage_extract.py
+def run_extract(cfg: RunConfig):
+    for model in cfg.models:
+        ext = ActivationExtractor(model, ...)
+        for ds_name in cfg.datasets:
+            base = load_dataset(ds_name, "test", cfg.max_examples_per_dist, seed=0)
+            dists = [base] + [filter_fidelity(s.apply(base), base) for s in shifts]
+            for d in dists:
+                if store.exists(model, ds_name, d.distribution, layer=0): continue  # 幂等
+                ext.extract(d, store, batch_size=cfg.extract_batch[model])
+        del ext; free_gpu()        # 关键:释放当前模型再载下一个
+
+# stage_probe.py / stage_score.py / stage_stats.py 全程只读缓存,CPU/小 GPU 即可
+```
+
+---
+
+## 3. 数据流与缓存格式
+
+### 3.1 数据流
+
+```
+原始数据集 ──load_dataset──► DistributionSet(iid)
+                                  │
+                    shift.apply ──┼──► bt_de / domain_x / len_*  (候选)
+                                  │
+              MNLI fidelity 过滤 ─┘──► label-preserving DistributionSet
+                                  │
+   ActivationExtractor(单模型常驻)│ 一次前向抓全层 → pool → fp16 → 流式
+                                  ▼
+                          ShardStore 落盘  (阶段1结束)
+                                  │  (此后 GPU 几乎闲置)
+       fit_and_eval(读缓存) ──────┼──► ProbeResult{acc[dist], direction}
+                                  ▼
+            stability.py ──► Score(双分量)
+                                  ▼
+            stats ──► Spearman/CI/Holm-BH/Cliff's δ ──► results/*.parquet ──► figures
+```
+
+### 3.2 缓存目录与 shard 命名
+
+分片粒度:`(model, dataset, distribution)` 一组目录,层维拆文件,便于按层流式与按需读取。
+
+```
+artifacts/cache/
+└── {model}/{dataset}/{distribution}/
+        ├── X_layer{LL}.npy        # 每层一个,float16,shape (N, d_model)
+        ├── y.npy                  # int8,  shape (N,)
+        ├── uids.npy               # <U32 字符串 id, shape (N,)
+        └── meta.json              # {n, n_layers, d_model, pooling, dtype, src_hash, created}
+```
+
+命名约定:
+
+| 占位符 | 取值示例 | 说明 |
+|---|---|---|
+| `{model}` | `pythia-70m`, `gpt2-medium`, `qwen2.5-0.5b` | HF id 安全化 |
+| `{dataset}` | `sst2`, `ag_news`, `ud_pos` | configs/datasets.yaml key |
+| `{distribution}` | `iid`, `bt_de`, `domain_books`, `len_long` | iid + shift 产物 |
+| `{LL}` | `00`..`24` | 层索引零填充两位(含 embedding 层为 0) |
+
+### 3.3 dtype / shape 约定
+
+| 数组 | dtype | shape | 备注 |
+|---|---|---|---|
+| `X_layer{LL}` | float16 | (N, d_model) | residual-stream pooled;训练时升 float32 |
+| `y` | int8 | (N,) | 标签;多分类 ≤127 类够用 |
+| `uids` | <U32 | (N,) | 跨分布对齐(fidelity 后 N 可不同) |
+| `manifest.parquet` | — | 一行一 shard | 列:model,dataset,distribution,layer,n,d_model,bytes,sha1 |
+
+约定:同一 `(model,dataset)` 下所有 distribution 的 `d_model`、`n_layers` 一致;`X` 列序 = `residual_hook_points` 返回顺序。
+
+---
+
+## 4. 显存与运行时控制策略
+
+1. **单模型常驻、串行换模型**:外层循环按 model,处理完所有 dataset/shift 后 `del model; gc.collect(); torch.cuda.empty_cache()` 再载下一个。任一时刻显存只住一个 LM。1.4B 在 fp16 约 2.8 GB 权重,远低于 24 GB。
+
+2. **激活抽取一律 `torch.inference_mode()` + fp16 autocast**,不建图、不存梯度。一次前向用 hook 抓全层,避免逐层多次前向。
+
+3. **流式落盘,绝不全量驻留**:每个 batch 抽取→pool 成 (B,D)→cast fp16→`np.memmap` 追加写,显存/内存只持有一个 batch 的激活。pool 后丢弃 (B,T,D) 中间张量。
+
+4. **batch 自适应降级**:`gpu.py` 包装 OOM 捕获,触发即 `batch//=2` 重试,直到成功;最优 batch 记入 `models.yaml` 复用。`max_len=256` 截断,长样本不爆显存。
+
+5. **辅助模型(NLLB / DeBERTa-MNLI)单独阶段跑**,与 LM 抽取错峰,跑完即释放;同样 fp16 + batch 流式。
+
+6. **探针阶段几乎不吃显存**:LogReg/mass-mean 走 CPU(numpy/sklearn);MLP 用小 GPU batch(X 已是 (N,D) 向量,512 batch 占用极小)。可与抽取分进程。
+
+7. **mmap 读缓存**:`np.load(..., mmap_mode='r')`,大矩阵不一次性进内存;训练时按需切片升 float32。
+
+8. **幂等跳过**:`store.exists()` 命中即跳过,中断后续跑零浪费;`hashing.py` 用 config+输入指纹做 key。
+
+---
+
+## 5. 4090 上各阶段运行时估算(目标 ≤30 GPU·h)
+
+规模假设:8 数据集 × (1 IID + 3 shift 各取代表) ≈ 每数据集 ~4 distribution;每 distribution 上限 4000 样本;7 模型;`max_len=256`。
+
+| 阶段 | 主要开销 | GPU·h 估算 |
+|---|---|---|
+| **辅助:back-translation (NLLB-600M)** | 8 数据集 × ~4000 × fp16,~往返 2 次 | ~4.0 |
+| **辅助:MNLI fidelity (DeBERTa-base)** | 仅对 shift 样本,分类一次/样本 | ~1.5 |
+| **阶段1:激活抽取(7 模型)** | 小模型(70M/160M/410M/gpt2-s/m/qwen0.5b)合计 ~3.5;1.4B 单独 ~3.0;总样本 ~8×4×4000=128k 前向/模型 | ~8.0 |
+| **阶段2:探针训练+评估** | LogReg/mass-mean 走 CPU(忽略 GPU);MLP 全配置 GPU 小 batch | ~3.0 |
+| **阶段3:Score 聚合** | 纯 CPU | ~0 |
+| **阶段4:统计(bootstrap/检验)** | 纯 CPU(可多进程) | ~0 |
+| **缓冲(重跑/调参/失败重试)** | — | ~5.0 |
+| **合计** | | **≈ 24.5 GPU·h ✓ ≤30** |
+
+降本旋钮:`max_examples_per_dist` 调到 2000 可再省 ~40% 抽取与翻译时间;shift 数从 3 减到 2;1.4B 仅在子集数据集上跑。
+
+---
+
+## 6. 复现入口(一条命令跑通 MVE)
+
+`configs/experiments/mve.yaml`:`pythia-70m` × `sst2` × (`iid` + `bt_de` 一个 shift) × `{logreg, mass_mean, control}` × `seeds=[0,1]`,`max_examples_per_dist=500`。预计 **< 10 分钟、< 4 GB 显存**。
+
+`scripts/run_mve.sh`:
+
+```bash
+#!/usr/bin/env bash
+set -euo pipefail
+python -m probe_stability.cli run \
+  --exp configs/experiments/mve.yaml \
+  --stages extract,probe,score,stats \
+  --out artifacts/
+```
+
+一条命令:
+
+```bash
+bash scripts/run_mve.sh
+# 等价于
+python -m probe_stability.cli run --exp configs/experiments/mve.yaml --stages all
+```
+
+产物:`artifacts/cache/pythia-70m/sst2/{iid,bt_de}/X_layer*.npy`、`artifacts/results/{probe_results,scores,stats}.parquet`、`artifacts/figures/score_vs_drop.png`。
+
+全量:`python -m probe_stability.cli run --exp configs/experiments/full.yaml --stages all`。
+
+CLI 子命令(便于分阶段重跑、节省 GPU·h):
+
+```bash
+probe-stability extract --exp ...     # 仅写缓存(吃 GPU)
+probe-stability probe   --exp ...     # 仅读缓存训探针(几乎不吃 GPU)
+probe-stability score   --exp ...
+probe-stability stats   --exp ...
+probe-stability verify-cache          # 校验分片完整性
+```
+
+---
+
+关键工程取舍:**抽取与训练彻底解耦**(缓存为界)是 GPU·h 预算可控的核心——GPU 只在阶段1+辅助阶段被占用,探针的多 seed × 多探针 × 多分布组合爆炸全部落在读缓存的 CPU 侧;**单模型常驻串行**保证 1.4B 也安全;**流式 fp16 落盘 + 幂等跳过**让中断可续、零重复计算。如果需要,我可以接着把任意单个模块(如 `activations.py` 的 hook 抽取或 `stability.py` 的 Score 计算)落成可运行代码。

repro_bundle/EXPERIMENTS_PLAN.md

@@ -0,0 +1,99 @@
+# ProbeShift —— 完整实验规划(5-seed A 跑完之后)
+
+预算状态:4090 总 200 GPU·h,目前累计用掉 ~15h;API $0(全本地)。余量充足。
+原则:**绝大多数补充实验是"在已缓存激活上重分析",几乎零额外 GPU**;少数需要新 GPU 的单独标注。
+
+---
+
+## Tier 0 —— 当前 A 跑(进行中)直接交付
+
+- **G1 基准**:drop + rotation 双维 ground truth,12 概念 × 7 模型 × 5 独立 seed。
+- **G2 评测**:7 预测器 × 4 目标(acc-drop / naive-rotation / IID-resample 安慰剂 / EXCESS)排名 + bootstrap CI + LOCO。
+- **多 seed 复现** + **EXCESS 上 aug vs raptor 的配对显著性检验**。
+- 缓存:`cache_seed0..4`(激活),`results/`(eval/predictors/stats/audit jsonl)。
+
+---
+
+## Tier 1 —— 必需,且只需"重分析已缓存激活"(零/极少新 GPU)
+
+这些都是从 `cache_seed*` 重算指标,新增 `run_pipeline.py` 分析子命令即可,**不重抽激活**。
+
+1. **Shift-type 拆分**(回应:核心 framing + reviewer 必问)
+   现在 rotation 是 paraphrase/domain/length 的均值。**分别**报告每类 shift 下"谁能预测"。预期:
+   EXCESS(去采样噪声)只对 paraphrase 有意义;domain 主要体现在 acc-drop。→ 把"解耦"讲细。
+   成本:重分析(eval row 里已有 per-dist rotation/drop)。
+
+2. **Estimator 外部效度(S2)**:预测器现基于 logreg 方向。复算基于 **mass-mean 方向** 的 rotation/EXCESS,
+   验证结论跨 estimator 成立(mass-mean 在多分类上弱,需谨慎)。
+   成本:重分析(mass-mean 方向已可从缓存重拟合)。
+
+3. **度量消融(M3)**:rotation 目标现用裸 cosine。增加 **ID-白化 cosine 作为 rotation 目标**,
+   验证"循环性 + EXCESS 不可预测"结论不依赖度量取巧(Park 2024 / Truth-Spectrum 都指白化更稳)。
+   成本:重分析。
+
+4. **Layer 稳健性**:现报 IID-best 层。补 **layer-wise 曲线**(每层的 EXCESS 可预测性),
+   证明结论非"挑层"产物。成本:重分析(缓存含全层激活)。
+
+5. **预注册阈值 + 判定表(H1–H6)**:把 PROPOSAL §5 的阈值写成 `preregistration.md`,冻结,
+   对 A 结果逐条报 pass/fail。成本:写文档 + 一次性核对。
+
+6. **Label-fidelity 数值(S3/M5)**:从 `audit.jsonl` 汇报每 shift-type 的 NLI 通过率;
+   作者内部抽检(每类 50 条,**零新众包**)一致率。主动引用 Falsesum/DISCO。成本:重分析 + 少量人工核对。
+
+---
+
+## Tier 2 —— 必需,需要少量新 GPU(都在预算内,API 仍 $0)
+
+7. **多-pivot 多样化增广(最重要的补强)**(回应:EXCESS 赢家是 augmentation;M5)
+   现 n_aug=1(单 de pivot,多 aug 相同)。补 **de/fr/zh/ru 多 pivot 回译** 产生真正多样的增广,
+   重算 augmentation-consistency。预期:aug 在 EXCESS 上的 0.225 **变强/更稳**,坐实"基于增广的信号
+   才捕捉 shift 特异脆性"这一反转结论。
+   成本:下 3 对 opus-mt(fr/zh/ru,本地,$0)+ 对 train 多翻译几遍(~10-20 GPU·h)+ 重算 predict。
+
+8. **补 1–2 个通用-OOD baseline(M4)**:加 **Deng 2023 Confidence&Dispersity** 或
+   **Baek Agreement-on-the-Line**,以及把 **control-task selectivity** 也作为预测器,
+   证明 PAC/aug 的增量超过这些。成本:实现 baseline + 重分析(基本零 GPU)。
+
+9. **OOD-using 上界对照(Truth-Spectrum 风格)**:实现"用 OOD covariance 白化的 cosine"作为
+   **放宽约束的上界**,与所有 IID-only 信号对比,诚实展示"免 OOD 的代价"。成本:重分析。
+
+---
+
+## Tier 3 —— 加分(显著提升说服力)
+
+10. **Size-ladder 可预测性曲线(C4 / M2)**:画"各预测器 ρ vs 模型规模(70M→1.4B + gpt2 家族)"。
+    显式检验 Pressure-Testing-Deception(2605.27958)的 inverse-scaling 是否出现 / 是否 artifact。
+    成本:重分析(A 已含 size ladder)。
+
+11. **更大模型抽检(泛化性,回应"只在小模型")**:对 1 个 ~7B 模型(4-bit 量化,frozen forward)
+    在 3–4 个概念上抽激活、复算核心指标,展示结论不限于 tiny 模型。
+    成本:7B 4-bit 前向 ~15-30 GPU·h(24GB 可载),API $0,零标注。
+
+12. **机制诊断:为什么 aug 行、dispersion 不行**:展示 augmentation-consistency 与 paraphrase-rotation
+    的"残差(扣掉 IID dispersion 后)"特异相关;dispersion 只与 IID-resample 相关。一张机制图。
+    成本:重分析。
+
+---
+
+## Tier 4 —— 锦上添花
+
+13. **实用性 demo**:把最佳 a-priori 信号当"部署前免 OOD 的探针层/estimator 选择器",
+    报告其 train/skip 决策的实际收益(或诚实的"目前都不够好")。成本:重分析。
+
+14. **跨 backbone 家族补充**:若 qwen/gpt2-medium 的 massive-activation 想纳入,加 per-feature
+    标准化(z-score)在 load 时,挽回被 SKIP 的格子。成本:重分析(改 cache.load_shard + 重跑 grid)。
+
+---
+
+## 建议执行顺序(A 跑完后)
+
+**第一波(出论文主体,几乎纯重分析,1 天内):** Tier 1 全部(1–6)→ 这直接撑起 Results/Discussion。
+**第二波(补强 + 防守,2–3 天):** Tier 2(7–9)——多-pivot 增广是重点,直接关系 EXCESS 结论强度。
+**第三波(加分):** Tier 3(10–12)。
+**收尾:** 出图(nature-figure)+ 写 LaTeX 初稿 + 预注册文件 + 复现包。
+
+## 对 ACML 接收的逻辑闭环
+- 核心贡献(G1 基准 + G2 循环性揭示 + EXCESS open problem)由 A 跑 + Tier 1 坐实。
+- EXCESS 上 aug 的反转结论由 Tier 2(多-pivot)补强到显著。
+- Tier 3 把"仅小模型/挑层/单度量"三类质疑全部堵死。
+- 全程 ≤200 GPU·h、API $0、零新标注 —— 预算闭环可复现,benchmark 随缓存发布。

repro_bundle/PRIOR_ART.md

@@ -0,0 +1,586 @@
+# Prior-Art 核查 + Novelty 定位 + 红队报告
+
+## 一、定位文档(逐 claim 撞车判决 + 必引 + 必比 baseline)
+
+Probe Stability Score(PSS)的 prior-art 定位文档如下。
+
+---
+
+## 1. Prior-art 地图(去重后按主题归类)
+
+**A. 探针准确率不可靠 / faithfulness(C1 的 motivation 基石)**
+- Hewitt & Liang 2019(control task / selectivity, EMNLP)
+- Belinkov 2022(Probing Classifiers 综述, CL)
+- Pimentel et al. 2020(信息论探针, ACL);Voita & Titov 2020(MDL, EMNLP)
+- Ravichander, Belinkov & Hovy 2021(EACL);Kumar/Ravfogel et al. 2022(Probing Classifiers are Unreliable, NeurIPS)
+- Boxo et al. 2026(探针依赖文本证据泄漏)
+
+**B. 真值几何 / 概念方向旋转 under shift(C1 核心现象的直接竞品)**
+- Marks & Tegmark 2023/24(Geometry of Truth, mass-mean)— 同时是 C3 的基石
+- **Dies, Maynard, Savcisens & Eliassi-Rad 2025(Representational Stability of Truth, arXiv:2511.19166)— 与 C1+C2 机制最接近的单篇**
+- Liang et al. 2025(LLM Knowledge is Brittle, 2510.11905, paraphrase/domain/length 三类偏移逐字一致)
+- Azizian et al. 2025(Geometries of Truth Are Orthogonal Across Tasks, 2506.08572)
+- Bürger/Levinstein 2024(truth directions context-sensitive, 2404.18865);Bao/Zhang 2025(2506.00823);2601.06599(context 旋转角 θ × size)
+- Park, Choe & Veitch 2024(Linear Representation Hypothesis, ICML)— cosine 度量理论依据
+
+**C. 探针方向稳定性度量(dispersion / augmentation-robustness)(C2 双分量分别被占)**
+- **Gao et al. 2026(RAPTOR, 2602.00158)— directional stability = K 次 ablation 后 mean |cos|,= C2 的 dispersion 分量,且含 size ladder**
+- **Lysnæs-Larsen, Eggen & Strümke 2025(Probing the Probes, 2511.04312)— augmentation robustness = C2 的 augmentation-consistency 分量**
+- Méloux et al. 2025(Mech Interp as Statistical Estimation, 2510.00845)— 重采样方向离散度做可靠性
+- CKA 可靠性系列(Ding et al. 2021;Davari ICLR 2023)— 相似度度量陷阱
+
+**D. 训练前 a priori / 免标注预测探针可靠性或 OOD(C2 定位被多方占据)**
+- **Huang 2025(Spectral Identifiability / SIP, 2511.16288)— eigengap 几何,部署前 a priori 判据**
+- **Reblitz-Richardson 2026(Fragility, 2606.11375)— forward-pass 崩溃点度量**
+- Chou et al. 2026(Geometry Markers, ICLR 2026 / 2603.01879)— ID-only 几何预测 OOD(视觉)
+- CCPS 2025(2505.21772)— 隐状态扰动做置信度校准
+
+**E. 免标注 OOD 性能预测(通用,非探针,C2 必比 baseline)**
+- **Xie et al. 2023(Dispersion Score, NeurIPS, 2303.15488)— 同名同质:label-free + forward-pass + feature dispersion 预测 OOD acc**
+- Deng et al. 2023(Confidence & Dispersity, ICML)— 用了 "dispersity" 一词
+- Baek et al. 2024(Agreement-on-the-Line, NeurIPS)
+- Aithal et al. 2021(Robustness to Augmentations as Generalization Metric, NeurIPS comp)
+
+**F. Transferability estimation 范式(C2 自认类比来源)**
+- LEEP(ICML 2020)、LogME(ICML 2021)、H-score(ICIP 2019)
+- Zhang et al. 2026(SAE as a Crystal Ball, 2603.02908)— LLM 版免训练事前预测跨域迁移
+- **Truthfulness Spectrum 2026(2602.20273)— Mahalanobis-白化方向相似度预测 OOD AUROC, R²≈0.98**
+
+**G. C3 专属(探针估计子比较)**
+- Marks & Tegmark 2023(mass-mean vs LogReg vs CCS);SAE-vs-LogReg 系列(2502.16681)
+
+**H. C4 专属(size ladder 基建)**
+- Biderman et al. 2023(Pythia, ICML);2505.21399(Factual Self-Awareness, Pythia 全 ladder × 扰动);2604.13386(探针 acc 随 size)
+
+**I. C5 专属(NLI label-fidelity 审计)**
+- NLI overgenerate-and-filter / roundtrip consistency 范式(Falsesum 2022、DISCO 2023 等);counterfactual augmentation(Kaushik 2020 ICLR、Gardner contrast sets 2020)
+
+---
+
+## 2. 逐 claim 撞车判决表
+
+| Claim | 最接近的已有工作 | 重叠程度 | 我们仍站得住的 defensible delta |
+|---|---|---|---|
+| **C1** 准确率高估稳定性 + accuracy-drop × cosine-rotation 矩阵 across shift×size | Liang 2025《LLM Knowledge is Brittle》(paraphrase/domain/length 逐字一致);Dies 2025(方向旋转 under rephrase);2505.21399(Pythia ladder × 扰动) | **高(crowded)**。"现象"已被占;Liang 2025 连三类 shift 都一致 | 不能作主贡献。唯一 delta:把 accuracy-drop 与 cosine-rotation **解耦** 成 shift-type × model-size **双指标二维矩阵**,且跨多概念(非仅 truth);Liang/Dies 都只做 truth 单概念、未做 size×shift 完整矩阵化、未把 rotation 与 drop 解耦量化 |
+| **C2** 免标注 forward-pass 双分量 PSS(dispersion ⊕ augmentation-consistency)训练前 a priori 预测 OOD 迁移 | dispersion 分量→RAPTOR(2602.00158);aug-consistency 分量→Probing the Probes(2511.04312);a priori 定位→SIP(2511.16288)、Fragility(2606.11375);命名→Xie 2023 Dispersion Score;Dies 2025(rotation+dispersion 已合做) | **高,但有缝**。两分量各自被占,a priori 框架被多方占,但**"双分量合成单一分数 + 验证它能 a priori 预测 OOD drop/rotation 的预测-验证闭环"无人完整做** | **论文命脉**��delta:(a) 双分量**合成**并做互补性消融(证明 aug-consistency 在 dispersion 之外有增量);(b) 预测目标是 **probe 方向几何稳定性**(非 RAPTOR 的诊断、非 Xie 的整体 acc、非 SIP 的 ID 可识别性);(c) **完全免新标注/免触碰 OOD 数据**(对比 Truthfulness Spectrum 2602.20273 需 OOD reference probe);(d) 建立**预测↔验证闭环**(用 PSS 排序 C1 矩阵里的真实 drop) |
+| **C3** mass-mean vs LogReg vs MLP 在 OOD 下相对可迁移性 | Marks & Tegmark 2023(mass-mean vs LogReg vs CCS,结论 mass-mean 更稳) | **高**。核心结论已发表 | 单独看几乎是复现。delta 只剩:加入 **MLP** 第三方;用 **label-preserving 输入偏移**(非换数据集)做 OOD;跨 **size ladder**;并把比较结果**绑定到 PSS 预测**。建议**降格为支撑性发现**,不作并列主贡献 |
+| **C4** model size ladder 对探针 OOD 稳定性影响 | 2505.21399(Pythia ladder × 扰动);2601.06599、2604.13386(方向一致性/acc 随 size);Pythia suite 本身 | **中-高**。"越大越一致/越准"已被多篇占;甚至有"越大越脆"反例(Pressure-Testing Deception 2605.27958) | 单独**不足以成贡献**。仅作 C1/C2 的交叉维度。可挖的 delta:size 对 **PSS 预测有效性** 的影响,及是否复现/反驳"越大越脆" |
+| **C5** 本地 DeBERTa-MNLI 做 label-fidelity 审计保证 shift label-preserving | NLI overgenerate-and-filter / roundtrip consistency(Falsesum 2022、DISCO 2023);Kaushik 2020;Gardner 2020 | **中**。"用 NLI 核验增广 label 一致性"是 2022–24 成熟范式 | **放弃作为新颖性**,定位为方法学 sanity check / 可信度保障。delta 仅在"应用到 probe-stability 评测的 shift 构造管线"这一工程实例化,需主动引用过滤文献以示知情 |
+
+---
+
+## 3. 整体 novelty 判决
+
+**判决:crowded,但存在一条可辩护的窄缝。**
+
+- C1、C3、C4、C5 单独看**均已被不同程度抢占**(C1/C3 = high)。
+- C2 的两个分量、a priori 框架、甚至 "dispersion/dispersity" 命名都已分别被占;Dies 2025 已把 rotation + dispersion 合做——这是最危险的单篇。
+
+**安全的收窄组合(必须按此重写 contribution statement):**
+
+> 把论文重心从"发现探针不稳 + 比较探针类型"(已被占)收窄到:
+> **"一个针对 linear/concept probe 方向几何、完全免新标注、仅 forward-pass 的双分量(dispersion ⊕ augmentation-consistency)合成分数,首次在训练前 a priori 预测探针在 label-preserving 语义偏移(paraphrase/domain/length)下的 OOD 迁移性,并通过 shift-type × model-size 矩阵 + 与 SIP/Fragility/Dispersion-Score/mass-mean 的 head-to-head baseline 对比,证明该分数提供超越单分量与现有几何/谱预测器的增量。"**
+
+护城河三要素,缺一即被判 incremental:
+1. **双分量合成 + 互补性消融**(无人合做两分量);
+2. **a priori 预测 ↔ 验证闭环**(用 PSS 预测 C1 矩阵真实 drop;Dies/RAPTOR/SIP 都停在诊断或 ID 判据);
+3. **完全免触碰 OOD 数据/新标注**(对比 Truthfulness Spectrum 的 OOD reference probe 需求)。
+将 C1 降为"被预测的真值",C3/C4 降为外部效度,C5 降为 sanity check。
+
+---
+
+## 4. 必引清单(must-cite,漏引会被直接攻击)
+
+**绝对必引(high-threat,逐篇做"我们 vs 他们"切割):**
+1. Dies, Maynard, Savcisens & Eliassi-Rad 2025, *Representational Stability of Truth in LLMs*(2511.19166)——最近邻,rotation+dispersion 已合做。
+2. Gao et al. 2026, *RAPTOR: Ridge-Adaptive Logistic Probes*(2602.00158)——dispersion 分量 + size ladder。
+3. Lysnæs-Larsen, Eggen & Strümke 2025, *Probing the Probes*(2511.04312)——augmentation-robustness 分量。
+4. Huang 2025, *Spectral Identifiability for Interpretable Probe Geometry*(2511.16288)——a priori 探针稳定性判据。
+5. Reblitz-Richardson 2026, *When Probing Accuracy Saturates, Fragility Resolves*(2606.11375)——forward-pass 崩溃点。
+6. Marks & Tegmark 2023/24, *The Geometry of Truth*(2310.06824)——mass-mean vs LogReg(C3 基石)。
+7. Xie et al. 2023, *Dispersion Score*(NeurIPS, 2303.15488)——同名同质 OOD 预测器。
+8. Liang et al. 2025, *LLM Knowledge is Brittle*(2510.11905)——C1 三类 shift 逐字一致。
+9. Truthfulness Spectrum 2026(2602.20273)——白化方向相似度预测 OOD(C2 预测目标最近邻)。
+10. Chou et al. 2026, *Diagnosing Generalization Failures from Geometry Markers*(ICLR 2026 / 2603.01879)——ID-only 预测 OOD 范式。
+
+**必引背景(定位用):** Hewitt & Liang 2019;Belinkov 2022;Kumar/Ravfogel et al. 2022(NeurIPS);Voita & Titov 2020;Azizian et al. 2025(2506.08572);Bürger 2024(2404.18865)/ Bao 2025(2506.00823);LEEP 2020 / LogME 2021;Deng et al. 2023(Confidence & Dispersity);Baek et al. 2024(Agreement-on-the-Line);Zhang 2026 SAE Crystal Ball;Biderman et al. 2023(Pythia)+ 2505.21399;NLI-filter 系列(C5)。
+
+> 注:所有 2025–2026 编号(2511.x / 2602.x / 2603.x / 2606.x / 2602.20273 / 2605.27958 等)venue 与作者多为"待核实",投稿前必须在 arXiv/OpenReview 逐条���核标题、作者、是否已被 ACML/NeurIPS/ICLR/ACL 接收——prior-art 效力取决于此。
+
+---
+
+## 5. 必比 baseline 清单(must-have baselines,不比会被质疑)
+
+**A. a priori 探针稳定性 / 可靠性预测器(同定位,直接 head-to-head,核心)**
+1. **SIP / Spectral Identifiability**(eigengap 几何判据)— 必比,否则"首个 a priori 预测器"主张崩塌。
+2. **RAPTOR directional stability**(K 次 ablation mean |cos|)— 即单独的 dispersion 分量,用作消融对照。
+3. **Fragility**(临界噪声崩溃点)— forward-pass 稳定性对照。
+4. **Augmentation-robustness**(Probing the Probes)— 即单独的 augmentation-consistency 分量,用作消融对照。
+
+**B. 通用免标注 OOD 性能预测器(证明探针方向几何 ≠ 输出/流形几何)**
+5. **Dispersion Score**(Xie 2023)— 同名,必跑作 baseline,否则被判换皮。
+6. **Confidence & Dispersity**(Deng 2023)/ **Agreement-on-the-Line**(Baek 2024)— 至少跑其一。
+7. **Geometry Markers**(effective manifold dim / utility)— ID-only 几何预测对照。
+
+**C. 探针估计子(C3 内部对照)**
+8. **mass-mean / difference-of-means**(Marks & Tegmark)— 必含。
+9. **LogReg** 与 **MLP** 探针 — C3 三方比较的另两方。
+
+**D. 方法学对照(C1/C5)**
+10. **control-task / selectivity probe**(Hewitt & Liang)— 作为"已知探针诊断"的对照基线,证明 PSS 提供 selectivity 之外的信息。
+11. **普通 cosine vs Mahalanobis-白化 cosine**(Truthfulness Spectrum)— 方向相似度度量的对照,回应"普通 cosine 不够"的已知结论。
+
+**消融(内部 must-have):** dispersion-only / augmentation-consistency-only / 双分量合成 三档,证明互补增量;以及"PSS 预测的 OOD drop vs 实际 OOD drop"的相关性/排序能力(Spearman ρ、Kendall τ)作为闭环验证主结果。
+
+---
+
+**一句话结论:** 现状是 crowded;唯一安全的护城河是把 C2 收窄成"双分量合成 + a priori 预测↔验证闭环 + 完全免 OOD 标注 + 针对 probe 方向几何",并以 SIP / Fragility / RAPTOR-stability / Dispersion-Score / Augmentation-robustness 五个 baseline 的 head-to-head 增量来支撑;C1 降为被预测真值,C3/C4 降为外部效度,C5 降为 sanity check。**关键风险点:Dies 2025(2511.19166)必须拿到全文确认它是否已做"预测 OOD"闭环——若已做,C2 护城河将进一步收窄,需立即二次检索核实。**
+
+
+## 二、Novelty 红队(最强攻击 + 必需防守)
+
+# ACML 苛刻 Reviewer 意见:Reject(Probe Stability Score,PSS)
+
+我已通读定位文档。下面分两档给出我作为对抗性 reviewer 准备的攻击,每条都点名撞车对象/设计漏洞,并配上作者"活下来"所需的最小防守动作。
+
+---
+
+## 一、致命级攻击(任一条不解决即建议 Reject)
+
+### F1. 与 Dies et al. 2025(2511.19166)的"组合新颖性"被整体抹平
+**攻击点:** 作者的全部卖点是"rotation + dispersion 双视角"。Dies 2025 已经把 representational stability of truth 用 rotation + dispersion 合做了。作者文档自己承认"这是最危险的单篇"。如果 Dies 已经做了"用 ID 几何预测 OOD 行为"的闭环,那 C2 护城河的(1)(2)两要素同时塌掉,PSS 只剩"换了个加权求和的公式 + 多了几个概念",这是教科书级的 incremental。
+**这撞了什么:** 直接撞 Dies 2025 的方法核心,且作者尚未确认对方是否已做预测闭环。
+**必需防守:**
+- 拿到 Dies 2025 全文,做一张**逐能力对照表**(rotation? dispersion? 二者合成单一分数? a priori 预测 OOD? 免 OOD 数据? 跨多概念? size ladder?),逐格标 Y/N。
+- 必须证明至少一格是"他们 N、我们 Y"且该格是**预测闭环或合成分数**,而非边角(概念数量、模型数量这类不算)。
+- 在同一 benchmark 上**复现 Dies 作为 baseline 并跑输/跑平**——若跑不赢,论文死。
+
+### F2. "首个 a priori 预测器"主张直接被 SIP / Fragility / Truthfulness Spectrum 证伪
+**攻击点:** 文档同时列了 SIP(2511.16288,部署前 a priori 几何判据)、Fragility(2606.11375,forward-pass 崩溃点)、Truthfulness Spectrum(2602.20273,白化方向相似度预测 OOD AUROC,R²≈0.98)。"训练前/免 OOD 预测探针稳定性"这一 framing 已被至少三方占据。任何"首次"措辞都可被一句话击穿。
+**这撞了什么:** SIP + Fragility + Truthfulness Spectrum,三方夹击 framing 新颖性。
+**必需防守:**
+- **删除一切"first / novel framing"措辞**,改为"a label-free composite that empirically outperforms SIP / Fragility / Dispersion-Score on probe-direction OOD prediction"。
+- 三者**全部作为 head-to-head baseline 跑数**,主结果用 Spearman ρ / Kendall τ 报告 PSS 的增量,且增量需带置信区间/显著性检验(bootstrap)。
+- 特别针对 Truthfulness Spectrum 的 R²≈0.98:若对方已能 R²=0.98 预测 OOD,作者必须说明 PSS 在"完全不碰 OOD reference probe"约束下还能逼近此精度,否则"免 OOD 数据"这条护城河不值钱。
+
+### F3. 与 Xie 2023 Dispersion Score 同名同质,涉嫌"换皮"
+**攻击点:** Xie 2023(NeurIPS)已有 label-free + forward-pass + feature dispersion 预测 OOD acc,名字都一样。reviewer 第一反应是"这就是把 Dispersion Score 套到 probe 方向上"。
+**这撞了什么:** Xie 2023,命名与机制双重撞车。
+**必需防守:**
+- **改名**,避免 "Dispersion" 作主词根,杜绝第一眼误判。
+- 用实验证明"probe 方向几何 ≠ feature 流形 dispersion":在同一数据上跑 Xie 的 Dispersion Score,展示二者预测目标/相关性显著分离(例如 Dispersion Score 在 paraphrase shift 上失效而 PSS 不失效)。
+
+### F4. 预测目标循环论证 / 闭环自证(实验设计漏洞)
+**攻击点:** "用 PSS 预测 C1 矩阵里的真实 drop"——但 C1 矩阵的 shift、概念、size 配置若与构造 PSS / 调权重所用的是同一批数据,则是在**训练集上验证**,Spearman ρ 没有意义。双分量"合成"必然涉及一个加权/聚合超参,若该超参在评测集上选,闭环就是过拟合。
+**这撞了什么:** 实验设计的数据泄漏 / 无 held-out。
+**必需防守:**
+- 严格 **split**:权重/聚合超参只在 dev 概念-shift 组合上选,报告结果只在**完全 held-out 的概念 × shift-type × size 组合**上算。
+- 提供 **leave-one-concept-out / leave-one-shift-type-out** 交叉验证,证明 PSS 跨概念、跨 shift-type 泛化,而非记忆 C1 矩阵。
+
+### F5. C3 几乎是 Marks & Tegmark 2023 的复现
+**攻击点:** "mass-mean vs LogReg 在 OOD 下谁更稳"结论已发表。加一个 MLP、换 label-preserving shift,不足以构成独立贡献。若作者把 C3 列为并列主贡献,reviewer 直接判 partial reproduction。
+**这撞了什么:** Marks & Tegmark 2023。
+**必需防守:** 按文档自己的结论,**C3 降格为支撑性发现/外部效度**,且只在"PSS 能否预测三类 estimator 谁更稳"这一从属问题下出现,绝不并列。
+
+---
+
+## 二、重要级攻击(不致命但累加可压到 Reject)
+
+### M1. 互补性消融若不显著,双分量合成失去存在理由
+**攻击点:** PSS 的核心增量主张是"aug-consistency 在 dispersion 之外有增量"。若 dispersion-only 与双分量合成的 ρ 无显著差异,则合成是噱头。
+**必需防守:** 三档消融(dispersion-only / aug-only / 合成)必须报告**配对显著性检验 + 效应量**,且合成显著优于两个单分量。否则坦诚降级为单分量方法。
+
+### M2. C4(size ladder)无独立贡献,且有反例风险
+**攻击点:** "越大越稳/越一致"已被多篇占;且存在"越大越脆"反例(Pressure-Testing Deception 2605.27958)。若作者下"越大越稳"结论而不处理反例,会被打成 cherry-picking。
+**必需防守:** C4 仅作交叉维度;明确测试并讨论"PSS 预测有效性是否随 size 单调",主动引用并复现/反驳反例,不下普适性结论。
+
+### M3. 度量选择(普通 cosine)已被已知结论否定
+**攻击点:** Park 2024 LRH 用 cosine 有前提;Truthfulness Spectrum 已表明需 Mahalanobis 白化。用裸 cosine 度量 rotation 会被质疑度量不当。
+**必需防守:** 加入**白化 cosine vs 裸 cosine** 的消融,说明 PSS 的稳健性不依赖度量取巧。
+
+### M4. baseline 覆盖不足即被判"未与最近邻比较"
+**攻击点:** 文档列出的 must-have baseline(SIP / RAPTOR-stability / Fragility / Aug-robustness / Dispersion-Score / mass-mean / Geometry Markers / control-task)若缺任一核心项,reviewer 可直接以"未与显然相关工作比较"拒稿。
+**必需防守:** 至少 A 组四个 a priori 预测器(SIP / RAPTOR-stability / Fragility / Aug-robustness)+ Xie Dispersion Score **全跑**;B/C 组至少各跑一个。缺项需在 limitation 明确说明理由。
+
+### M5. shift 的 label-preserving 保证仅靠单个 DeBERTa-MNLI,可信度不足
+**攻击点:** C5 用本地 DeBERTa-MNLI 审计 label fidelity。单模型审计、未引 overgenerate-and-filter / roundtrip(Falsesum 2022、DISCO 2023)文献,会被质疑审计本身不可靠且不知情。
+**必需防守:** C5 明确降为 sanity check;主动引用 NLI-filter 系列;报告 MNLI 审计的**通过率与人工抽检一致性**,证明 shift 确为 label-preserving。
+
+### M6. 跨概念泛化样本量与统计功效存疑
+**攻击点:** 若"多概念"只有 3–5 个概念,Spearman ρ 的置信区间会很宽,"跨多概念优于 truth 单概念"的卖点站不住。
+**必需防守:** 概念数足够大(建议 ≥10),报告 ρ 的 bootstrap CI;若概念少,撤回"跨多概念"作为新颖性主张。
+
+### M7. prior-art 编号未核实,存在 prior-art 效力风险(双刃)
+**攻击点:** 大量 2025–2026 arXiv 编号"待核实"。reviewer 若发现作者**未引用**已正式接收的某篇最近邻(如 Dies/RAPTOR 若已中 NeurIPS/ICLR),漏引即攻击点;反之若作者拿"未发表 arXiv"当挡箭牌排除 prior art,也会被驳。
+**必需防守:** 投稿前逐条核实标题/作者/venue/接收状态;对所有 high-threat 篇目无论是否正式发表均做切割讨论,不以"preprint"为由回避。
+
+---
+
+## 三、给作者的一句话总结(survival path)
+
+**当前状态 crowded,唯一活路:** 删掉所有"first/novel framing"措辞,把贡献彻底收窄为"label-free 双分量合成分数 + 严格 held-out 的 a priori 预测↔验证闭环 + 完全免 OOD 数据约束下对 SIP/Fragility/Aug-robustness/Dispersion-Score 五个 baseline 的可显著性检验的增量",并以 **F1(Dies 全文对照表跑赢)+ F4(无泄漏 split)+ M1(互补性显著)** 三者同时成立为生死线。C1 降为被预测真值,C3/C4 降为外部效度,C5 降为 sanity check。三条生死线缺一,我维持 Reject。
+
+
+## 三、原始检索明细(10 角度)
+
+
+### 角度: 探针方法学与 control task(linear probing / diagnostic classifiers / selectivity / cont...
+
+- **Hewitt & Liang, 2019** — *Designing and Interpreting Probes with Control Tasks* (EMNLP 2019) [威胁:medium]
+  - 概念上撞 C1 的核心论点(探针准确率会高估表征中真正编码的信息)与 C2/C3 的探针选择哲学。但它用『随机标签控制 + selectivity』这一轴,而非『label-preserving 输入分布偏移 + 方向 cosine 旋转 + dispersion/augmentation-consistency』。是我们必须正面区分的奠基工作,但范式不同。
+- **Belinkov, 2022** — *Probing Classifiers: Promises, Shortcomings, and Advances* (Computational Linguistics (MIT Press), 48(1)) [威胁:low]
+  - 背景层面覆盖 C1/C2/C5 所处问题空间(探针可靠性诊断),但不提供我们提议的具体度量(分布偏移下的方向旋转矩阵、免标注稳定性分数)。是定位 novelty 的必引背景,本身不抢占。
+- **Pimentel et al., 2020** — *Information-Theoretic Probing for Linguistic Structure* (ACL 2020) [威胁:low]
+  - 与 C1『准确率作为表征质量代理不可靠』同源动机,但走信息论方向,不涉及分布偏移、方向稳定性或免标注预测。背景相关。
+- **Marks & Tegmark, 2023/2024** — *The Geometry of Truth: Emergent Linear Structure in LLM Representations of True/False Datasets* (COLM 2024(arXiv 2310.06824)) [威胁:high]
+  - 直接、强烈撞 C3:mass-mean vs LogReg 在 OOD 下相对可迁移性的系统比较正是该文核心结果之一。C3 若不显著扩展(如加入 MLP、跨 shift type × model size 的系统矩阵、把比较绑定到我们的稳定性分数上)将被认为已被做过。
+- **Anonymous / 待核实, 2026** — *Pressure-Testing Deception Probes in LLMs: Scaling, Robustness, and the Geometry of Deceptive Representations* (arXiv 2605.27958(2026,预印本,待核实正式 venue)) [威胁:high]
+  - 高度撞 C1(探针准确率在 label-preserving 风格偏移下崩溃,高估稳定性)与 C4(model size ladder 对探针 OOD 稳定性的影响,且给出反直觉结论)。也部分触及 C3(线性 vs MLP)。区别:它未度量概念方向的 cosine 旋转,未提免标注 a-priori 稳定性分数,且把 augmentation 当作『修复手段』而非我们 C2 中的『预测信号』。这是本角度对 C1+C4 威胁最大的近期工作。
+- **待核实, 2025** — *Calibrating LLM Confidence by Probing Perturbed Representation Stability (CCPS)* (arXiv 2505.21772(2025,预印本)) [威胁:medium]
+  - 撞 C2 的『表征/方向稳定性作为可靠性信号』直觉,且同样是 forward-pass 轻量方案。但关键差异:扰动作用于隐状态(非 label-preserving 输入增广/paraphrase),目标是 per-answer 置信度校准而非『训练前预测探针 OOD 迁移性』,也无 dispersion(重采样方向离散度)分量。相关但目标错位。
+- **待核实, 2025** — *Probing the Geometry of Truth: Consistency and Generalization of Truth Directions in LLMs Across Logical Transformations and QA Tasks* (ACL Findings 2025(arXiv 2506.00823)) [威胁:medium]
+  - 撞 C1 的方向一致性部分:它确实在『变换后概念方向是否保持』这一轴上做了实证。差异:变换是逻辑/任务层面而非 paraphrase/domain/length 的 label-preserving 表层偏移;未给出 accuracy drop × cosine 旋转 的跨 shift×scale 矩阵,也未提免标注稳定性分数。是 C1 方向旋转主张需正面区分的工作。
+- **Ravichander et al. / 待核实, 2022** — *Probing Classifiers are Unreliable for Concept Removal and Detection* (NeurIPS 2022) [威胁:low]
+  - 概念上支撑 C1 的批判性前提(探针准确率高估表征中概念的真实存在/稳定性),但聚焦因果性与 concept removal,不涉及分布偏移下的方向稳定性度量或免标注预测分数。背景相关。
+- **待核实, 2026** — *Diagnosing Generalization Failures from Representational Geometry Markers* (arXiv 2603.01879(2026,预印本)) [威胁:medium]
+  - 强烈撞 C2 的方法论框架(免 OOD 标注、用 ID 侧量化指标 a-priori 预测 OOD 迁移性)。差异:用 manifold 几何指标而非探针方向 dispersion/augmentation-consistency,且为视觉模型、非语言探针、不比较探针类型。是 C2『免标注预测 OOD 迁移性』这一卖点的概念竞品,需说明我们的双分量方向指标与之正交。
+- **待核实, 2022** — *Predicting Fine-Tuning Performance with Probing* (EMNLP 2022(arXiv 2210.07352,待核实)) [威胁:low]
+  - 与 C2『用探针侧信号预测下游/迁移表现』同向,但预测目标是 fine-tuning 性能而非探针自身 OOD 稳定性,且不涉及方向 dispersion/augmentation-consistency。弱相关。
+
+> 评估: 已联网检索。本角度的结论:你们的五条贡献没有被任何单篇完整抢占,但每条都已被『部分前置』,且有两个高危撞车点。
+
+最大撞车点(高危):
+1) C3(mass-mean vs LogReg 的 OOD 相对可迁移性比较)几乎被 Marks & Tegmark, The Geometry of Truth(COLM 2024)直接做过——他们正是提出 mass-mean 并系统对比其与 LogReg/CCS 在跨数据集 transfer 下更鲁棒。C3 单独看 novelty 很弱,必须靠『加入 MLP 第三方、跨 shift type × model size 的系统矩阵、并把比较结果绑定到你们的稳定性分数』来续命,否则会被审稿人直接判定重复。
+2) C1+C4 的核心实证(label-preserving 风格偏移下探针准确率崩塃、且随 model size 出现非单调/逆向 scaling、augmentation 可缓解)与 2026 的 Pressure-Testing Deception Probes 高度重叠。你们相对它的唯一硬区分是『概念方向 cosine 旋转矩阵』这一几何量(它只看 AUROC 崩塃,不看方向旋转)以及把 augmentation 当预测信号而非修复手段——这两点必须在论文里前置强调,否则 C1/C4 会被认为是该文的子集。
+
+仍然空白(你们真正的护城河):
+- C2 的具体形式——『dispersion(同分布 K 次重采样方向离散度)+ augmentation-consistency(label-preserving 增广后方向一致性)』的双分量、免新标注、纯 forward-pass、用于训练前 a-priori 预测探针 OOD 迁移性——在本角度未找到完全对应工作。CCPS(隐状态扰动做置信度校准)与 Geometry Markers(ID 几何预测 OOD,视觉域)分别占了『稳定性信号』和『免标注 a-priori 预测』两个相邻 niche,但都不是『探针方向的重采样离散度 + 增广一致性』。这是你们最干净的 novelty,应作为主卖点突出,并显式与这两篇划界。
+- C5(用本地 DeBERTa-MNLI 做 label-fidelity 审计来保证 shift 是 label-preserving)在本角度未见任何探针文献采用;这是方法学增量(虽不足以单独成贡献),建议作为支撑 C1 严谨性的卖点而非独立 claim。
+
+行动建议:把论文 framing 从『我们发现探针不稳定 + 我们比较探针类型』(已被占)转向『我们提出免标注、训练前可计算的方向级稳定性分数,并证明它能预测 OOD 迁移』(C2 为核心),把 C1/C3/C4 降格为支撑性诊断证据,并在 related work 中点名 Marks & Tegmark、Pressure-Testing Deception Probes、CCPS、Geometry Markers 做正面切割。注意:多篇 2026 预印本(2605.27958、2603.01879、ICLR 2026 kHVzEjThKE)venue 与作者仍待核实,投稿前请二次确认。
+
+### 角度: 探针到底测了什么、探针可靠性争议(amnesic probing / information-theoretic probing / probe vs mode...
+
+- **Elazar, Ravfogel, Jacovi & Goldberg, 2021 (arXiv 2020)** — *Amnesic Probing: Behavioral Explanation with Amnesic Counterfactuals* (TACL 2021) [威胁:medium]
+  - 与 C1 的动机层强重叠:都在质疑探针准确率高估了特征的真实地位。但 Amnesic Probing 质疑的是'被模型使用与否(causal usage)',而非'在 label-preserving 输入分布偏移下是否稳定(robustness/OOD)';它需要反事实擦除而非纯 forward-pass,也不给 a priori 预测分数。
+- **Voita & Titov, 2020** — *Information-Theoretic Probing with Minimum Description Length* (EMNLP 2020) [威胁:medium]
+  - 与 C1+C2 概念层重叠:同样主张'accuracy 不足以衡量特征被编码的可靠程度',并提出替代度量。但 MDL 衡量的是同分布下的编码努力/压缩,不涉及 paraphrase/domain/length 偏移,也不预测 OOD 迁移、不含方向稳定性或增广一致性。是你度量动机的最重要先验之一,需在 related work 明确区分。
+- **Hewitt & Liang, 2019** — *Designing and Interpreting Probes with Control Tasks* (EMNLP 2019) [威胁:medium]
+  - 与 C1 动机重叠(accuracy 会高估)且与 C3 间接相关(线性 vs 非线性探针的可信度差异,呼应你 mass-mean/LogReg/MLP 比较)。但 selectivity 衡量的是'探针自身容量混淆',不是输入分布偏移下的稳定性,也无方向旋转、无 OOD 预测分数。
+- **Belinkov, 2022** — *Probing Classifiers: Promises, Shortcomings, and Advances* (Computational Linguistics (MIT Press) 2022) [威胁:medium]
+  - 与 C1 高度重叠在'问题陈述'层面——你要量化的'accuracy 高估特征稳定性'正是该综述列出的核心缺陷之一。威胁不在于它做了你的实验,而在于审稿人会用它说'这个问题早已是共识'。需要你用具体的偏移矩阵+方向旋转把'已知缺陷'落成'可量化诊断'来体现增量。
+- **Marks & Tegmark, 2023/2024** — *The Geometry of Truth: Emergent Linear Structure in LLM Representations of True/False Datasets* (COLM 2024 (arXiv 2310.06824)) [威胁:high]
+  - 对 C3 是最直接威胁:'mass-mean vs LogReg 的相对(跨分布)可迁移性比较'核心结论它已给出,且方法上引入了 difference-of-means 探针的现代用法。你的 C3 若仅复述'mass-mean 更可迁移'会被判为已知。差异点:它聚焦 truth 这一概念、用跨数据集泛化而非系统化的 paraphrase/domain/length label-preserving 偏移矩阵,也无 MLP、无 model-size ladder、无 a priori 稳定性分数。
+- **Gao, Zhu, Zeng, Zhao, Wang, Ruan & Ding, 2026** — *RAPTOR: Ridge-Adaptive Logistic Probes* (arXiv 2602.00158 (2026, venue 待核实)) [威胁:high]
+  - 对 C2 的 dispersion 分量是最近、最危险的撞车:它已经把'对训练数据扰动下的方向 cosine 稳定性'作为正式评测指标。区别在于 RAPTOR 把方向稳定性当作改进探针训练的目标/评估,而非'训练前、免标注、用来预测 OOD 迁移的双分量预测分数',也不含 augmentation-consistency 分量、不做 paraphrase/domain 偏移迁移、不做 size ladder。务必引用并清楚切割:你是 a priori 预测器,它是 robustness 评估+方法改进。
+- **Boxo, Neelappa & Raval, 2026** — *Linear Probes Rely on Textual Evidence (leakage mitigation in LLM probes)* (ICML 2026 (arXiv 2509.21344, venue 待核实)) [威胁:medium]
+  - 与 C1 动机强重叠的'近作':直接给出'探针准确率虚高、表征不稳健'的量化证据,会被用来质疑你 C1 的新颖性。但它走的是 token-level 证据泄漏路径,不是 label-preserving 输入分布偏移,也无方向旋转、无 a priori 分数、无 mass-mean/size 比较。是必须 cite 并区分的同期工作。
+- **Anonymous/待核实, 2025-2026** — *No Answer Needed: Predicting LLM Answer Accuracy from Question-Only Linear Probes* (OpenReview (待核实)) [威胁:low]
+  - 与 C2 的'仅 forward-pass、事前预测'范式精神相近,可能被审稿人援引为'事前探针预测'已有先例。但它预测的是答案正确性,不是探针自身的 OOD 可迁移性,目标对象完全不同,无方向稳定性/增广一致性。
+- **Ravichander, Belinkov & Hovy, 2021** — *Probing the Probing Paradigm: Does Probing Accuracy Entail Task Relevance?* (EACL 2021) [威胁:medium]
+  - 与 C1 动机重叠:再次证明 accuracy 会高估特征的真实地位。但关注'任务相关性'而非分布偏移稳健性,无方向旋转、无预测分数、无 C3/C4。属背景到中等的反例支撑文献。
+
+> 评估: 已联网检索(WebSearch/WebFetch 可用)。本角度结论如下。
+
+最大撞车点(按风险排序):
+1) C3 撞 Marks & Tegmark《The Geometry of Truth》(COLM 2024, high)。'mass-mean(difference-of-means)vs LogReg 在跨分布下更可迁移'这一核心结论已被发表,且 mass-mean 探针的现代用法即源于此。C3 不能停在'mass-mean 更稳';必须靠'系统化 shift type × model size 矩阵 + 加入 MLP + label-fidelity 审计'来制造增量,否则会被直接判为已知。
+2) C2 的 dispersion 分量撞 RAPTOR(arXiv 2026, high)。'对训练数据扰动的方向 cosine 稳定性'已被它作为正式 directional-robustness 指标。你的护城河必须落在 RAPTOR 没有的三点:(a) 这是训练前、免新标注的 a priori 预测器,目标是预测 OOD 迁移性而非改进探针;(b) 双分量,额外含 augmentation-consistency;(c) 跨 paraphrase/domain/length 偏移与 size ladder 的预测有效性验证。强烈建议把'我们的分数能 a priori 预测 C1 矩阵里的 accuracy drop / 方向旋转'做成核心实验,这是与所有先验最干净的切割。
+3) C1 的'问题本身'被大量经典+近作占据(Belinkov 综述、Hewitt&Liang、Ravichander、Voita&Titov,以及 2026 的 leakage 论文)。'探针准确率高估特征可靠性'已是领域共识级命题,审稿人极可能说'not new'。C1 的新颖性只能来自'量化方式'——即把已知缺陷落成 accuracy drop + 概念方向 cosine 旋转 的双轴矩阵,且明确限定在 label-preserving 输入分布偏移(而非 token 泄漏、控制任务或因果擦除)。务必在 related work 用一句话各自切割上述四篇。
+
+仍有空白(你的真实可辩护新颖性):
+- 没有任何一篇把'同分布重采样方向离散度 + label-preserving 增广方向一致性'组合成单一、免标注、训练前的探针稳定性预测分数,并验证其对 OOD 迁移的预测力。这是 C2 最干净的空白,应作为论文第一卖点。
+- 没有人用'概念方向 cosine 旋转'作为跨 paraphrase/domain/length 偏移的稳定性轴(现有方向-cosine 工作都在'数据扰动/seed'维度,如 RAPTOR;偏移维度是空白)。C1 的方向旋转矩阵因此有真增量。
+- C4(model-size ladder 对探针 OOD 稳定性的影响)在本角度文献里基本无人系统做(本角度多为单模型/单规模),是相对安全的贡献,但单独不足以撑论文。
+- C5(本地 NLI/DeBERTa-MNLI 做 label-fidelity 审计以保证 shift 是 label-preserving)在本角度未见对应工作,属方法学卫生措施而非独立新颖点,审稿人不会视为贡献但会因其缺失而质疑 C1 的有效性——保留但别当卖点。
+
+行动建议:把叙事重心从 C1(易被判已知)转到 C2(a priori 免标注预测器)为主轴,用 C1 矩阵作为'被预测的真值'、C3/C4 作为外部效度验
+
+### 角度: 线性表示假设 / 真值几何 / 概念方向(linear representation hypothesis, geometry of truth, concep...
+
+- **Liang et al. (2025), 作者名待核实** — *LLM Knowledge is Brittle: Truthfulness Representations Rely on Superficial Resemblance* (arXiv:2510.11905(2025-10);会议归属待核实) [威胁:high]
+  - C1 几乎完全重叠(accuracy drop + 概念方向 cosine 旋转 across shift type,且 shift 类型 paraphrase/domain/length 与我们逐字一致);很可能也涉及 C4(多模型/规模)。未见 C2(无 forward-pass 先验预测分数)、未明确 C5(NLI 审计)、C3 仅部分(用了多种探针但未见系统的 mass-mean/LogReg/MLP 可迁移性对比)。
+- **Bao, Zhang et al. (2025)** — *Probing the Geometry of Truth: Consistency and Generalization of Truth Directions in LLMs Across Logical Transformations and Question Answering Tasks* (Findings of ACL 2025(arXiv:2506.00823);有公开代码) [威胁:high]
+  - C1 中等重叠(跨变换的方向一致性/泛化,但其变换是逻辑变换/QA,而非我们的 paraphrase/domain/length 分布偏移);C4 重叠(模型能力/规模越强方向越一致);C3 部分(讨论简单 vs 复杂探针是否必要)。未涉及 C2 的先验预测分数,未涉及 C5。
+- **作者待核实 (2026)** — *How Context Shapes Truth: Geometric Transformations of Statement-level Truth Representations in LLMs* (arXiv:2601.06599(2026-01)) [威胁:high]
+  - C1 中-高重叠(方向旋转角 θ 作为核心度量,与我们的概念方向 cosine 旋转同构,只是触发因素是上下文而非 paraphrase/domain/length);C4 高度重叠(明确给出 model size × 方向变化的结论)。未涉及 C2 的免标注先验稳定性分数,未涉及 C3、C5。
+- **Marks & Tegmark (2023/2024)** — *The Geometry of Truth: Emergent Linear Structure in LLM Representations of True/False Datasets* (arXiv:2310.06824;COLM 2024(待核实)) [威胁:high]
+  - C3 高度重叠(正是 mass-mean vs LogReg vs CCS 的可迁移性/鲁棒性比较的来源,且已声称 mass-mean 对分布偏移更鲁棒——我们的 C3 很大程度是在复刻+扩展);C4 部分(规模才出现线性结构);C1 背景(transfer 实验)。未涉及 C2、C5。
+- **Aithal, Kashyap et al. (2021)** — *Robustness to Augmentations as a Generalization Metric* (arXiv:2101.06459(NeurIPS 2020 Predicting Generalization in Deep Learning 竞赛亚军方案)) [威胁:medium]
+  - C2 中等重叠:augmentation-consistency 作为 label-free 泛化预测量的核心思想已被提出(但对象是分类网络整体泛化,而非探针方向在 OOD 的可迁移性,也无 dispersion 分量、无 LLM 概念方向语境)。是我们 C2「双分量分数」中 augmentation-consistency 一支的最直接先验思想。
+- **作者待核实 (2024)** — *Truth-value judgment in language models: 'truth directions' are context sensitive* (arXiv:2404.18865(2024)) [威胁:medium]
+  - C1 中等(为「探针/方向不稳定」提供早期证据,但未做 accuracy drop × 方向旋转的系统矩阵,也无 paraphrase/domain/length 三类偏移设计);C4 弱相关。未涉及 C2/C3/C5。
+- **作者待核实 (2025)** — *Are Sparse Autoencoders Useful? / SAE probes vs logistic regression baselines* (arXiv:2502.16681(2025)) [威胁:medium]
+  - C3 中等(探针方法系统比较的范式与基线选择,虽未含 mass-mean/MLP 的 OOD 专门对比);C1 弱相关。未涉及 C2、C4(规模)、C5。
+- **作者待核实 (2025)** — *The Impact of Off-Policy Training Data on Probe Generalisation* (arXiv:2511.17408(2025)) [威胁:low]
+  - C1 弱-中(探针泛化受分布影响,但聚焦训练数据 on/off-policy 而非输入端 label-preserving 偏移,也无方向旋转度量);未涉及 C2/C3/C4/C5。
+- **Park, Choe & Veitch (2023/2024)** — *The Linear Representation Hypothesis and the Geometry of Large Language Models* (ICML 2024(arXiv:2311.03658)) [威胁:low]
+  - 全为背景(C1 方向 cosine 度量的理论依据);不直接做探针稳定性/OOD 实验。无 C2-C5 重叠。
+
+> 评估: 已联网检索(WebSearch + WebFetch arXiv 全文)。诚实结论:本角度下我们的 novelty 处于高度拥挤区,C1、C3、C4 基本已被抢占,真正未被占领的护城河只剩 C2 的「双分量、免标注、训练前先验预测」整体打包,以及 C5 的方法学细节。
+
+最大撞车点(逐条):
+- C1 = 最危险。arXiv:2510.11905「LLM Knowledge is Brittle」几乎逐字做了我们 C1 要做的事:in-distribution 探针准确率高估真值特征鲁棒性,且在 paraphrase/domain/length(与我们三类偏移完全一致)下度量 accuracy drop + 概念方向 cosine 漂移。再叠加 2601.06599 把「方向旋转角 θ × model size」这条结论也做了。我们不能再把 C1 当作主贡献,最多作为复现+扩展,且必须显式区分于这两篇(例如:我们是 shift-type × model-size 的完整矩阵化、并把方向旋转与 accuracy drop 解耦量化)。务必注意 2510.11905 的「superficial resemblance」叙事会被审稿人直接拿来质疑 C1 的新颖性。
+- C3 = 已被 Marks & Tegmark(2310.06824)定调:mass-mean vs LogReg vs CCS、且已声称 mass-mean「对分布偏移更鲁棒、更因果」。我们的 C3 若只比 mass-mean/LogReg/MLP 的 OOD 可迁移性,基本是复刻;增量必须落在 MLP(非线性)这一未被充分比较的轴上,或与 C2 分数的相关性验证上,否则审稿人会判为已知结论。
+- C4 = 2506.00823 与 2601.06599 都已给出「模型越大/越强→方向越一致/靠方向变化区分」的规模结论;C4 单独不构成贡献,只能作为辅助维度。
+- C5(NLI label-fidelity 审计)= 检索未见同角度有人明确用本地 NLI(DeBERTa-MNLI)审计 shift 的 label-preserving 性;但这是方法学卫生措施,难以单独支撑 novelty。
+
+仍留白(可主打的差异化):
+1) C2 的整体打包是目前唯一未见先例的核心。检索到的最接近先验是 2101.06459(augmentation-robustness 作为 label-free 泛化度量),但那是「分类网络整体泛化」非「探针方向 OOD 可迁移性」,且没有 dispersion(同分布 K 次重采样方向离散度)这一分量、没有 LLM 概念方向语境。建议把论文重心明确转向 C2:强调它是 a priori、forward-pass-only、双分量、专门预测探针方向 OOD 迁移的分数,并实证它与 C1 观测到的 accuracy drop / 方向旋转高度相关(把 C1/C3/C4 降级为支撑 C2 的证据,而非并列贡献)。
+2) 必须补充与 2101.06459 的区分与(若可能)对照实验,否则 C2 的 augmentation-consistency 分量会被指为已知思想的迁移。
+3) dispersion 分量(重采样方向离散度作为稳定性预测量)在本角度检索中未见直接对应,是 C2 内最干净的新点,建议作为卖点之一并单独消融。
+
+未能确认的点(建议进一步精读全文):2510.11905 是否做了 mass-mean/LogReg/MLP 系统对比(WebFetch 显示
+
+### 角度: 概念擦除与方向估计:INLP / RLACE / LEACE / mass-mean(difference-of-means)/ steering vector...
+
+- **Marks & Tegmark, 2023** — *The Geometry of Truth: Emergent Linear Structure in LLM Representations of True/False Datasets* (COLM 2024 (arXiv:2310.06824)) [威胁:high]
+  - C3 直接重叠(mass-mean vs LogReg 的相对可迁移性正是它的核心卖点之一);C1 部分重叠(展示了探针跨数据集迁移 accuracy 的变化)。但它把 mass-mean 当作'更好'来推销,没有把跨分布 accuracy 高估当成系统性诊断问题,也无方向旋转矩阵、无 a priori 预测分数、无 NLI 审计。
+- **Anonymous/Hazratian & Zia (待核实作者), 2026** — *The Truthfulness Spectrum Hypothesis* (arXiv:2602.20273(2026-02,预印本,待核实正式 venue)) [威胁:high]
+  - C2 最强威胁:它正面解决了'预测探针 OOD 迁移性',且明确指出普通 cosine 不够、需白化——这会侵蚀你 C2 中'方向一致性/旋转'分量的新意。C4 也重叠(model size ladder)。关键差异(你的护城河):它的预测器需要在 OOD 数据上训练一个 reference probe 才能算 MCS,不是 forward-pass-only / a priori / 免新标注;它评估的是真值这一单一概念域、用 LogReg、无 paraphrase/length/domain 的 label-preserving shift 框架、无 NLI 审计、无 dispersion(K 次同分布重采样)分量。
+- **Ravfogel, Goldberg et al. (待核实完整作者), 2022** — *Probing Classifiers are Unreliable for Concept Removal and Detection* (NeurIPS 2022 (arXiv:2207.04153)) [威胁:medium]
+  - C1 的基石性背景:它确立了'探针准确率可被虚假相关误导、高估概念存在'这一核心命题。你 C1 的增量在于把它具体化为 label-preserving 分布偏移(paraphrase/domain/length)下的 accuracy drop × cosine 旋转矩阵 × model size,而该文不做分布偏移矩阵、不做方向旋转量化、不做预测分数。属于必须正面区分并引用的 prior art。
+- **Tan, Chanin et al. (待核实完整作者), 2024** — *Analysing the Generalisation and Reliability of Steering Vectors* (NeurIPS 2024 (arXiv:2407.12404)) [威胁:medium]
+  - C1/C3 中等重叠:steering vector = mass-mean 方向,该文已展示这类方向对输入改动的脆性(与你'稳定性被高估'同源),并触及可控性可被某些信号预测。差异:它面向 steering 干预效果(behavioural)而非探针分类 accuracy/方向旋转矩阵;无 a priori forward-pass 双分量分数;无 NLI label-fidelity 框架;模型规模非主轴。
+- **(待核实作者), 2025** — *Understanding (Un)Reliability of Steering Vectors in Language Models* (arXiv:2505.22637(2025,预印本,待核实 venue)) [威胁:medium]
+  - C2 中高威胁:它的'cosine 一致性预测有效性'+ d' 可分性指标,在思想上非常接近你 C2 的两个分量(方向一致性 + 离散度/可分性),且都是从激活几何 a priori 推断。差异:它针对 steering 有效性而非探针 OOD accuracy;无 dispersion(K 次同分布重采样)正式定义;无 paraphrase/domain/length 的 label-preserving shift 协议;无 NLI 审计;�� mass-mean/LogReg/MLP 三者系统比较;无 model size ladder。
+- **Méloux, Portet & Peyrard, 2025** — *Mechanistic Interpretability as Statistical Estimation: A Variance Analysis* (arXiv:2510.00845(2025,预印本,待核实 venue)) [威胁:medium]
+  - C2 dispersion 分量的概念前身:'用 K 次重采样的方向离散度衡量探针稳定性'这一核心机制与你高度同构。差异(待全文确认):它把 dispersion 当作'可信度/可复现性'诊断,未必把它接到'a priori 预测 OOD 迁移 accuracy'这一具体目标上,且未必含 augmentation-consistency 第二分量、label-preserving shift 协议与 NLI 审计。建议优先精读全文判定是否需降为护城河外或升为 high。
+- **Azizian, Kirchhof, Ndiaye, Bethune, Klein, Ablin, Cuturi, 2025** — *The Geometries of Truth Are Orthogonal Across Tasks* (arXiv:2506.08572(2025,预印本,待核实 venue)) [威胁:medium]
+  - C1/C3 重叠:它正是'方向旋转/正交 ↔ accuracy 掉落'的量化,与你 C1 矩阵中'cosine 旋转'一列同源,并涉及多探针类型。差异:它跨任务(task)而非你强调的 label-preserving 同任务输入分布偏移(paraphrase/domain/length);无 forward-pass a priori 预测分数;无 NLI 审计;无 model size ladder 为主轴。
+- **Ravfogel, Twiton, Goldberg, Cotterell (待核实), 2022** — *Linear Adversarial Concept Erasure (RLACE)* (ICML 2022 (arXiv:2201.12091)) [威胁:low]
+  - 背景相关:为你'方向估计本身不稳定'提供方法学锚点(RLACE 优化不稳),但它不研究 label-preserving 分布偏移下探针 accuracy 高估,不提预测分数,不做 paraphrase/NLI。属应引用的方向估计可靠性脉络,但不构成 novelty 撞车。
+- **Belrose, Schneider-Joseph, Ravfogel, Cotterell, Raff, Biderman, 2023** — *LEACE: Perfect Linear Concept Erasure in Closed Form* (NeurIPS 2023 (arXiv:2306.03819)) [威胁:low]
+  - 背景相关:concept erasure 角度的 SOTA 锚点,但聚焦'擦除得多干净'而非'探针方向在 label-preserving 偏移下是否稳定/可迁移',与 C1–C5 无直接方法撞车。
+- **Ravfogel, Elazar, Goldberg, Cotterell (待核实), 2020** — *Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection (INLP)* (ACL 2020 (arXiv:2004.07667)) [威胁:low]
+  - 背景相关:与 RLACE 一道支撑'线性方向估计本身脆弱'的论点,但不涉及分布偏移下的探针 accuracy 高估、预测分数、paraphrase/NLI。
+
+> 评估: 已联网检索(WebSearch/WebFetch)。重要可靠性提示:WebFetch 的小模型摘要出现过明显的"附和我的检查清单全中"幻觉——我已对每篇下载 PDF 抽样核验,确认 arXiv:2602.03951(TorRicc)其实是 CIFAR 视觉模型 + 谱复杂度/Ricci 曲率做无监督 checkpoint 选择,与本角度几乎无关(故未列入,顶多算 label-free OOD 预测哲学的远亲);2510.00845 与 2602.20273 的"全中"部分也按怀疑态度下调或标注待核实。
+
+最大撞车点(按威胁排序):
+1) C2(a priori 预测探针 OOD 迁移性)是你被抢占风险最高的贡献。arXiv:2602.20273(Truthfulness Spectrum, 2026-02)已用 Mahalanobis-白化后的方向相似度近完美预测 OOD AUROC(R²≈0.98),并明确说普通 cosine 不够——这几乎正面命中你"用方向一致性预测迁移"的卖点,且跨 model size。你唯一且关键的护城河是:它的预测器需要在 OOD 数据上额外训练 reference probe(需 OOD 标注),而你主张 forward-pass-only / 免新标注 / 训练前 a priori,且含 dispersion(同分布 K 次重采样)这一不需要任何 shift 数据的分量。务必在论文里把"是否需要触碰目标分布/新标注"作为核心区分轴写死,并直接对比 MCS。
+2) C3(mass-mean vs LogReg vs MLP 的 OOD 相对可迁移性)已被 Marks & Tegmark(2023)实质性占据 mass-mean vs LogReg 这一半;你的增量只剩"加入 MLP + 在系统化 label-preserving shift 矩阵下比较"。建议把 C3 从"比较"重新定位为"在你的 shift × size 矩阵框架内的受控复现+扩展",否则单独作为贡献偏弱。
+3) C1(accuracy 高估 + 方向旋转矩阵)在'探针不可靠/会高估'(NeurIPS'22)与'方向跨任务正交 ↔ 迁移失败'(2506.08572)之间被两头夹。但没有一篇把它组织成 shift type(paraphrase/domain/length)× model size 的 accuracy-drop × cosine-rotation 双矩阵,这个具体形态仍可主张为新颖呈现。
+4) C2 的 dispersion 分量思想已有前身(2510.00845 把方向重采样离散度当可靠性指标;2505.22637 用激活差 cosine 一致性 + d' 预测 steering 有效性)。"双分量(dispersion + augmentation-consistency)合成一个分数、专门用于训练前预测探针 OOD"这个组合目前未见完全重合者,是你最干净的立足点。
+
+仍存的空白(= 你可主张的真正 novelty):
+- forward-pass-only / 免任何 OOD 数据与新标注 的 a priori 预测(vs 2602.20273 需 OOD reference probe);
+- dispersion + augmentation-consistency 双分量的显式合成与互补性消融;
+- label-preserving 的 paraphrase/domain/length shift 受控协议,尤其 C5
+
+### 角度: 探针在分布偏移/OOD 下的泛化与失效(probing under distribution shift; robustness of probing clas...
+
+- **Xie, Wei, Feng, Cao, An (2023)** — *On the Importance of Feature Separability in Predicting Out-Of-Distribution Error (Dispersion Score)* (NeurIPS 2023(arXiv:2303.15488,待核实最终录用会议)) [威胁:high]
+  - 对 C2 威胁最大:与我们 PSS 的第一个分量(dispersion/方向离散度)在名称、机制(feature dispersion)、目标(免标注+仅前向预测 OOD 性能)上几乎正面撞车。差异:他们用于一般图像分类器的整体精度估计,不是 concept-direction/线性探针,不含 augmentation-consistency 第二分量,也不做 a priori(他们是在给定 OOD 测试集上估精度,我们是训练前预测探针迁移性)。
+- **Levinstein & Herrmann / Bürger 等 (2024)（'truth directions are context sensitive'）** — *Truth-value judgment in language models: 'truth directions' are context sensitive* (arXiv:2404.18865(待核实正式 venue)) [威胁:high]
+  - 对 C1 威胁高:正是'探针准确率高估了概念方向稳定性、在 label-preserving 上下文/改述偏移下方向漂移'这一核心观察。差异:聚焦真值方向单一概念、未给出 accuracy-drop × direction-rotation 跨 shift×模型规模 的系统矩阵,也未提出预测性 score(不碰 C2)。
+- **Azizian, Kirchhof, Ndiaye, Bethune, Klein, Ablin, Cuturi (2025)** — *The Geometries of Truth Are Orthogonal Across Tasks* (arXiv:2506.08572(待核实)) [威胁:high]
+  - 对 C1/C3 中等偏高:用方向几何(正交=cosine 旋转)量化探针跨分布不可迁移,与 C1 的'方向旋转'指标同源。差异:跨任务而非 label-preserving 输入偏移(paraphrase/domain/length),且无免标注预测 score、无 mass-mean/LogReg/MLP 比较、无规模阶梯。
+- **Marks & Tegmark (2023)** — *The Geometry of Truth: Emergent Linear Structure in LLM Representations of True/False Datasets* (arXiv:2310.06824(后续 COLM/会议版本待核实)) [威胁:high]
+  - 对 C3 威胁高:已直接做了 mass-mean vs LogReg 的跨数据集(可视为 OOD)相对迁移性比较,且结论是 difference-of-means 更稳。我们的 C3 若只停在'mass-mean vs LogReg vs MLP 谁更稳'会被认为增量;需靠 MLP 纳入 + 系统化 shift 类型 + 与 PSS 预测对齐来区分。差异:无 MLP,无 paraphrase/length 偏移设计,无预测 score。
+- **多篇 (2025-2026)：'generalization of LLM truth directions on conversational formats'(2505.09807)、'Probing the Geometry of Truth: Consistency and Generalization across Logical Transformations'(2506.00823)、'Testing the Limits of Truth Directions'(2604.03754)** — *Truth-direction generalization across formats / logical transformations 系列* (arXiv(2025-2026,均待核实)) [威胁:medium]
+  - 对 C1/C4 中等:覆盖了 length/format 这类 label-preserving 偏移并量化泛化,与 C1 的 length shift、与 C4 的'结构变化下稳定性'有重叠。差异:仍局限于真值任务,无统一的 accuracy-drop×rotation 矩阵,无免标注 a priori score,规模阶梯不是主轴。
+- **Müller 等 / 待核实 (2025)** — *Intermediate Layer Classifiers for OOD Generalization (ICLR 2025)* (ICLR 2025(proceedings.iclr.cc,待核实作者)) [威胁:medium]
+  - 对 C1/C4 中等:直接研究线性探针 OOD 泛化、层与规模因素,且量化 accuracy 在 shift 下的保持率。差异:视觉/表征学习设定为主,概念方向旋转、免标注预测 score、NLI label-fidelity 审计均不涉及。
+- **Hewitt & Liang (2019)** — *Designing and Interpreting Probes with Control Tasks (selectivity)* (EMNLP 2019) [威胁:low]
+  - 对 C1/C2 低-中:是'探针准确率会误导、需要额外诊断量'这一论点的奠基性背景,与我们 PSS 作为'诊断量'同属一脉,但 selectivity 针对记忆/复杂度而非分布偏移稳定性,且非免标注预测 OOD 迁移。
+- **Belinkov (2022)** — *Probing Classifiers: Promises, Shortcomings, and Advances* (Computational Linguistics 48(1)) [威胁:low]
+  - 对 C1 低:为我们的动机(探针指标不可尽信)提供权威背景,但不提供任何偏移矩阵、score 或方法比较,属定位/引用价值。
+- **Ravichander, Belinkov, Hovy (2021)** — *Probing the Probing Paradigm: Does Probing Accuracy Entail Task Relevance?* (EACL 2021) [威胁:low]
+  - 对 C1 低:支撑'准确率高估特征有用性/稳定性'的总论点,但聚焦任务相关性而非分布偏移下的方向稳定性,无 score 提案。
+
+> 评估: 已联网检索。本角度下你们最大的撞车点有两处,均需在论文中正面处理:
+
+1) C2(免标注、仅前向的双分量 PSS)与 Xie et al. 2023 的 "Dispersion score" 高度同名同质——同样是 label-free、forward-pass、用 feature dispersion 预测分布偏移下性能。区分点必须明确且可被审稿人一眼看到:(a) 我们是对'线性/概念探针的方向稳定性'而非'整体分类精度'打分;(b) 我们是训练前 a priori 预测探针迁移性,而非在给定 OOD 测试集上估精度;(c) 我们加了第二个正交分量 augmentation-consistency,且需用消融证明它在 dispersion 之外提供增量。否则 C2 会被定性为 Dispersion score 的换皮迁移。建议直接把 Dispersion score 作为 baseline 跑对比。
+
+2) C1(准确率高估方向稳定性 + accuracy-drop×cosine-rotation 矩阵)的核心现象已被 "truth directions" 系列(Levinstein/Bürger 2024、Azizian et al. 2025 正交性、conversational-format/logical-transformation 系列 2025-2026)在真值方向这一特例上反复证明,包括用方向旋转/正交性量化。你们的真正新意只能落在'系统性矩阵'(shift type × model size × probe type)+ 跨多概念(不止真值)+ 与可预测 score 挂钩,而非'发现方向会漂移'这一已知事实。C3(mass-mean vs LogReg)也已被 Marks & Tegmark 2023 部分做掉,纯比较是增量;靠纳入 MLP、统一 shift 设计、并把比较结果用 PSS 预测出来才有区分度。
+
+仍存在的空白(你们可主打的护城河):没有任何一篇把(i)免标注 a priori 探针稳定性 score、(ii)dispersion+augmentation-consistency 双分量、(iii)label-preserving 偏移的 accuracy-drop×direction-rotation 系统矩阵、(iv)mass-mean/LogReg/MLP 三探针×规模阶梯、(v)本地 NLI(DeBERTa-MNLI)做 shift 的 label-fidelity 审计——五者整合在统一框架里。尤其 C5(NLI 审计保证 shift label-preserving)在本角度文献中几乎无人配套使用,是相对干净的方法学贡献点,但单独 novelty 不足以支撑论文,应作为严谨性卖点而非核心 claim。
+
+诚实结论:C1、C2、C3 单独看都已被相邻工作不同程度抢占(C2 最危险,直接撞 Dispersion score 之名与机制),论文的可辩护新颖性集中在'a priori 预测 + 双分量 + 系统矩阵 + NLI 审计'的整体打包与'方向稳定性视角的 score 化'上,务必在 related work 显式区分上述四篇 high-threat 工作并用实验(Dispersion score、mass-mean 作 baseline)证明增量。
+
+### 角度: 表示/方向的稳定性与旋转:线性探针方向在 seed/重采样/扰动/分布偏移下的不稳定性与旋转(cosine 漂移)、direction consistency、...
+
+- **Gao, Zhu, Zeng, Zhao, Wang, Ruan, Ding — 2026** — *RAPTOR: Ridge-Adaptive Logistic Probes* (arXiv:2602.00158 (preprint;投稿目标未知,待核实)) [威胁:high]
+  - 与 C2 的 dispersion 分量高度重叠(其 eq.9 几乎就是我们的'同分布重采样方向离散度',只是用 1−|cos| 还是 |cos| 的符号约定差异);与 C4(model size ladder 影响方向稳定性)直接重叠;与 C3 部分重叠(比较不同探针估计子的方向稳定性,但不是 mass-mean/LogReg/MLP 这条具体轴线);与 C1 在动机表述上重叠(accuracy 高估稳定性)。缺口:没有 augmentation-consistency 分量、不做 label-preserving 输入偏移(paraphrase/length/domain)、不构建 accuracy-drop×cosine-rotation 矩阵、不把分数当作训练前 a priori 的 OOD 迁移预测器(它的稳定性是诊断而非预测 OOD)。
+- **Lysnæs-Larsen, Eggen, Strümke (NTNU) — 2025** — *Probing the Probes: Methods and Metrics for Concept Alignment* (arXiv:2511.04312 (待核实是否已被会议接收)) [威胁:high]
+  - 与 C2 的 augmentation-consistency 分量高度重叠:同样是'label-preserving 增广后方向一致性、forward-pass-only、归一化到 0–1'的探针质量分数。区别:这是视觉/ResNet50/CAV 场景,增广是图像变换而非文本 paraphrase/length/domain;不含同分布重采样 dispersion 分量;不把两分量合成单一 Probe Stability Score;不预测 OOD 迁移;不比较 mass-mean/LogReg/MLP;不变 model size。
+- **Huang (William Hao-Cheng) — 2025** — *Spectral Identifiability for Interpretable Probe Geometry* (arXiv:2511.16288 (preprint,待核实)) [威胁:high]
+  - 与 C2 的核心动机/定位撞车最危险的一篇:同样是'在训练/部署前 a priori 预测探针方向可靠性'的免标注判据。但机制不同——它用 spectral/eigengap 几何而非 dispersion+augmentation 双分量;聚焦 in-distribution 可识别性而非显式预测 OOD/label-preserving shift 迁移;不构建 accuracy-drop×rotation 矩阵;不比较探针族;model size 未系统化。属'同一空白被另一种方法占住',对 C2 的'首个 a priori 探针稳定性预测器'新颖性主张构成实质威胁。
+- **Marks & Tegmark — 2023/2024** — *The Geometry of Truth: Emergent Linear Structure in LLM Representations of True/False Datasets* (COLM 2024(待核实);arXiv:2310.06824) [威胁:high]
+  - 与 C3 直接、强重叠:正是'mass-mean vs LogReg(+CCS)在跨数据集/OOD 下相对可迁移性'的系统比较,且给出 mass-mean 更可迁移的结论——这会削弱 C3 的新颖性,需明确我们相对它的增量(加入 MLP、label-preserving 输入偏移、规模 ladder、以及把可迁移性与稳定性分数挂钩)。与 C4 部分相关(它有 scale 讨论)。不覆盖 C2 的稳定性分数,也不做 paraphrase/length 偏移。
+- **Bürger / Levinstein 等(Truth-value judgment)— 2024,及 Probing the Geometry of Truth — 2025** — *Truth-value judgment in language models: 'truth directions' are context sensitive (2404.18865);Probing the Geometry of Truth: Consistency and Generalization of Truth Directions across Logical Transformations and QA (2506.00823)* (arXiv:2404.18865 / arXiv:2506.00823(会议状态待核实)) [威胁:high]
+  - 与 C1 核心现象强重叠:同时记录了'准确率下降'与'方向旋转/近正交'这两件事,正是我们 C1 想系统化的 accuracy-drop + direction-rotation。区别:它们的偏移是逻辑/上下文/否定与跨域,而非我们刻意构造的 label-preserving paraphrase/length/domain 矩阵,也未把它组织成 shift-type × model-size 的二维矩阵,且未提出稳定性分数。仍显著压缩 C1 '系统量化高估'的新颖空间。
+- **Kramár, Engels, Wang, Chughtai, Shah, Nanda, Conmy (Google DeepMind)— 2026** — *Building Production-Ready Probes for Gemini* (arXiv:2601.11516 (cs.LG, preprint)) [威胁:medium]
+  - 与 C1 在'准确率高估稳定性 / 输入长度偏移下崩溃'上重叠,且是 length-shift 的高质量证据点。但偏工程/安全部署视角,不量化方向 cosine 旋转、不做 paraphrase/domain 矩阵、不提稳定性分数、不比较探针族、不做规模 ladder。主要威胁 C1 的'长度偏移导致高估'这一子论点的首发性。
+- **Ding & Koehn 等(On Model Stability as a Function of Random Seed)— 2019** — *On Model Stability as a Function of Random Seed* (CoNLL 2019 (aclanthology K19-1087)) [威胁:low]
+  - 背景层面支撑 C2 的 dispersion 动机(同条件多次训练方向/性能会漂移),但针对的是任务模型而非线性探针方向、不涉及 cosine 旋转、不预测 OOD、不提稳定性分数。仅作 related work 引用。
+- **Ding, Denain & Steinhardt(Grounding Representation Similarity with Statistical Testing);Davari 等(Reliability of CKA, ICLR 2023);Deconfounded CKA — 2021–2023** — *Grounding Representation Similarity with Statistical Testing / Reliability of CKA as a Similarity Measure / Deconfounded Representation Similarity* (NeurIPS 2021 / ICLR 2023 / 2022(arXiv:2108.01661, 2202.00095 等)) [威胁:low]
+  - 方法论背景:为'用相似度/方向一致性衡量表示稳定性'提供工具与告诫,与 C2 用 cosine 度量方向一致性的思路同源。但研究对象是整层表示相似度而非单个探针概念方向、不针对 OOD 探针迁移预测、不涉及 label-preserving 文本偏移。低威胁,但建议引用以表明我们知道相似度度量的可靠性陷阱。
+- **Ravichander, Belinkov, Hovy 等(probing 方法论批评线)— 2021–2022** — *Probing-Classifier 方法论批评(Probing the Probing / Belinkov 2022 survey 等),及 Probing Classifiers are Unreliable for Concept Removal and Detection (NeurIPS 2022)* (NeurIPS 2022 / CL 2022 等(待核实具体条目)) [威胁:medium]
+  - 与 C1 的'准确率高估特征稳定性/存在性'动机重叠(高估的是 presence/causal,我们强调的是 across-shift stability),并被 RAPTOR 直接引用为'accuracy≠可靠方向'的依据。属强背景:压缩 C1 '高估'论点的概念新颖性,但未做 shift×size 矩阵、无旋转量化、无稳定性分数。
+
+> 评估: 已联网检索(WebSearch/WebFetch + 本地读取 RAPTOR PDF),结论可信。
+
+最大撞车点(按威胁排序):
+1) C2 的两个分量已被分别占住,且各自都强。RAPTOR(2602.00158)的 directional-stability 度量(eq.9,K 次 ablation 后方向 mean |cos|)几乎就是我们的 dispersion 分量;"Probing the Probes"(2511.04312)的 Augmentation Robustness(label-preserving 变换、forward-pass-only、0–1)几乎就是我们的 augmentation-consistency 分量。也就是说 C2 的"双分量"里每一个分量单独看都已有人做过,真正可主张的新颖性只剩:(a) 把两者合成单一 Probe Stability Score、(b) 把它定位为训练前 a priori 的 OOD 迁移预测器(而非事后诊断)、(c) 用在文本 LLM 的 paraphrase/length/domain label-preserving 偏移上。
+
+2) C2 的"a priori 预测探针可靠性"定位本身也已被 Spectral Identifiability(2511.16288)抢占——它同样是免标注、部署前可验证的判据,只是用 eigengap 几何而非我们的双分量。这对"首个 a priori 探针稳定性预测器"的措辞最危险,必须改写主张、并做正面 baseline 对比(我们的分数 vs SIP vs RAPTOR-stability 谁更能预测真实 OOD drop)。
+
+3) C1 的核心现象(准确率下降 + 方向旋转)在 truth-direction 文献(2404.18865、2506.00823)里已被同时记录(约 25% OOD drop、跨域方向近正交),length 偏移崩溃在 Gemini probes(2601.11516)里也有强证据。C1 仍可主张的是"系统化的 shift-type × model-size 二维矩阵 + 同一概念跨多种 label-preserving 偏移的受控量化",而非"首次发现高估"。
+
+4) C3 与 Marks & Tegmark(Geometry of Truth)直接撞:mass-mean vs LR(+CCS)的跨数据集可迁移性比较已有,且结论是 mass-mean 更可迁移。我们的增量必须明确为:加入 MLP 探针、用受控 label-preserving 输入偏移而非换数据集、跨规模 ladder、并把可迁移性与我们的稳定性分数关联。
+
+仍存的空白(可成为真正卖点):
+- 没有任何一篇把 dispersion + augmentation-consistency 合成单一分数,并实证证明该分数能在训练前预测 label-preserving OOD 下的 accuracy-drop 与 cosine-rotation(即把稳定性分数当 OOD 迁移的预测器并报告相关性/排序能力)。这是 C2 最干净的立足点。
+- 没有一篇把 (accuracy drop × concept-direction cosine rotation) 组织成 shift-type(paraphrase/domain/length)× model-size 的完整矩阵(C1 的呈现形式仍是空白)。
+- mass-mean vs LogReg vs
+
+### 角度: 可迁移性的免训练事前预测(transferability estimation: LEEP/LogME/H-score/GBC 及其在 probe 方向上的类比...
+
+- **Nguyen, Hassner, Seeger, Archambeau (2020)** — *LEEP: A New Measure to Evaluate Transferability of Learned Representations* (ICML 2020) [威胁:medium]
+  - 与 C2 的方法论范式(免训练 + 仅 forward-pass 预测可迁移性)直接对齐,是 C2 明确自认的类比来源。但目标完全不同:LEEP 预测的是【整模型迁移/微调后准确率】,需要源标签头/伪标签,且不涉及探针方向(direction)的稳定性,更不涉及 label-preserving 分布偏移下的方向旋转。C2 的新意正是把这一范式迁移到 probe 方向几何上并提出 dispersion+augmentation-consistency 两个新分量。
+- **You, Liu, Wang, Long (2021)** — *LogME: Practical Assessment of Pre-trained Models for Transfer Learning* (ICML 2021) [威胁:medium]
+  - 与 C2 范式同源(免训练打分预测下游可用性),且 LogME 本质就是在特征上做线性证据评估,与 probe 场景天然接近,因此是范式撞车点。差异:LogME 用【目标任务标签】拟合,属于事后/需要目标标注的估计,预测对象是 in-distribution 下游准确率,而非 OOD 下方向稳定性;不做重采样离散度,也不做增广一致性。C2 的『完全免新标注 + 专测 OOD 方向稳定』是区分点。
+- **Bao, Li, Huang, et al. (2019)** — *An Information-Theoretic Approach to Transferability in Task Transfer Learning (H-score)* (ICIP 2019) [威胁:low]
+  - 与 C2 共享『用特征几何/统计量免训练预测迁移』的思路。但 H-score 仍是任务级、需目标标签的可迁移性度量,与 probe 方向稳定性、label-preserving 输入偏移无关。属于范式背景而非具体抢占。
+- **Zhang, Wang, Wang, Chai, Yin, Lin, Wang (2026)** — *SAE as a Crystal Ball: Interpretable Features Predict Cross-domain Transferability of LLMs without Training* (arXiv 2603.02908(预印本,待核实最终去向)) [威胁:high]
+  - 这是本角度下时间最近、立意最接近 C2 的工作:同样主张【免训练 + 仅 forward-pass + 训练/微调前 a priori 预测 LLM 跨域(OOD)迁移性】。撞车点在于『事前可迁移性预测』这一核心卖点的 LLM 表征版本已被占。但机制不同——它用 SAE 可解释特征维度的偏移,而非 C2 的『同分布重采样方向离散度 + label-preserving 增广方向一致性』;预测对象是 SFT 后整体迁移,而非线性探针方向几何的 OOD 稳定性。C2 仍有方法差异,但 novelty 表述需明显避让此文。
+- **(作者待核实) (2025)** — *Feature Space Perturbation: A Panacea to Enhanced Transferability Estimation* (arXiv 2502.16471(待核实)) [威胁:medium]
+  - 与 C2 的第二分量(augmentation-consistency:扰动/增广后方向一致性)在直觉上最接近——都把『对扰动的鲁棒性』当作可迁移性信号。但它扰动的是特征空间用于模型选择评分,不是 label-preserving 输入增广,也不针对 probe 方向的旋转或 OOD 稳定性。是 augmentation-consistency 思想的近邻先例,需在 related work 中正面区分。
+- **Kirch, Dower, Skapars, Yannakoudakis, Lubana, Krasheninnikov (2025)** — *The Impact of Off-Policy Training Data on Probe Generalisation* (arXiv 2511.17408(待核实)) [威胁:medium]
+  - 与 C1(探针在分布偏移下泛化失败)主题高度相关,也触及『探针 OOD 迁移性』这一对象。但它是【诊断性/因果研究】(探究什么影响泛化),而非提出免训练的 a priori 预测分数,因此不抢 C2 的预测器贡献;对 C1 构成部分主题重叠(探针 OOD 不稳定的现象记录),但没有 C1 的 shift type × model size 矩阵 + 方向 cosine 旋转量化。
+- **Lee, Lee, Shin, et al. (2021)** — *Improving Transferability of Representations via Augmentation-Aware Self-Supervision (AugSelf)* (NeurIPS 2021) [威胁:low]
+  - 与 C2 augmentation-consistency 分量共享『增广与可迁移性挂钩』的底层假设,提供了『增广不变性≠总是好』的重要注意点(对 C2 增广一致性作为稳定性代理的有效性边界有警示意义)。但它是表征学习方法,不预测 probe 方向稳定性。属背景与潜在 limitation 讨论来源。
+
+> 评估: 已联网检索(WebSearch/WebFetch 可用)。
+
+【本角度最大撞车点】两层:
+1) 范式层(medium):LEEP/LogME/H-score/GBC 早已确立『免训练、仅前向、用特征统计量事前预测可迁移性』这一完整范式。C2 不能宣称『免��注/免训练事前预测可迁移性』本身为新——必须把 C2 明确定位为『该范式在 linear probe 方向几何 + label-preserving 输入偏移下的 OOD 专用类比』,并强调新的预测量(方向离散度、增广方向一致性)和新的预测目标(方向稳定性而非微调准确率)。
+
+2) 最近的具体抢占(high):Zhang et al. 2026《SAE as a Crystal Ball》是威胁最大的一篇——它已经在 LLM 表征上主张『免训练、仅 forward-pass、训练前 a priori 预测跨域(OOD)迁移性』,并给出强相关性证据。C2 与它的核心卖点高度撞句。差异点(C2 用 probe 方向的重采样离散度+增广一致性,目标是 probe 方向几何稳定性;该文用 SAE 特征维度偏移,目标是 SFT 整体迁移)是可辩护的,但论文 novelty 句必须显式引用并区分此文,否则审稿人极可能据此判定 incremental。建议把 C2 的卖点收窄为『首个针对 linear probe 方向(direction)在 label-preserving 偏移下 OOD 稳定性的免标注前向预测分数,且双分量分别捕捉 epistemic(重采样离散度)与 invariance(增广一致性)两类不稳定来源』。
+
+【留白/可主张的空白】(a) 现有 transferability estimators 全部预测『迁移后准确率/性能』,无一预测『概念方向的几何稳定性(cosine 旋转)』——C2 的预测目标是真空地带;(b) 没有工作把『同分布 K 次重采样方向离散度』作为 probe 稳定性的事前信号——dispersion 分量看起来确属新;(c) 没有工作把 label-preserving 增广方向一致性 + NLI 标签保真审计(C5)接到 transferability 预测上;(d) 把 mass-mean/LogReg/MLP 探针类型(C3)与 model size(C4)纳入同一可迁移性预测框架,在本角度文献中未见。
+
+【行动建议】Related Work 必须正面引用并区分:LEEP、LogME、H-score(范式),Zhang 2026 SAE Crystal Ball(最近 LLM-OOD 事前预测,重点避让),Feature Space Perturbation 2502.16471(扰动一致性的近邻),Kirch 2025 off-policy probe generalisation(探针 OOD 现象,支撑 C1 但需声明其非预测器)。最大风险是 C2 与 SAE Crystal Ball 的卖点表述撞车,务必收窄措辞。注:部分 arXiv 编号(2603.x/2602.x/2601.x)和作者列表为检索片段所得,投稿前需逐条人工核实标题/作者/venue。
+
+### 角度: label-preserving 文本增广(back-translation / EDA / contrast sets / counterfactual au...
+
+- **Lysnæs-Larsen, Eggen & Strümke, 2025** — *Probing the Probes: Methods and Metrics for Concept Alignment* (arXiv preprint (cs.AI), 待核实是否已投会议) [威胁:high]
+  - 对 C1 与 C2 构成最直接的概念性撞车:C1 的核心论点'准确率高估特征稳定性'与该文'准确率不可靠'几乎同义;C2 的 augmentation-consistency 分量与其'augmentation robustness'度量同源。差异:该文偏视觉/spatial 概念定位,未做文本 paraphrase/domain/length 的 shift 矩阵,未提出免标注的 a priori OOD 迁移性'预测分数',也未含 dispersion 分量。
+- **Kaushik, Hovy & Lipton, 2020 (arXiv 2019)** — *Learning the Difference that Makes a Difference with Counterfactually-Augmented Data* (ICLR 2020) [威胁:medium]
+  - 为 C1 提供核心思想基础:分类器对 label-preserving 编辑敏感、依赖虚假特征。但其对象是端到端分类器而非探针的概念方向,且不量化 cosine 旋转、不提稳定性分数,也无 NLI 审计。属强背景威胁而非方法抢占。
+- **Gardner et al., 2020** — *Evaluating Models Local Decision Boundaries via Contrast Sets* (Findings of EMNLP 2020) [威胁:medium]
+  - 与 C1 的'准确率高估'叙事高度同构(标准评测 vs 扰动评测的 gap)。差异:面向端到端模型而非探针/概念方向,扰动多为改标签的对比而非纯 label-preserving 分布偏移,无 forward-pass 稳定性分数、无概念旋转度量。
+- **Wei & Zou, 2019** — *EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks* (EMNLP-IJCNLP 2019) [威胁:low]
+  - 提供 C2 增广分量与 C5 审计动机的工具与'增广未必真 label-preserving'的证据。本身不涉及探针稳定性或 NLI 审计,属背景。
+- **多篇 (e.g. Falsesum 2022; On-the-fly Denoising 2022; DISCO 2023; A Synthetic Data Approach for NLI 2024)** — *NLI/entailment 作为增广质量过滤器(overgenerate-and-filter / roundtrip consistency)* (ACL/EMNLP/NAACL 系列, 各篇待核实) [威胁:high]
+  - 对 C5 构成方法层面的强威胁——'用 NLI 核验 label 是否被保留'已是公认做法,DeBERTa-MNLI 做 fidelity 审计难算新方法,只能算工程实例化。C5 应降级为方法学保障而非核心新颖性主张。
+- **待核实 (检索命中, OOD probing transferability)** — *关于 probing 在 distribution shift 下迁移性的研究(binary vs multiclass probe OOD F1 显著下降)* (待核实) [威胁:medium]
+  - 直接触及 C1 的实证核心(探针 OOD 掉点)并部分触及 C3(探针类型对比)。需进一步核实该文是否已给出按 shift type × model size 的矩阵;若有则 C1/C3 威胁升至 high��当前证据为 medium。
+- **Sen et al. / Huang et al. 等, 2023 (待核实)** — *Counterfactually-Augmented SNLI Training Data Does Not Yield Better Generalization* (openreview / EMNLP 系列, 待核实) [威胁:low]
+  - 与 C1/C5 相关:佐证'增广的 label 质量与稳定性需被审计'的动机,但不提出探针稳定性分数,属支持性背景。
+
+> 评估: 未联网核验全部 venue 的部分以"待核实"标注;以下基于本轮真实检索(WebSearch)结果与既有知识。
+
+最大撞车点(按威胁排序):
+1) C5 基本已被抢占。"用 NLI/entailment 模型核验增广/paraphrase 是否 label-preserving、不一致就过滤"是 2022–2024 的成熟范式(overgenerate-and-filter、roundtrip consistency、DeBERTa-MNLI 判同义)。把它作为核心新颖性会被审稿人直接驳回。建议把 C5 明确定位为"方法学保障/sanity check",而非贡献点,并主动引用这些过滤工作以示知情。
+
+2) C1+C2 的概念框架面临 "Probing the Probes" (2025) 的正面威胁。该文已独立提出:探针准确率不可靠、需用 "augmentation robustness" 度量概念对齐。这与你们"准确率高估稳定性"(C1)和"augmentation-consistency 分量"(C2)在叙事和度量上同源。务必把它作为最近邻工作正面对比,凸显差异:(a) 你们在文本 NLP 上系统化 paraphrase/domain/length × model size 的偏移矩阵 + 概念方向 cosine 旋转;(b) 你们的 dispersion 分量(同分布 K 次重采样离散度)它没有;(c) 关键差异化——你们做的是 a priori、免新标注、训练前预测探针 OOD 迁移性的分数,而它停留在事后诊断度量。这条"预测性 + 双分量(dispersion ⊕ augmentation-consistency)"是目前看最可能站得住的真新颖点。
+
+3) C1 的"准确率高估真实能力"叙事不是新发现:contrast sets (Gardner 2020)、counterfactually-augmented data (Kaushik 2020) 早已确立同构论点,且检索到的 OOD probing transferability 工作已实证探针 OOD 掉点。你们的增量必须落在"专门针对线性探针的概念方向、用 cosine 旋转量化、并跨 shift type × scale 成矩阵",而非"发现了掉点"本身。
+
+仍存的空白(可主打的真新颖性):
+- 双分量、forward-pass-only、免新标注、训练前 a priori 预测探针 OOD 迁移性的"分数"——未见同形态工作;dispersion(同分布重采样方向离散度)作为稳定性信号在文本探针语境下尚属空白。
+- "accuracy drop 与概念方向 cosine 旋转解耦"的矩阵化呈现(同一 shift 下准确率稳但方向已旋转)——未见直接对应工作,是较强卖点。
+- C3(mass-mean vs LogReg vs MLP 在 OOD 下相对可迁移性的系统比较)与 C4(model size ladder 对探针 OOD 稳定性影响)在本角度检索中未见被系统覆盖,威胁较低,可作为扎实的实证贡献,但需另行核查"OOD probing transferability"系列是否已部分触及 C3/C4。
+
+诚实结论:C5 应放弃作为新颖性、改为方法保障;C1 需从"现象发现"转为"探针概念方向的量化方式";C2 是核心命脉,但必须在 "Probing the Probes" 之上明确切出"预测性 + dispersion 分量"的差异,否则有被判为增量的风险。
+
+### 角度: 小语言模型/Pythia/GPT-2 上的可解释性与探针 size-ladder 研究(scaling of representations、Pythia in...
+
+- **Reblitz-Richardson, 2026** — *When Probing Accuracy Saturates, Fragility Resolves: A Complementary Metric for LLM Pre-Training Analysis* (arXiv 2606.11375 (cs.CL),待核实是否已投会议) [威胁:high]
+  - 对 C2 与 C1 威胁最大。C2:它同样是 forward-pass、免新标注、对‘准确率会高估稳定性’做补救的探针稳定性度量,与你的 dispersion 分量(对扰动的离散度/崩溃点)精神高度重合,只是它用激活噪声而非 label-preserving 增广/重采样。C1:与你‘准确率高估特征稳定性’的核心动机几乎同款表述。差异:它是噪声注入而非分布偏移(paraphrase/domain/length),不做方向 cosine 旋转矩阵,也不做训练前 a priori 预测 OOD 迁移。
+- **Huang, 2025** — *Spectral Identifiability for Interpretable Probe Geometry* (arXiv 2511.16288 (stat.ML),待核实) [威胁:high]
+  - 对 C2 威胁高。这正是‘训练前/无需新标注、用内在几何量 a priori 预测探针会不会失稳’的同类工作,与你 Probe Stability Score 的预测性目标直接撞车。差异:用 eigengap/Fisher 谱判据而非 dispersion+augmentation-consistency 双分量;主要在合成数据上验证,未明确做 Pythia/GPT-2 size-ladder、未做 label-preserving NLP 偏移、未给 shift×size 准确率+方向旋转矩阵。
+- **Anonymous/待核实, 2025** — *Factual Self-Awareness in Language Models: Representation, Robustness, and Scaling* (arXiv 2505.21399,待核实) [威胁:high]
+  - 对 C1+C4 威胁高。它已经做了‘Pythia size-ladder × 输入扰动 → 探针准确率下降’的矩阵式分析,且强调‘对表层噪声稳健、对语义偏移敏感’,与你 C1 的 accuracy-drop across shift×size 直接重叠,并覆盖 C4 的规模阶���。差异:只测扰动下的准确率,不测概念方向 cosine 旋转;不提出预测性 stability score(C2);扰动非严格 NLI 审计的 label-preserving 集合(C5);单一任务(事实自知)。
+- **待核实, 2026** — *Linear Probe Accuracy Scales with Model Size and Benefits from Multi-Layer Ensembling* (arXiv 2604.13386,待核实) [威胁:medium]
+  - 对 C4 中度、C1/C3 低度。它直接做了‘探针准确率随 model size 提升’的 scaling 结论(C4 核心),并触及方向跨层旋转(与你的 cosine 旋转测量相邻但你的是跨 shift 旋转、它是跨层旋转)。差异:模型族非 Pythia/GPT-2;不做 label-preserving 分布偏移、不做免标注预测性 stability score、不比较 mass-mean/LogReg/MLP。
+- **Chou, Kirsanov, Yang & Chung, 2026** — *Diagnosing Generalization Failures from Representational Geometry Markers* (ICLR 2026) [威胁:medium]
+  - 对 C2 中度。这是‘用内在几何标记、免 OOD 标注、事先预测泛化失败’的同范式工作,与你 stability score 的预测性主张方法论撞车。差异:领域是视觉而非 LLM/Pythia,不用 augmentation-consistency/dispersion 双分量,不涉及 NLP label-preserving 偏移、方向旋转或 size-ladder。可作为‘几何预测泛化’这一思路的强先例,削弱 C2 的范式新颖性。
+- **Marks & Tegmark, 2023** — *The Geometry of Truth: Emergent Linear Structure in LLM Representations of True/False Datasets* (arXiv 2310.06824(后续亦见会议版,待核实)) [威胁:high]
+  - 对 C3 威胁高、C4 低。它已经做了 mass-mean vs LogReg(vs CCS)在‘跨数据集/分布’下相对可迁移性的系统比较,结论是 mass-mean 更稳——这正是你 C3 的核心命题之一(只是它没纳入 MLP 探针,且‘OOD’指跨数据集而非 paraphrase/domain/length label-preserving 偏移)。差异:不做 size-ladder 矩阵、不做方向旋转量化、不提出预测性 stability score。你的 C3 若不强调与 MLP 的三方比较 + label-preserving shift 设定,很容易被认为是其复现。
+- **待核实(EmergentMind/相关综述指向), 2023** — *Gaussian Process Probes (GPP) for Uncertainty-Aware Probing* (arXiv 2305.18213,待核实) [威胁:low]
+  - 对 C2 低度背景相关。同样关注‘探针给出的判断有多可信/稳定’,但走贝叶斯不确定性路线,不预测 OOD 迁移、不用增广一致性/重采样离散度、不涉及 size-ladder 或方向旋转。可作为‘探针可靠性度量’的相关背景引用。
+- **Biderman et al., 2023** — *Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling* (ICML 2023) [威胁:low]
+  - 背景相关(C4 的实验底座)。本身不主张探针 OOD 稳定性结论,但你的 C4 实验若用 Pythia 则必须引用;它界定了‘size-ladder’这一研究范式,使单纯‘随规模变化’的发现难称新颖。
+
+> 评估: 联网完成。本角度下你们面临的撞车风险偏高,且分散在 C1/C2/C3/C4 四条主张上,需要重新收紧定位。
+
+最大的三个撞车点:
+1) C1(准确率高估稳定性 + Pythia size-ladder × 输入扰动)已被 "Factual Self-Awareness"(2505.21399, 2025)在 Pythia 全 size-ladder 上基本做掉:它已给出‘扰动 × 规模 → 探针准确率下降’的矩阵,并明确‘对表层稳健、对语义偏移敏感’。你们若仅停留在 accuracy-drop,几乎是复现。你们尚存的差异——概念方向 cosine 旋转(direction rotation)、严格用本地 NLI 审计的 label-preserving shift 集合(C5)、覆盖 paraphrase/domain/length 三类偏移而非零散扰动——必须被抬成 C1 的主卖点,否则 C1 会被审稿人判为已有。
+
+2) C2(免标注、forward-pass-only、训练前预测 OOD 迁移的探针稳定性 score)是你们最值钱也最危险的主张:同一‘事前预测探针可靠性/泛化失败’的范式已有至少三篇——Spectral Identifiability(2511.16288, eigengap/Fisher 判据)、Diagnosing Generalization Failures from Geometry Markers(ICLR 2026,几何量预测 OOD,免标签)、以及 Fragility(2606.11375,forward-pass 的崩溃点度量)。范式新颖性已被占。你们真正未被占的是具体机制:dispersion(同分布 K 次重采样方向离散度)+ augmentation-consistency(label-preserving 增广后方向一致性)这个‘双分量、基于方向而非准确率/谱/流形维度’的组合,且明确用于 NLP 探针的 a priori OOD 预测。建议把 C2 的 novelty 严格锚定在‘双分量方向一致性度量’及其相对上述谱/几何/噪声度量的预测增益对比(强烈建议做 head-to-head baseline:SIP、Fragility、manifold dimensionality 都要作为对照)。
+
+3) C3(mass-mean vs LogReg vs MLP 的 OOD 相对可迁移性)中,mass-mean vs LogReg(vs CCS)的跨分布比较已由 Geometry of Truth(Marks & Tegmark 2023)做过,结论 mass-mean 更稳。你们的增量只剩:加入 MLP 探针构成三方比较 + 在 label-preserving paraphrase/domain/length 偏移(而非跨数据集)下重做 + 与 size-ladder 交叉。务必如此措辞,否则 C3 易被视为复现。
+
+留有的空白(可作为你们真正的差异化护城河):
+- 没有任何一篇把‘概念方向 cosine 旋转矩阵’作为 shift×size 的系统量化主轴——这是 C1 的真空白,应重点强化。
+- 没有人用‘重采样离散度 + 增广方向一致性’这个双分量、纯方向几何的免标注 score 来预测 NLP 探针 OOD 迁移,且没人把它和谱/流形/噪声类预测器做过对比——这是 C2 的真空白,但必须靠 baseline 对比来证明增益。
+- C4 单独看价值最低:‘探针随规模更稳/更准’已被多篇(2
+
+### 角度: Probe reliability / calibration / trustworthiness — "when not to trust probes" a...
+
+- **Dies, Maynard, Savcisens & Eliassi-Rad, 2025** — *Representational Stability of Truth in Large Language Models* (arXiv:2511.19166 (preprint; venue待核实)) [威胁:high]
+  - Directly overlaps C1 (concept-direction rotation under paraphrase/perturbation) and C2 (dispersion-across-resamples as a stability signal). This is the single closest paper to our core mechanism — it already operationalizes 'direction rotation under rephrasing' + 'dispersion across resamples'. Differs in that it targets one concept (truth), does not frame the metric as an a-priori label-free OOD-transfer predictor, and lacks C3/C4/C5.
+- **Kumar, Tan et al., 2022** — *Probing Classifiers are Unreliable for Concept Removal and Detection* (NeurIPS 2022 (arXiv:2207.04153)) [威胁:high]
+  - Overlaps C1 (probe accuracy overstates feature quality/stability) and the spirit of C2 (a label-free quality metric for probes). Establishes the 'do not trust probe accuracy' premise our paper depends on, and pre-empts the framing novelty of 'a metric to gauge probe quality.' Differs: focuses on concept removal/spurious correlation, not label-preserving distribution shift, not direction rotation, no OOD-transfer prediction, no estimator/scale ladder.
+- **Lysnæs-Larsen, Eggen & Strümke, 2025** — *Probing the Probes: Methods and Metrics for Concept Alignment* (arXiv:2511.04312 (preprint; cs.AI)) [威胁:high]
+  - Overlaps C2's augmentation-consistency component directly (they call it 'augmentation robustness' as a probe-quality metric) and C1 (accuracy overstates faithfulness). Strong threat to our 'augmentation-consistency as a probe-reliability signal' framing. Differs: vision/CAV-centric, no dispersion-across-resamples component, no a-priori OOD-transfer prediction, no NLP paraphrase/domain shift, no scale ladder.
+- **Liu, Chen, Cheng & He, 2024** — *On the Universal Truthfulness Hyperplane Inside LLMs* (EMNLP 2024 (arXiv:2407.08582)) [威胁:medium]
+  - Overlaps C1 (documents probe accuracy dropping OOD / cross-domain) and partially C3 (compares probing setups across datasets). Establishes the empirical 'probes don't transfer OOD' result we want to extend. Differs: requires labeled OOD datasets to measure transfer (no label-free a-priori predictor), no direction-rotation metric, no dispersion/augmentation score, no scale ladder.
+- **Marks & Tegmark, 2023** — *The Geometry of Truth: Emergent Linear Structure in LLM Representations of True/False Datasets* (arXiv:2310.06824 (later COLM/workshop线; venue待核实)) [威胁:high]
+  - Directly overlaps C3 — it is the canonical mass-mean vs LogReg generalization comparison. Our C3 must position as an extension (adding MLP, label-preserving shift types, a stability-score lens) rather than a first comparison. Differs: no MLP, no paraphrase/length/domain shift taxonomy, no a-priori stability score, no scale ladder.
+- **Baek, Raghunathan et al., 2024** — *Predicting the Performance of Foundation Models via Agreement-on-the-Line* (arXiv:2404.01542 (NeurIPS 2024线; venue待核实)) [威胁:medium]
+  - Overlaps C2's goal (label-free, forward-pass-only prediction of OOD performance) — establishes a strong prior-art baseline for 'predict OOD reliability without OOD labels.' Reviewers will demand comparison. Differs: uses inter-model agreement rather than direction dispersion/augmentation consistency; not probe/concept-direction specific; no rotation metric, no estimator/scale study.
+- **Deng, Zheng et al., 2023** — *Confidence and Dispersity Speak: Characterising Prediction Matrix for Unsupervised Accuracy Estimation* (arXiv:2302.01094 (ICML 2023线; venue待核实)) [威胁:medium]
+  - Overlaps C2 conceptually — a label-free, dispersion-flavored predictor of OOD accuracy. Notably it uses the word 'dispersity,' close to our 'dispersion' component, which weakens the terminology/concept novelty of that sub-claim. Differs: operates on output prediction matrices, not on probe-direction geometry; no concept-direction rotation, no augmentation-consistency, not probe-faithfulness framed.
+- **Belrose et al. / SAE-probe baseline studies, 2025** — *SAE probes underperform the baseline of logistic regression (SAE Features for Classifications and Transferability line)* (arXiv:2502.16681 / 2502.11367 (preprints)) [威胁:low]
+  - Background-to-medium overlap with C3 (probe-estimator transferability comparison under shift). Adds to the crowded space of 'which probe estimator transfers best' that our C3 sits in. Differs: SAE-vs-LogReg focus, no mass-mean/MLP triad, no a-priori stability score, no rotation metric.
+- **Chou, Kirsanov, Yang & Chung, 2026** — *Diagnosing Generalization Failures from Representational Geometry Markers* (ICLR 2026 (arXiv:2603.01879)) [威胁:medium]
+  - Overlaps C2's high-level claim (predict OOD reliability from ID-only signal, no OOD labels). A conceptual competitor for the 'a-priori predictor' framing. Differs: vision/ImageNet, manifold-geometry not probe-direction dispersion/augmentation, no NLP paraphrase shift, no estimator/scale ladder. Adjacent companion: 'Representation Geometry as a Diagnostic for OOD Robustness' (arXiv:2602.03951) makes the same ID-only-predicts-OOD claim.
+- **Belinkov, 2022** — *Probing Classifiers: Promises, Shortcomings, and Advances* (Computational Linguistics (MIT Press) 48(1)) [威胁:low]
+  - Background for C1/C5 — frames the 'probes can mislead' problem and the need for controls (relevant to our NLI label-fidelity audit in C5). Differs: survey, no quantitative shift matrix, no stability score, no OOD predictor.
+
+> 评估: 已联网检索(WebSearch/WebFetch可用)。结论:本角度对我们novelty构成实质威胁,但没有任何一篇把我们的完整组合(C2双分量 a-priori 免标注 forward-pass 探针稳定性分数 = dispersion + augmentation-consistency,用于训练前预测探针OOD迁移性)做完整。
+
+最大撞车点(按优先级):
+1) Representational Stability of Truth in LLMs (Dies et al., 2025-11) 是头号威胁——它已经把"概念方向在改写/扰动下的旋转"与"跨重采样的dispersion"两件事合在一起做了。这几乎覆盖了我们C1的旋转矩阵和C2的dispersion分量。我们必须把它当作最近的prior art正面引用并明确切割:它(a)只针对truth单一概念、(b)没有把这个量当成"训练前、免新标注的OOD迁移性预测器"来validate、(c)没有augmentation-consistency第二分量、(d)没有C3/C4/C5。建议把C2的卖点从"测量稳定性"上移到"用稳定性分数a-priori预测下游OOD迁移,并做预测-验证闭环",这是它没做的。
+
+2) 关于C2的"免标注预测OOD"框架,存在一整条成熟的unsupervised-OOD-accuracy-prediction文献(Agreement-on-the-Line 2024;Confidence-and-Dispersity 2023——后者甚至用了"dispersity"一词,直接削弱我们"dispersion"术语新颖性;Representation-Geometry-as-Diagnostic 2026;Geometry Markers ICLR 2026)。这些不是探针专用,但reviewer几乎一定会要求对比/区分。我们的护城河必须是:作用在探针方向几何(不是输出预测矩阵或流形几何)、专为label-preserving语义shift设计、且做probe-faithfulness而非泛OOD acc。
+
+3) C2的augmentation-consistency分量已被 "Probing the Probes" (2025-11) 以"augmentation robustness"之名作为探针质量度量提出(视觉/CAV域)。需引用并以"NLP label-preserving paraphrase + 与dispersion组合 + 作为预测器"区分。
+
+4) C3(mass-mean vs LogReg)已被 Geometry of Truth (Marks & Tegmark 2023) 做过,是canonical对比,不能宣称首次;加MLP和label-preserving shift分类法是我们的增量,但增量较薄,建议把C3降格为支撑性发现而非主贡献。
+
+5) C1的"探针accuracy高估特征稳定性/faithfulness"premise已被 Kumar et al. NeurIPS 2022 与 Belinkov 2022 survey 牢固确立,不可作为novelty主张,只能作为motivation。
+
+留有的空白(我们仍可主张的真正新意):
+- 把 dispersion + augmentation-consistency 显式组合成单一双分量分数,并在 shift-type × model-size 二维上验证它能a-priori排序探
+
+---
+
+## 四、独立验证记录(2026-06-22,直接抓取 arXiv abstract 复核)
+
+> 由 Claude 主回路用 WebFetch 逐条核对工作流给出的高危引用,确认其**真实存在**(并非 agent 幻觉)。
+
+| arXiv ID | 实际标题 / 作者 | 核实结论 | 对我们的影响 |
+|---|---|---|---|
+| 2511.19166 | *Representational and Behavioral Stability of Truth in LLMs*(Dies, Maynard, Savcisens, Eliassi-Rad)— P-StaT 框架 | ✅ 真实 | 测 truth 信念在语义重述下的稳定性(fictional vs synthetic),**未见**双分量免标注 a-priori 预测闭环。最近邻,但 C2 仍有缝。 |
+| 2602.00158 | *RAPTOR: Ridge-Adaptive Logistic Probes*(Gao 等) | ✅ 真实 | ridge 自适应 logistic 探针 + 概念向量 + stability;偏方法,非直接 OOD 预测。须作 baseline。 |
+| 2510.11905 | *LLM Knowledge is Brittle: Truthfulness Representations Rely on Superficial Resemblance*(Haller, Ibrahim, Kirichenko, Sagun, Bell @ Meta) | ✅ 真实 | 仅覆盖 **paraphrase/reformulation + 仅 truthfulness**;**工作流高估**——未明确做 domain/length/多概念。C1 的"paraphrase×domain×length×多概念×drop⊥rotation 解耦矩阵"仍可区分。 |
+| 2303.15488 | *On the Importance of Feature Separability in Predicting OOD Error*(Xie, Wei, Feng, Cao, An, NeurIPS 2023) | ✅ 真实 | label-free + feature dispersion 预测 OOD acc。标题非"Dispersion Score"但机制重叠 → **避免用 "Dispersion" 作主名 + 必跑对照**。 |
+| 2602.20273 | *The Truthfulness Spectrum Hypothesis*(Ying, Ravfogel, Kriegeskorte, Hase) | ✅ 真实 | **Mahalanobis cosine 预测跨域泛化 R²=0.98** —— C2 最强竞品。但它需"探针间"相似度(隐含 OOD 侧 reference probe);我们 PSS 是 **IID-only**,这是唯一硬区分。 |
+| 2511.16288 | *Spectral Identifiability for Interpretable Probe Geometry (SIP)*(W. Huang) | ✅ 真实 | eigengap a-priori 探针稳定性判据;偏理论+合成实验。必作 baseline。 |
+
+**结论**:该子领域(probe stability / truth directions)是 2025Q4–2026Q2 的高热赛道,多为 Meta/Northeastern 等强组工作。纯 novelty 竞争**高风险**。但所有竞品都可转化为 baseline —— 这为"benchmark-first"重定位提供了基础。

repro_bundle/PROPOSAL.md

@@ -0,0 +1,159 @@
+# ProbeShift —— ACML 2026 主会投稿提案(benchmark-first 定位)
+
+> **定位决策(2026-06-22):** 经深度 prior-art 核查 + 直连 arXiv 验证(见 [PRIOR_ART.md](PRIOR_ART.md)),
+> `probe stability / truth direction` 是 2025Q4–2026Q2 高热赛道,纯 novelty 竞争高风险。
+> 因此**把基准(benchmark)作为第一贡献**,所有竞品转化为在统一基准上评测的 baseline;
+> PAC(我们的 label-free 复合分数)**降为"有竞争力的 IID-only 入选项,不宣称 SOTA"**。
+> 这样即使 PAC 不夺冠,"系统基准 + 统一评测"仍独立成立 —— 这是当前最稳的 accept 路线。
+> predictor-first 的备选 framing 仍保留在 git 历史 / `PROPOSAL_predictor_framing.md`。
+
+---
+
+## 1. 标题与一句话卖点
+
+**中文标题:**《ProbeShift:线性探针方向在语义保真偏移下 OOD 稳定性的系统免标注基准,及对训练前预测信号的统一评测》
+
+**英文标题:** *ProbeShift: A Systematic, Label-free Benchmark for the OOD Directional Stability of Linear Probes under Label-preserving Semantic Shift, with a Unified Evaluation of A-priori Predictors*
+
+**一句话卖点:** 我们**不**主张"首次发现探针不稳",也**不**主张某个新预测器最强;我们提供该方向**第一个系统、免新标注、跨多概念 × 多偏移类型 × 多模型规模 × 多探针类型**的探针 OOD 方向稳定性基准,并在统一基准上对 **5 个现有训练前 label-free 信号**(SIP / RAPTOR-stability / Fragility / augmentation-robustness / Xie-dispersion)做首次 apples-to-apples 评测,诚实回答"**到底什么信号能在训练前预测探针会不会迁移**"。
+
+---
+
+## 2. Abstract(中文,约 190 词)
+
+线性探针被广泛用作 LLM 内部表示的"免标注读出器",但其 in-distribution(ID)准确率系统性高估了它在语义保真偏移(paraphrase / domain / length)下的方向稳定性。近期涌现的诊断与预测工作各自为政、benchmark 与协议互不可比:有的只度量不稳(Dies et al. 2025 的 P-StaT、Haller et al. 2025),有的需触碰 OOD 数据做白化(Truthfulness Spectrum 2026),有的只给单一几何判据(SIP eigengap、Fragility、RAPTOR directional stability、Probing-the-Probes augmentation-robustness)。本文提供 **ProbeShift**:一个完全免新标注、随激活缓存一并发布、可在数分钟内复现的探针 OOD 方向稳定性基准,系统覆盖 concept × shift-type × model-size × estimator,并将 accuracy-drop 与 direction-rotation **解耦**为两维 ground truth。在此基准上,我们首次对 5 个现有训练前 label-free 预测信号做统一 head-to-head 评测(Spearman ρ / Kendall τ,leave-one-concept/shift-out 交叉验证),给出"何种信号预测探针 OOD 稳定性"的可比结论。我们另提供一个简单的 IID-only 复合分数 **PAC**(dispersion ⊕ augmentation-consistency)作为入选项,诚实报告其在何处带来增量、何处现有单分量信号已足够。全部实验受限于单卡 4090 ≤200 GPU·h、API <$100、零新人工标注。
+
+---
+
+## 3. Introduction:研究缺口与贡献
+
+### 3.1 缺口
+探针被当作真实性/欺骗/概念检测的免标注评测器,但在分布偏移下方向几何崩塌。实践者需要**部署前**判断"这个探针方向能否迁移",理想情况下还**不应需要 OOD 数据**。问题在于:相关信号(SIP/Fragility/RAPTOR/aug-robustness/Xie/Truth-Spectrum/P-StaT)散落在**不同概念、不同模型、不同偏移定义、不同评测口径**下,**无人在一个统一基准上比较它们**,导致"到底什么能预测探针 OOD 稳定性"这一实践问题无答案。
+
+### 3.2 贡献(formal,编号)
+- **(G1,主)ProbeShift 基准。** concept(≥12)× shift-type(paraphrase/domain/length)× model-size(Pythia ladder + GPT-2 + Qwen)× estimator(LogReg/mass-mean/MLP)的网格,记录**解耦的** accuracy-drop 与 direction cosine-rotation 作为 ground truth。完全免新标注;随**激活缓存**发布,reviewer 数分钟可复现。
+- **(G2,主)统一 a-priori 预测信号评测。** 首次在同一基准上 head-to-head 评测 SIP(eigengap)、RAPTOR directional stability、Fragility、augmentation-robustness、Xie feature-dispersion 五个现有 label-free 训练前信号对探针 OOD 方向 drop 的预测力(Spearman ρ / Kendall τ + held-out CV),并与"需 OOD 数据"的 Truth-Spectrum 白化 cosine 作上界对照。**诚实结论而非夺冠**。
+- **(G3,次)PAC 入选项。** 一个简单、IID-only、免 OOD 数据的双分量复合分数(dispersion ⊕ augmentation-consistency),作为基准上的一个 entry;做互补性消融,报告其增量边界。**不宣称 SOTA。**
+- **(支撑)** estimator × size 的外部效度;基于本地 NLI 的 label-fidelity sanity check(主动引用 overgenerate-and-filter 文献)。
+
+> **关键 framing:** G1+G2 即使 PAC(G3)毫无增量也独立成立 —— 这是把"赛道很挤"从威胁转成贡献("我们统一比较了所有相关工作")的核心。
+
+---
+
+## 4. Related Work(逐能力对照表 + 逐篇切割)
+
+下表中所有 high-threat 文献均经直连 arXiv 核实(见 PRIOR_ART.md §四);ProbeShift 的定位是**容纳**它们为 baseline,而非取代。
+
+| 能力维度 | Dies/P-StaT | SIP | Fragility | RAPTOR | Probing-Probes | Xie | Truth-Spectrum | **ProbeShift(本文)** |
+|---|---|---|---|---|---|---|---|---|
+| 统一可比基准 | N | N | N | N | N | N | N | **Y(G1)** |
+| 多预测器 head-to-head | N | N | N | N | N | N | N | **Y(G2)** |
+| accuracy-drop ⊥ rotation 解耦 | 部分 | N | N | N | N | N | 部分 | **Y** |
+| 跨多概念(≥12) | N(truth) | 合成 | N | concept | 视觉 | 视觉/分类 | 5 truth | **Y** |
+| 跨 size ladder | N | N | Y | Y | N | N | N | **Y** |
+| 完全免 OOD 数据 | N | Y | Y | Y | Y | Y | N | **Y** |
+| 随激活缓存发布、分钟级复现 | N | N | N | N | N | N | N | **Y** |
+
+**逐篇切割(正文必写):** P-StaT 触碰 OOD 扰动、仅 truth、不预测特定 drop → 我们将其 perturbation 纳入 ground truth 并复现为 baseline;SIP/Fragility/RAPTOR/aug-robustness 各为单一信号 → 我们统一评测之;Xie 预测整体 acc 而非探针方向几何 → §7.4 分离实验证明二者机制不同;Truth-Spectrum 需 OOD covariance → 列为"放宽约束上界"。**全文删除一切 "first predictor / novel framing" 措辞,只保留 "first unified benchmark + evaluation"。**
+
+---
+
+## 5. Formal Claims(可证伪 + 预注册阈值,看到 held-out 前冻结)
+
+- **H1(基准有效性):** ID accuracy 高估方向稳定性。判定:跨配置 Spearman ρ(ID-acc, OOD-retention) 的 bootstrap 95% CI 上界 < 0.5。
+- **H2(G2 主结论):** 存在至少一个现有 label-free 信号能 held-out 预测 OOD 方向 drop。判定:至少一个 baseline 的 held-out ρ 95% CI 下界 > 0.4;**报告全部信号的排名表**(这是 G2 的产出,无论谁赢)。
+- **H3(G3 互补性):** 若 PAC 入选,合成需显著优于其两单分量。判定:配对 Wilcoxon(Holm 后 p<0.05)+ Cliff's δ ≥ 0.33;**否则诚实降级为单分量 entry**。
+- **H4(外部效度):** 排名结论跨 estimator 与跨 size 稳健;不下"越大越稳"普适结论,显式检验 Pressure-Testing-Deception(2605.27958)的 inverse-scaling 是否为 training-distribution artifact。
+- **H5(sanity):** shift 确为 label-preserving。判定:NLI 审计通过率 ≥95%,作者内部抽检(每 shift-type 50 条,**非新众包**)一致率 ≥90%。
+
+---
+
+## 6. Method:基准构造 + PAC 定义
+
+### 6.1 基准 ground truth(G1)
+对每个 (concept c, shift s, size m, estimator e):在 ID 训练集得探针方向 **w̄**;在 OOD(shift)集得 **w_s**。记录两维:
+- **accuracy-drop** = ID_acc − OOD_acc(用 ID 探针在 OOD 集上预测)。
+- **direction-rotation** = 1 − |cos(**w̄**, **w_s**)|(在 OOD 集上重拟合方向,度量几何旋转)。
+解耦报告二者(回应"准确率与方向不是一回事")。
+
+### 6.2 a-priori 预测信号(G2,全部 label-free、IID-only)
+统一接口下实现并评测:
+- **SIP**:LDA 相关子空间的 eigengap / Fisher 误差(越大越稳)。
+- **RAPTOR directional stability**:K 次 bootstrap 方向 mean|cos|(= dispersion 分量)。
+- **Fragility**:对 ID 激活注入递增各向同性噪声,临界崩溃 σ(越大越稳)。
+- **augmentation-robustness**:label-preserving 文本增广后方向一致性(= aug 分量)。
+- **Xie feature-dispersion**:inter-class 特征离散度(预测整体 acc 的通用信号,作机制对照)。
+- **(上界对照)Truth-Spectrum 白化 cosine**:Mahalanobis 白化(注:需 OOD covariance,故列为放宽约束上界)。
+
+### 6.3 PAC 入选项(G3)
+PAC(c,ℓ,e) = σ( α·z(1−D_disp) + (1−α)·D_aug ),α 为唯一聚合超参,**仅在 dev 概念-shift 组合上选**,选定冻结。D_disp 为方向重采样离散度,D_aug 为增广一致性,白化仅用 **ID** covariance(守住免 OOD 约束)。
+
+### 6.4 防泄漏(生死线)
+所有 G2/G3 结果只在**完全 held-out 的 concept × shift-type × size 组合**上算;主结果用 **leave-one-concept-out / leave-one-shift-type-out** 双交叉验证;split 清单随码发布。
+
+### 6.5 Shift 构造 + label-fidelity(sanity)
+paraphrase(LLM API 改写)/ domain(领域措辞改写)/ length(扩缩写),逐字对齐 Haller et al. 2025 三类定义。overgenerate-and-filter:API 过量生成 → 本地 DeBERTa-v3-MNLI 双向 entailment(roundtrip)过滤。主动引用 Falsesum 2022 / DISCO 2023。
+
+---
+
+## 7. Experiments
+
+### 7.1 数据/模型
+- **概念 ≥12**:truth、sentiment、formality、toxicity、language-ID、tense、negation、topic(≥5 类)等,覆盖 factual + stylistic,保证统计功效。
+- **size ladder**:Pythia 70M–2.8B + GPT-2 small/medium + Qwen2.5-0.5B(单卡可载)。
+- 激活一次前向缓存,所有信号共享缓存。
+
+### 7.2 主实验
+1. 构建 ProbeShift ground-truth 矩阵(G1)。
+2. 计算全部 a-priori 信号(含 PAC)。
+3. 报告每个信号 held-out **Spearman ρ / Kendall τ** + bootstrap CI + 排名表(G2)。
+
+### 7.3 关键消融/分离
+- **互补性(H3):** dispersion-only / aug-only / PAC 合成,配对 Wilcoxon + Cliff's δ + Holm。
+- **方向几何 ≠ 流形 dispersion:** paraphrase shift 上对比 PAC 与 Xie,预期 Xie 失效、PAC 不失效。
+- **度量消融:** 裸 cosine vs ID-白化 cosine。
+- **size 效应:** 预测有效性随 size 是否单调;检验 2605.27958 artifact。
+
+### 7.4 统计
+每配置 **5 seeds**;相关系数报 **bootstrap(10k)95% CI**;多重比较 **Holm-Bonferroni**;效应量 **Cliff's δ / Cohen's d**;预注册阈值 §5。
+
+---
+
+## 8. Expected Results 与负结果叙事
+
+**预期产出(无论谁赢都成立):** ProbeShift 排名表给出"现有信号谁能预测探针 OOD 方向 drop";很可能发现单一信号(如 SIP 或 aug-robustness)在某些 shift 类型上强、在另一些上弱,而**没有单信号通吃** —— 这本身就是有价值的 G2 结论。
+
+**诚实负结果(预先写入):**
+- 若 PAC 无显著增量 → 降为"现有单分量已足够 + PAC 仅作统一框架下的参照",G1+G2 不受影响。
+- 若所有信号 held-out ρ 都低 → 这是更强的 G2 结论:"探针 OOD 稳定性目前不可被现有 label-free 信号可靠预测",指明 open problem(对 benchmark 论文是加分而非减分)。
+- 若 size 出现 inverse scaling → 不下普适结论,引 2605.27958 解释为 artifact。
+
+---
+
+## 9. Reproducibility
+发布:激活缓存、探针权重、全部 baseline + PAC 实现、预注册阈值、shift 数据集 + NLI 通过率、seed 列表、leave-one-out split 清单、bootstrap 脚本、一键复现入口。
+
+## 10. Budget(单卡 4090 ≤200 GPU·h,API <$100)
+
+| 项目 | 估算 | 说明 |
+|---|---|---|
+| 激活缓存(12 概念 × 7 模型)| ~40 GPU·h | 一次前向,全 baseline 复用 |
+| 探针训练(3 estimator × bootstrap × 5 seed)| ~50 GPU·h | LogReg/mass-mean 极廉,MLP 主成本 |
+| Augmentation 重训(3 shift)| ~45 GPU·h | 增广激活 + 重训 |
+| Baselines(SIP/Fragility/RAPTOR/aug/Xie/Geometry)| ~35 GPU·h | Fragility 噪声扫描最贵 |
+| P-StaT baseline 复现 | ~15 GPU·h | 用其公开 code/data |
+| 缓冲 | ~15 GPU·h | 合计 **~200 GPU·h** |
+| **API(3 shift × 12 概念 × overgenerate 3×)** | **<$80** | GPT-class API;NLI 过滤本地零成本 |
+
+## 11. Risk & 退路
+
+| 风险 | 缓解 / 退路 |
+|---|---|
+| benchmark 被指"只是把已有数据拼一起" | 强调三个新东西:**解耦 drop⊥rotation**、**跨 ≥12 概念统一口径**、**首次多预测器 head-to-head**;且随缓存发布、分钟级复现是实打实的社区价值。 |
+| 数据泄漏/循环验证(F4) | held-out + 双交叉验证 + split 随码发布。 |
+| PAC 无增量(M1) | 已 framing 为次要 entry,G1+G2 独立成立;诚实降级。 |
+| size inverse-scaling 反例(M2) | 仅交叉维度,复现/检验 2605.27958 artifact,不下普适结论。 |
+| baseline 覆盖不足(M4) | A 组 4 信号 + Xie 全跑,B/C 各 ≥1;缺项写 limitation。 |
+| P-StaT 修订版加入预测闭环 | 不影响本 framing(我们是基准+评测,不与单篇争"首个预测器");投稿前二次复核其 venue/版本。 |
+
+**已核实文献(投稿前复核 venue):** Dies/P-StaT [2511.19166](https://arxiv.org/abs/2511.19166) · SIP [2511.16288](https://arxiv.org/abs/2511.16288) · Fragility [2606.11375](https://arxiv.org/abs/2606.11375) · RAPTOR [2602.00158](https://arxiv.org/abs/2602.00158) · Probing-the-Probes [2511.04312](https://arxiv.org/abs/2511.04312) · Xie [2303.15488](https://arxiv.org/abs/2303.15488) · Truth-Spectrum [2602.20273](https://arxiv.org/abs/2602.20273) · Haller/Brittle [2510.11905](https://arxiv.org/abs/2510.11905) · Pressure-Testing-Deception [2605.27958](https://arxiv.org/abs/2605.27958)

repro_bundle/PROPOSAL_predictor_framing.md

@@ -0,0 +1,219 @@
+# [备选 framing] Predictor-first 提案(PAC 作主贡献)
+
+> 此为 predictor-first 备选;当前主选为 benchmark-first(见 PROPOSAL.md)。保留以备 PAC 增量显著时升级 framing。
+
+---
+
+All verified. The counterexample paper (2605.27958) actually concludes the inverse-scaling pattern is a *training-distribution artifact*, not genuine — which I can use to frame M2 carefully. I now have fully verified ground truth. Writing the complete proposal.
+
+# 《Probe Stability Score》ACML 2026 主会投稿提案(完整版)
+
+> **撰写前置说明(已完成 prior-art 核验)**:本提案撰写前已对定位文档第 4 节"必引清单"中全部 10 篇 high-threat 文献逐条在 arXiv 核实标题、作者与机制。核实结果直接改写了若干防守策略,关键修正如下(投稿前仍须复核 venue/接收状态):
+> - **Dies et al. 2025(2511.19166)** 真实标题为 *Representational **and Behavioral** Stability of Truth in LLMs*,框架名 **P-StaT**。**核实确认:它是"主动扰动数据 + 度量 retraction(最高 32.7%)"的评测框架,既未把 rotation+dispersion 合成单一分数,也未做 a priori OOD 预测,且全程触碰 OOD 数据、仅限 truth 单概念。** 这是 F1 生死线的关键有利事实。
+> - **Xie 2023(2303.15488)** 真实标题为 *On the Importance of Feature Separability in Predicting OOD Error*;其分数为 "feature dispersion",**并未正式命名为 "Dispersion Score"**,且预测目标是整体 test accuracy 而非探针方向几何。F3 改名 + 机制切割成立。
+> - **Truthfulness Spectrum(2602.20273)** 确认其 R²=0.98 依赖 **test-set covariance(即 OOD reference)** 做 Mahalanobis 白化;PSS "免 OOD 数据"护城河有效。
+> - **Probing the Probes(2511.04312)** 真实副标题 *Methods and Metrics for Concept Alignment*,其 augmentation-robustness 面向 **视觉 CAV / 空间概念(translation invariance)**,非文本 paraphrase。切割成立。
+> - **Liang/Brittle(2510.11905)** 第一作者实为 **Patrick Haller et al.**,范围 4 LLM 家族 × 5 数据集 × 3 探针方法,只做 truth。
+> - **Pressure-Testing Deception(2605.27958)** 结论是"inverse scaling 是 training-distribution artifact 而非真实 scale 现象"——M2 据此改为"不下普适 scaling 结论,并复现该 artifact 解释"。
+> - RAPTOR(2602.00158)、SIP(2511.16288)、Fragility(2606.11375)、Marks & Tegmark(2310.06824)均核实属实。
+
+---
+
+## 1. 标题与一句话卖点
+
+**中文标题:** 《探针锚定一致性分数:一种完全免 OOD 标注、仅前向传播的双分量复合指标,用于训练前 a priori 预测线性探针方向在语义保真偏移下的几何稳定性》
+
+**英文标题:** *Probe Anchoring Consistency (PAC): A Label-free, Forward-pass Composite for A-priori Prediction of Linear-probe Directional Stability under Label-preserving Semantic Shift*
+
+> **重要改名(回应 F3):** 弃用 "Probe Stability Score / Dispersion" 作主词根,改用 **Probe Anchoring Consistency(PAC)**,避免与 Xie 2023 的 feature dispersion 第一眼撞名。"dispersion" 仅作内部分量技术名出现,绝不作总分数名。下文 PAC = 原 PSS。
+
+**一句话卖点:** 我们不主张"首次发现探针不稳",而主张提供一个**在严格 held-out、完全不触碰 OOD 数据的约束下,对 SIP / Fragility / Augmentation-robustness / Xie-dispersion / mass-mean 五个 baseline 取得可经 bootstrap 显著性检验之增量**的 label-free 复合预测器,并闭环验证它能排序真实 OOD drop。
+
+---
+
+## 2. Abstract(中文,约 190 词)
+
+线性探针(linear/concept probe)被广泛用于读取 LLM 内部表示,但其 in-distribution(ID)准确率系统性高估了它在语义保真偏移(paraphrase / domain / length)下的方向稳定性。已有工作或停留在"诊断不稳"(Dies et al. 2025 的 P-StaT、Haller et al. 2025),或需触碰 OOD 数据做白化(Truthfulness Spectrum 2026,R²=0.98),或仅给单分量几何判据(SIP、Fragility、RAPTOR directional stability、Probing-the-Probes augmentation-robustness)。我们提出 **Probe Anchoring Consistency(PAC)**:一个仅用前向传播、完全免新标注与免 OOD 数据的双分量复合分数,分量一为方向重采样离散度(dispersion),分量二为文本增广一致性(augmentation-consistency),二者经 held-out 校准聚合为单一分数。我们的核心主张不是 framing 新颖性,而是:(i) 两分量**合成**相对各自单分量有统计显著增量(配对检验 + effect size);(ii) PAC 能在训练前 a priori **排序** 一个 shift-type × model-size × concept 矩阵中的真实 OOD drop(Spearman ρ / Kendall τ,leave-one-concept/​shift-out 交叉验证);(iii) 在零 OOD 数据约束下,对上述五个 baseline 取得 bootstrap 显著的预测增量。全部实验受限于单卡 4090 ≤200 GPU·h。
+
+---
+
+## 3. Introduction:研究缺口与贡献列表
+
+### 3.1 研究缺口
+
+线性探针正被当作"免标注评测器"用于真实性、欺骗、概念检测(Marks & Tegmark 2023;Pressure-Testing Deception 2026)。但探针在 clean benchmark 上 AUROC > 0.96,而在分布偏移下方向几何崩塌(Haller et al. 2025;2605.27958)。实践者需要一个**部署前**就能判断"这个探针方向能否迁移"的指标,且理想情况下**不应需要 OOD 数据**(否则与直接在 OOD 上评测无异)。
+
+现有候选各缺一角:
+- **诊断型**(Dies/P-StaT、Fragility):度量不稳,但不预测特定 OOD drop,且 P-StaT 触碰 OOD 扰动数据。
+- **需 OOD reference 型**(Truthfulness Spectrum):R²=0.98 但需 test-set covariance 白化。
+- **单分量 a priori 型**(SIP eigengap、RAPTOR mean|cos|、Probing-the-Probes augmentation):各占一个分量或一个几何判据,无人**合成两分量**并验证互补增量。
+- **通用免标注 OOD 预测**(Xie 2023 feature dispersion、Deng 2023、Baek 2024):预测整体 accuracy / 流形几何,非**探针方向几何**。
+
+**缺口一句话:** 无人在"双分量合成 + 完全免 OOD 数据 + a priori 预测 ↔ 验证闭环 + 针对探针方向几何"四约束同时成立下,给出经显著性检验、跑赢上述 baseline 的预测器。
+
+### 3.2 贡献列表(formal,编号;已按红队收窄)
+
+- **(主贡献 G1)** 提出 **PAC**——dispersion ⊕ augmentation-consistency 的 label-free、forward-pass、免 OOD 数据复合分数,并在**严格 held-out(leave-one-concept-out / leave-one-shift-type-out)**下证明其 a priori 排序真实 OOD directional drop 的能力(Spearman ρ / Kendall τ + bootstrap CI)。
+- **(主贡献 G2)** **互补性消融**:证明双分量合成相对 dispersion-only、augmentation-only 有**配对显著**增量(Wilcoxon + Cliff's δ),即 augmentation-consistency 在 dispersion 之外携带增量信息。
+- **(主贡献 G3)** **Head-to-head baseline 增量**:在零 OOD 数据约束下,PAC 对 SIP / Fragility / RAPTOR-stability / Probing-the-Probes-augmentation / Xie-feature-dispersion 取得 bootstrap 显著的预测增量;并与需 OOD 的 Truthfulness-Spectrum 白化 cosine 做"放宽约束上界"对照。
+- **(支撑发现 S1=原 C1,降为被预测真值)** 构建 shift-type × model-size × concept 的 **accuracy-drop / cosine-rotation 解耦二维矩阵**,作为 PAC 预测的 ground-truth,而非独立贡献。
+- **(外部效度 S2=原 C3+C4,降级)** mass-mean / LogReg / MLP 三类 estimator 与 size ladder 仅作为"PAC 预测有效性是否跨 estimator、跨 size 成立"的外部效度检验,不并列。
+- **(Sanity check S3=原 C5,降级)** 基于本地 NLI 的 label-fidelity 审计协议,定位为方法学 sanity check,主动引用 NLI-filter 文献。
+
+---
+
+## 4. Related Work(含逐 claim 定位表 + 逐篇切割)
+
+### 4.1 主题归类(略写,完整见附录引用)
+A. 探针准确率不可靠 / faithfulness;B. 真值几何 / 方向旋转 under shift;C. 探针方向稳定性度量;D. 训练前 a priori 判据;E. 免标注 OOD 性能预测;F. transferability estimation;G–I. estimator 比较 / size ladder / NLI 审计。
+
+### 4.2 与每篇最接近工作的逐能力对照表(直接回应 F1/F2/F3)
+
+| 能力维度 | Dies/P-StaT (2511.19166) | SIP (2511.16288) | Fragility (2606.11375) | RAPTOR (2602.00158) | Probing-Probes (2511.04312) | Xie (2303.15488) | Truth-Spectrum (2602.20273) | **PAC(本文)** |
+|---|---|---|---|---|---|---|---|---|
+| rotation 分量 | Y | N | N | N | N | N | Y(cos) | **Y** |
+| dispersion 分量 | Y | 谱gap | 噪声崩点 | Y(mean\|cos\|) | N | feature流形 | N | **Y** |
+| augmentation-consistency 分量 | N | N | N | N | Y(视觉CAV) | N | N | **Y** |
+| **两分量合成单一分数** | **N** | N | N | N | N | N | N | **Y** |
+| a priori **预测特定 OOD drop** | **N(只度量retraction)** | N(只给ID判据) | N(只给崩点) | N(只诊断) | N | Y(整体acc) | Y(需OOD cov) | **Y(探针方向drop)** |
+| **完全免 OOD 数据** | **N** | Y | Y | Y | Y | Y | **N(需test cov)** | **Y** |
+| 预测目标=探针方向几何 | 部分 | ID可识别 | 崩点 | 诊断 | CAV对齐 | **N(流形)** | Y | **Y** |
+| 跨多概念(≥10) | N(truth) | 合成 | N | concept | 视觉 | 视觉/分类 | 5 truth类型 | **Y(目标≥12)** |
+| size ladder | N | N | Y | Y | N | N | N | **Y** |
+| **预测↔验证闭环** | **N** | N | N | N | N | 部分 | 部分 | **Y** |
+
+**逐篇 one-paragraph 切割(论文正文必写):**
+- **vs Dies/P-StaT:** 经核实其为评测框架,度量 retraction 而**不预测**,主动**触碰 OOD 扰动数据**,且**不合成单一分数**、仅 truth。PAC 的"合成 + a priori 预测 + 免 OOD"三格对其为 N→Y。我们将 P-StaT 的 perturbation 集合作为 ground-truth drop 来源之一(被预测真值),并复现其作为 baseline。
+- **vs SIP / Fragility:** 二者是单分量 a priori 判据(eigengap / 崩点),不预测特定 shift 下的 drop 排序。二者作为 baseline 跑数;PAC 主张是**增量**而非取代,删除一切 "first" 措辞。
+- **vs Xie 2023:** 已核实非 "Dispersion Score" 命名、预测整体 acc。我们改名规避撞名,并实验证明"探针方向几何 ≠ feature 流形 dispersion"(见 §7.4 分离实验)。
+- **vs Truthfulness Spectrum:** 其 R²=0.98 **依赖 OOD test covariance**。PAC 在更苛约束(零 OOD)下逼近;我们诚实将白化 cosine 列为"放宽约束的上界对照",而非声称跑赢。
+
+---
+
+## 5. Formal Claims & Hypotheses(可证伪 + 预注册阈值)
+
+所有阈值在**看到 held-out 测试结果前**冻结,写入预注册文件随代码发布。
+
+- **H1(原 C1,被预测真值):** 对每个 (concept, shift-type, size),探针 ID accuracy 与 OOD directional retention(1−rotation)弱相关。**判定:** 若跨配置 Spearman ρ(ID-acc, OOD-retention) 的 bootstrap 95% CI 上界 < 0.5,则 H1 成立(准确率确实高估稳定性)。
+- **H2(主,G1 闭环):** PAC 在 held-out 配置上排序真实 OOD directional drop。**判定:** held-out Spearman ρ(PAC, −drop) ≥ 0.6 且 bootstrap 95% CI 下界 > 0.4。
+- **H3(主,G2 互补性):** 合成 PAC 显著优于两个单分量。**判定:** 配对 Wilcoxon(合成 vs dispersion-only,以及 vs augmentation-only)p < 0.05(Holm 校正后),且 Cliff's δ ≥ 0.33(中等以上效应)。**若不成立,坦诚降级为单分量方法(回应 M1)。**
+- **H4(主,G3 baseline 增量):** PAC 对每个 a priori baseline 的预测 ρ 有正增量。**判定:** Δρ(PAC − baseline) 的 bootstrap 95% CI 下界 > 0,对 SIP/Fragility/RAPTOR-stab/Aug-robust/Xie 五者分别检验,Holm 校正。
+- **H5(外部效度,S2):** PAC 的预测有效性跨 estimator(mass-mean/LogReg/MLP)与跨 size 成立。**判定:** 各 estimator 子集与各 size 子集的 ρ 均 > 0.5;**不下"越大越稳"普适结论**,并显式测试 2605.27958 的 inverse-scaling 是否为 training-distribution artifact。
+- **H6(sanity,S3):** 构造的 shift 确为 label-preserving。**判定:** NLI 审计通过率 ≥ 95%,且与人工抽检(每 shift-type 抽 50 条,作者内部标注,**非新众包标注**)一致率 ≥ 90%。
+
+---
+
+## 6. Method:PAC 的精确定义
+
+### 6.1 记号
+设隐状态层 ℓ、概念 c、估计子 e。在 ID 训练集 D 上得探针方向 **w**(单位向量)。语义保真增广算子集 𝒜 = {paraphrase, domain-reframe, length}(label-preserving)。
+
+### 6.2 分量一:Dispersion(方向重采样离散度,免 OOD)
+对 D 做 B 次 bootstrap 重采样,得方向 {**w**_b}。以 mean direction **w̄** 为锚:
+
+  D_disp(c,ℓ,e) = 1 − (1/B) Σ_b |cos(**w**_b, **w̄**)|
+
+值越小越稳。**度量消融(回应 M3):** 同时计算裸 cosine 与 Mahalanobis 白化 cosine(白化协方差仅用 **ID** covariance,绝不用 OOD,以守住免 OOD 约束),报告 PAC 对度量选择的稳健性。
+
+### 6.3 分量二:Augmentation-consistency(文本增广一致性,免 OOD、免标注)
+对 ID 样本施加 𝒜 中算子得增广版本(**仍是 ID 语义、标签由构造保持**,经 §6.4 审计),在增广前向激活上**重训**探针得 **w**_a:
+
+  D_aug(c,ℓ,e) = 1 − (1/|𝒜|) Σ_{a∈𝒜} |cos(**w**_a, **w̄**)|
+
+关键:这里增广是对 **ID 数据**做 label-preserving 变换以探测方向脆性,**不是 OOD 测试数据**——这正是与 Truthfulness Spectrum 触碰 OOD 的区别。
+
+### 6.4 合成(单一分数,held-out 校准)
+  PAC(c,ℓ,e) = σ( α·z(D_disp) + (1−α)·z(D_aug) )
+其中 z(·) 为基于 **dev 配置** 统计量的标准化,α∈[0,1] 为唯一聚合超参。
+
+**严格防泄漏(回应 F4,生死线):**
+- α **只在 dev 概念-shift 组合上**用 grid {0,0.1,…,1} 选定,选定后冻结。
+- 所有报告结果只在**完全 held-out 的 concept × shift-type × size 组合**上计算。
+- 主结果用 **leave-one-concept-out** 与 **leave-one-shift-type-out** 双交叉验证,证明 PAC 跨概念、跨 shift-type 泛化而非记忆 S1 矩阵。
+
+### 6.5 Shift 构造与 label-fidelity 协议(S3,sanity)
+- 三类 label-preserving shift:**paraphrase**(LLM API 改写)、**domain**(改写为新领域措辞)、**length**(扩写/缩写)。逐字对齐 Haller et al. 2025 的三类定义以便 head-to-head。
+- **生成-过滤管线(overgenerate-and-filter,主动引用 Falsesum 2022 / DISCO 2023):** API 过量生成 → 本地 DeBERTa-v3-MNLI 双向 entailment 过滤(roundtrip consistency)→ 仅保留双向 entailment 的样本。
+- 审计定位为 **sanity check**,报告通过率与作者内部抽检一致率(**零新众包标注**,遵守约束)。
+
+---
+
+## 7. Experiments
+
+### 7.1 数据集与模型
+- **概念(≥12,回应 M6 统计功效):** truth(Cities/Counterfact 风格)、sentiment、formality、toxicity、language-ID、tense、negation、topic(≥5 类)等,涵盖 factual 与 stylistic,确保"跨多概念"卖点有功效;报告每个 ρ 的 bootstrap CI。
+- **模型 / size ladder:** Pythia 全 ladder(70M–2.8B,单卡可载)+ Gemma-2-2B(若显存允许),对齐 2505.21399 / 2605.27958 以复现/检验 scaling。
+- **激活提取:** 一次前向缓存隐状态,全部 baseline 共享缓存(省 GPU)。
+
+### 7.2 Baselines(纳入 must-have,回应 M4)
+- **A 组 a priori(全跑):** SIP(eigengap)、RAPTOR directional stability(K 次 ablation mean|cos|,= dispersion-only 对照)、Fragility(临界噪声崩点)、Probing-the-Probes augmentation-robustness(= augmentation-only 对照,适配为文本增广)。
+- **B 组通用免标注 OOD:** Xie feature-dispersion(必跑,证明非换皮)、Deng 2023 Confidence&Dispersity 或 Baek 2024 Agreement-on-the-Line(至少一)、Geometry Markers(effective manifold dim)。
+- **C 组 estimator:** mass-mean、LogReg、MLP。
+- **D 组方法学:** Hewitt & Liang control-task/selectivity probe;裸 cosine vs Mahalanobis 白化 cosine(Truthfulness-Spectrum 风格,作为放宽 OOD 约束的上界对照)。
+
+### 7.3 主实验
+1. 构建 S1 矩阵(真实 OOD accuracy-drop / cosine-rotation,解耦)。
+2. 计算所有 a priori 预测器(含 PAC)。
+3. 报告 held-out **Spearman ρ / Kendall τ**(PAC vs 真实 drop),双交叉验证。
+
+### 7.4 关键消融与分离实验
+- **互补性(H3):** dispersion-only / augmentation-only / 合成 三档,配对 Wilcoxon + Cliff's δ + Holm 校正。
+- **方向几何 ≠ 流形 dispersion(回应 F3):** 在 paraphrase shift 上对比 PAC 与 Xie feature-dispersion 的预测相关性,预期 Xie 在 paraphrase 上失效(因流形整体不变但方向旋转),PAC 不失效——以图证明机制分离。
+- **度量消融(M3):** 裸 vs 白化 cosine。
+- **size 效应(M2):** PAC 预测有效性随 size 是否单调;复现/检验 2605.27958 的 inverse-scaling-as-artifact。
+
+### 7.5 统计方案
+- 每配置 **5 seeds**;所有相关系数报 **bootstrap(10k)95% CI**;多重比较 **Holm-Bonferroni**;效应量报 **Cliff's δ / Cohen's d**;预注册阈值见 §5。
+
+---
+
+## 8. Expected Results 与可能的负结果叙事
+
+**预期正结果:** H2(ρ≈0.6–0.75 held-out)、H3(合成显著优,δ≈0.4)、H4(对 SIP/Fragility/Xie 有正增量,对 RAPTOR-stab 因其=dispersion-only 故增量主要来自 augmentation 分量)。
+
+**可能负结果的诚实叙事(预先写入,回应 M1/M2):**
+- **若 H3 不成立**(合成不优于 dispersion-only):坦诚降级为"单分量 dispersion 预测器 + 一项 augmentation 失败的负结果",仍贡献"augmentation 分量在文本探针上无增量"这一可发表 negative finding。
+- **若对某 baseline 无显著增量**:在 limitation 明确该 baseline 已足够,PAC 的价值退守到"统一框架 + 闭环验证"。
+- **若 size 出现 inverse scaling**:不下普适结论,引 2605.27958 解释为 training-distribution artifact 并复现其 style-augmented 恢复实验。
+- **若白化 cosine(用 OOD cov)大幅超过 PAC**:诚实陈述"免 OOD 的代价",将 PAC 定位为"约束受限场景下的最优",而非全局最优。
+
+## 9. Reproducibility
+- 发布:全部**激活缓存**、探针权重、PAC/baseline 实现、**预注册阈值文件**、shift 数据集 + NLI 审计通过率、seed 列表、leave-one-out split 清单、bootstrap 脚本。
+- 一键复现脚本 + 固定 commit;所有 baseline 在共享缓存上运行以保证可比。
+
+## 10. Budget 明细(单卡 4090 ≤200 GPU·h,API <$100)
+
+| 项目 | 估算 | 说明 |
+|---|---|---|
+| 激活缓存(12 概念 × ladder 6 size × 缓存)| ~40 GPU·h | 一次性前向,全 baseline 复用 |
+| 探针训练(3 estimator × bootstrap B=50 × 5 seed)| ~50 GPU·h | LogReg/mass-mean 极廉,MLP 主成本 |
+| Augmentation 重训探针(3 shift × 上述)| ~45 GPU·h | label-preserving 增广激活 + 重训 |
+| Baselines(SIP/Fragility/RAPTOR/Aug/Xie/Geometry)| ~35 GPU·h | Fragility 噪声扫描最贵 |
+| Dies/P-StaT baseline 复现 | ~15 GPU·h | 用其公开 code/data |
+| 缓冲 | ~15 GPU·h | 共计 **~200 GPU·h** |
+| **API(paraphrase/domain/length 过量生成)** | **<$80** | 12 概念 × 3 shift × overgenerate 3x,GPT-class API;DeBERTa-MNLI 过滤本地零成本 |
+
+## 11. Risk & 退路(逐条对应红队致命攻击)
+
+| 攻击 | 缓解 / 退路 |
+|---|---|
+| **F1(Dies 抹平)** | 已核实 P-StaT 不预测、触碰 OOD、不合成、仅 truth → 逐能力对照表 3 格 N→Y(合成/预测/免OOD,均非边角);复现 P-StaT 作 baseline。**退路:** 若 P-StaT 修订版加入预测,立即二次检索并将 PAC 收窄至"免 OOD + 合成"双格。 |
+| **F2(首个被证伪)** | **全文删除 first/novel framing**,改 "empirically outperforms ... under zero-OOD constraint";SIP/Fragility/Truth-Spectrum 全跑;增量带 bootstrap CI。 |
+| **F3(Xie 换皮)** | 改名 PAC;§7.4 分离实验证明方向几何 ≠ 流形 dispersion(paraphrase 上 Xie 失效)。 |
+| **F4(循环论证/泄漏)** | α 仅 dev 选;结果仅 held-out;leave-one-concept/shift-out 双交叉验证;split 清单随码发布。 |
+| **F5(C3 复现)** | C3 降为 S2 外部效度,仅在"PAC 能否预测三 estimator 谁更稳"从属问题下出现,绝不并列。 |
+| **M1** | H3 配对显著 + 效应量;不显著则降级 + negative finding。 |
+| **M2** | C4 仅交叉维度;主动复现/检验 2605.27958 artifact,不下普适 scaling 结论。 |
+| **M3** | 裸 vs 白化 cosine 消融(白化仅用 ID cov)。 |
+| **M4** | A 组四 baseline + Xie 全跑,B/C 各 ≥1;缺项写 limitation。 |
+| **M5** | C5 降 sanity;引 Falsesum/DISCO;报通过率 + 作者内部抽检一致率(零新众包)。 |
+| **M6** | ≥12 概念;ρ 报 bootstrap CI;概念不足则撤回"跨多概念"主张。 |
+| **M7** | 已逐条核实 10 篇 high-threat(本提案顶部);投稿前再复核 venue/接收状态,对所有篇目无论是否发表均切割。 |
+
+**三条生死线状态:** F1(对照表 3 格 N→Y + 复现跑赢)✓ 设计就绪;F4(无泄漏 split + 双交叉验证)✓ 设计就绪;M1(互补性显著或诚实降级)✓ 含退路。三者同时成立,具备从 Reject 翻盘的最小充分条件。
+
+---
+
+**关键风险提示(交付给通讯作者):** 唯一未闭合的风险仍是 **Dies/P-StaT 是否会在修订版加入预测闭环**。本提案已核实其当前版本不预测,但投稿前须二次检索其最新版与是否被 NeurIPS/ICLR 接收;若其新增预测,PAC 须立即收窄至"合成 + 免 OOD 数据"双格护城河,并强化 §7.4 的方向几何 vs 流形分离实验作为差异化主轴。
+
+**已核实文献链接(供投稿前复核 venue):** Dies/P-StaT [2511.19166](https://arxiv.org/abs/2511.19166) · SIP [2511.16288](https://arxiv.org/abs/2511.16288) · Fragility [2606.11375](https://arxiv.org/abs/2606.11375) · RAPTOR [2602.00158](https://arxiv.org/abs/2602.00158) · Probing-the-Probes [2511.04312](https://arxiv.org/abs/2511.04312) · Xie [2303.15488](https://arxiv.org/abs/2303.15488) · Truthfulness-Spectrum [2602.20273](https://arxiv.org/abs/2602.20273) · Haller/Brittle [2510.11905](https://arxiv.org/abs/2510.11905) · Factual-Self-Awareness [2505.21399](https://arxiv.org/abs/2505.21399) · Pressure-Testing-Deception [2605.27958](https://arxiv.org/abs/2605.27958)

repro_bundle/README.md

@@ -0,0 +1,90 @@
+# ProbeShift
+
+> *ProbeShift: A Systematic, Label-free Benchmark for the OOD Directional Stability of Linear
+> Probes under Label-preserving Semantic Shift, with a Unified Evaluation of A-priori Predictors.*
+>
+> 投稿目标:ACML 2026 主会。定位:**benchmark-first**(见 [PROPOSAL.md](PROPOSAL.md))。
+> 预算锚:单卡 4090 ≤ 200 GPU·h、LLM API <$100、零新人工标注。
+
+## 核心贡献(benchmark-first,见 PROPOSAL.md)
+
+- **G1(主)ProbeShift 基准**:concept(≥12)× shift(paraphrase/domain/length)× size(Pythia ladder
+  + GPT-2 + Qwen)× estimator(LogReg/mass-mean/MLP)的网格,记录 **解耦的** accuracy-drop 与 direction
+  cosine-rotation 作为 ground truth。完全免新标注;随激活缓存发布,分钟级复现。
+- **G2(主)统一 a-priori 预测信号评测**:在同一基准上 head-to-head 评测现有 label-free 训练前信号
+  —— **SIP**(eigengap)、**RAPTOR-stability**、**Fragility**、**augmentation-robustness**、**Xie
+  feature-dispersion** —— 对探针 OOD 方向 drop 的预测力(Spearman/Kendall + leave-one-concept-out CV)。
+- **G3(次)PAC 入选项**:简单、IID-only、免 OOD 数据的双分量复合分数(dispersion ⊕ augmentation-
+  consistency),作为基准上的一个 entry,做互补性消融。**不宣称 SOTA。**
+- **支撑**:estimator × size 外部效度;本地 NLI 的 label-fidelity sanity check。
+
+> 关键 framing:G1+G2 即使 PAC 毫无增量也独立成立 —— 把"赛道很挤"从威胁转成贡献。
+> 所有竞品均经直连 arXiv 核实为真(见 [PRIOR_ART.md](PRIOR_ART.md) §四),并已转化为 baseline。
+
+## Pipeline
+
+```
+extract   一次性抽取 residual-stream 激活,float16 分片缓存          (唯一大头 GPU)
+eval      训练 LogReg/mass-mean/MLP 探针,IID/OOD accuracy + 方向旋转  (读缓存) -> G1 ground truth
+predict   对每个 config 计算全部 a-priori 预测器(SIP/Fragility/.../PAC)(读缓存) -> G2 信号
+score     PAC 双分量明细 + 互补性消融                                (读缓存) -> G3
+stats     每个预测器 vs 真实 OOD drop 的 Spearman/Kendall + LOCO 排名表 (读缓存) -> G2 主结果
+```
+
+**关键设计:激活只抽一次、落盘、之后全部读缓存** —— 这是整个项目守住算力预算 + 可复现的命门。
+缓存可直接随论文发布,reviewer 几分钟即可复现探针实验。
+
+## 安装
+
+```bash
+cd probe_stability
+python -m venv .venv && source .venv/bin/activate     # 或用 conda
+pip install -r requirements.txt
+```
+
+## 最小可行实验(M2 go/no-go,Day 3)
+
+一条命令在单模型单数据集上验证"探针脆性"是否存在(OOD drop + 方向旋转):
+
+```bash
+python run_pipeline.py smoke --model EleutherAI/pythia-410m --dataset sst2
+```
+
+若代表性概念上 **OOD drop < 5pt 或方向 cosine > 0.9**(探针根本不脆)→ 触发退路:切提案②
+(Demonstration Order Illusion),M0/M1 基础设施可复用。
+
+## 全量运行
+
+```bash
+python run_pipeline.py extract --all     # 抽取所有 (model,dataset,distribution) 激活
+python run_pipeline.py eval    --all      # 训练探针 + IID/OOD + 旋转矩阵 (G1 ground truth)
+python run_pipeline.py predict --all      # 计算全部 a-priori 预测器 (G2 信号)
+python run_pipeline.py score   --all      # PAC 双分量明细 + 互补性消融 (G3)
+python run_pipeline.py stats              # 多预测器 vs OOD drop 排名表 (G2 主结果;读全部结果)
+```
+
+## 目录结构
+
+```
+probe_stability/
+├── README.md
+├── requirements.txt
+├── config.py                # 模型/数据/shift 注册表 + 路径 + 超参
+├── cache.py                 # 激活分片的命名/保存/mmap 读取 + manifest
+├── metrics.py               # accuracy / 方向旋转 / Spearman / bootstrap CI / Cliff's delta / Holm-BH
+├── probes.py                # LogReg / mass-mean / MLP / control-task 探针
+├── stability_score.py       # dispersion + augmentation-consistency
+├── extract_activations.py   # HF 模型 hooks,residual-stream 抽取 + float16 分片落盘
+├── run_pipeline.py          # CLI 编排 M1–M5
+└── data/
+    ├── __init__.py
+    ├── datasets.py          # 统一加载 HF 数据集 -> (text, label, concept)
+    ├── shifts.py            # label-preserving shift 构造(back-translation / domain / length)
+    └── label_fidelity.py    # DeBERTa-MNLI 过滤 label-flip
+```
+
+## 状态
+
+骨架代码:核心数据流、缓存、探针、Stability Score、指标已实现;以下处标注 `TODO` 待在 4090 上迭代:
+back-translation 模型选型与显存调参、各数据集 domain-shift 配对、MLP 探针方向旋转的处理。
+`PROPOSAL.md` / `PRIOR_ART.md` 由研究工作流产出后补入。

repro_bundle/RESULTS.md

@@ -0,0 +1,203 @@
+# ProbeShift —— 全量跑结果
+
+## ⚠️ 关键诚实校正(对抗式评审驱动,2026-06-23)
+
+评审指出 EXCESS=naive−placebo 的差值构造会**机械制造负相关**(dispersion 高度预测 placebo,从被减数减去它→残差对 dispersion 必然趋负)。用**偏相关**(无差值)复核:
+
+| 预测器 | 偏相关 ρ(signal, para-rot \| placebo) | EXCESS ρ(差值) |
+|---|---|---|
+| pac | **+0.374** | −0.029 |
+| augmentation_robustness | **+0.345** | +0.046 |
+| raptor_stability | **+0.323** | −0.195(artifact) |
+| whitened_cosine | −0.015 | −0.245 |
+
+**结论修正:** (1) "dispersion 主动误导/anti-predict" 是**差值 artifact**——偏相关下 dispersion 正 +0.323;
+(2) 戏剧"反转"大幅缩水——偏相关下 aug(0.345)≈raptor(0.323),PAC 最高(0.374);
+(3) cluster bootstrap(按 12 概念,修正伪重复)Δρ[aug−raptor] on EXCESS 仍 +0.241 [+0.118,+0.363] SIGNIFICANT,
+但 EXCESS 目标本身被质疑。**诚实 headline:循环性真实重要;但 shift 脆性不是"完全不可预测",dispersion/aug 都有适度残余预测力(~0.32–0.37,PAC 最好),原"open problem + 反转"被差值构造夸大。**
+
+---
+
+## ★★★ 原始结果:Option A,5 独立 seed,n=472(差值/EXCESS 口径,需配合上方校正解读)
+
+12 概念 × 7 模型 × **5 个完全独立的 seed**(各自独立抽样 + 独立回译,gold-standard 复现)。
+gpt2-medium/qwen 的少数 massive-activation 格子被容错跳过(每 seed ~4 个)。
+
+### 四目标 + 配对显著性(n=472)
+| 预测器 | acc-drop | naive-rot | **PLACEBO(IID-resample)** | **EXCESS(非循环)** |
+|---|---|---|---|---|
+| raptor_stability(纯dispersion) | -0.06 | 0.945 [.93,.95] | **0.966 [.96,.97]** | **-0.195 [-.29,-.12]** |
+| pac | 0.12 | 0.839 | 0.799 | -0.029 [-.11,.06] |
+| augmentation_robustness | 0.18 | 0.711 | 0.653 | **0.046 [-.04,.13]** |
+| whitened_cosine_id | -0.08 | 0.769 | 0.814 | -0.245 [-.34,-.16] |
+| fragility | 0.11 | 0.640 | 0.643 | -0.136 |
+| xie_feature_dispersion | 0.06 | 0.458 | 0.464 | -0.117 |
+| sip_eigengap | 0.00 | -0.072 | -0.057 | 0.020 |
+
+**配对检验(bootstrap 95% CI,全 SIGNIFICANT):**
+- Δρ[**aug − raptor**] on EXCESS = **+0.241 [+0.170,+0.309]** ✓
+- Δρ[**pac − raptor**] on EXCESS = **+0.166 [+0.116,+0.215]** ✓
+- Δρ[raptor − aug] on naive-rot = +0.234 [+0.189,+0.284] ✓(naive 上 dispersion 赢——但循环)
+
+### 最终结论(5-seed,最严格、最诚实)
+1. **naive rotation 的强预测是循环的**:raptor 对 naive 0.945,对**纯采样安慰剂 0.966(更高)** → 它预测的是采样噪声地板。红队 F4 循环性被自家对照坐实(多 seed,CI 极紧)。
+2. **真·shift 脆性(EXCESS)无人能正向预测**:最佳 aug 仅 0.046(CI 跨 0);dispersion/whitened **显著负相关**(-0.195/-0.245)→ 它们在诚实目标上**主动误导**。→ 这是有价值的 **open problem**。
+3. **反转显著(headline)**:在诚实的 EXCESS 上,**基于增广的信号显著优于基于离散度的**(Δρ aug−raptor=+0.241,5-seed 稳健),naive 上则相反。即"哪个信号有用"完全取决于是否扣掉采样噪声。
+4. acc-drop:多 seed 下 aug 0.18 弱显著但 LOCO 0.06(跨概念不泛化)。
+
+### Tier 1 重分析(纯缓存,2026-06-23)
+- **Shift-type 拆分**:paraphrase(n=472,raptor 0.945 但循环)、domain(n=65,弱:raptor 0.59)、
+  length(n=30,全高:aug 0.95/pac 0.95/raptor 0.92)。预测器行为强烈依赖 shift 类型。
+- **Estimator 外部效度**:用 logreg 预测器去预测 **mass-mean 方向旋转** → 全部弱(xie 0.47 最高,
+  其余 0.25-0.29)且 **LOCO≈0(跨概念不泛化)**。→ 可预测性是 estimator 特异的,非一般属性(重要 caveat)。
+- **Label-fidelity(NLI 通过率)**:整体 80.6%;counterfact 96%/sst2 89% 高,**dbpedia 50% / imdb 64% 偏低**
+  (实体/长文回译易翻车)→ 按 H7 标注,这两个数据集 rotation 解读需谨慎。
+- **Layer 稳健性(refit)**:paraphrase rotation 按相对深度 0.68→0.71→0.79→0.78,**所有层都普遍(非挑层产物)**。
+- **白化度量消融(M3,refit)**:用 ID-白化 cosine 作目标时,raptor 从裸 cosine 的 0.945 **掉到 0.698**
+  (whitened_cosine_id 0.851 居首)→ dispersion 的循环优势**部分由度量造成**(裸 cosine 忽略协方差);
+  换白化度量后优势缩小。诚实的"结果依赖度量"caveat,非度量取巧。
+
+### Tier 2 多-pivot 增广(de/fr/ru,5 seed × 7 模型,n=472/490)
+- aug 在 EXCESS 上从 1-aug 的 0.046 仅升到 **0.070(CI 仍跨 0)** → 多样化增广**没把它推到显著正**。
+  → **强化 open-problem**:即使用 de/fr/ru 多样增广,shift 特异脆性仍不可正向预测(排除"增广太弱"解释)。
+- 反转**对增广多样性稳健**:Δρ[aug−raptor] on EXCESS = **+0.247**(1-aug 是 +0.241),SIGNIFICANT。
+- 结论:Tier 2 是干净的鲁棒性确认——核心发现不依赖增广质量。
+
+### Tier 3 #11 7B 抽检(pythia-6.9b,seed 0,7 数据集,n=7)
+- naive-rot:raptor **+0.93**(仍循环);excess:raptor **−0.39**(仍反预测)、aug **+0.68**(更强正)。
+- → **循环性 + aug>dispersion 反转在 6.9B 仍成立且 aug 优势放大**(n=7 偏粗,方向一致)→ 堵死"仅小模型"质疑。
+- size-ladder(70M→6.9B):raptor naive 全程 0.83–0.95、excess 全程 ≤0;**可预测性结构对规模不变,无 inverse-scaling**。
+
+### 图(figures/)
+- fig1_circularity:预测器 × {naive / 安慰剂 / excess} 分组柱(killer figure:naive≈安慰剂=循环,excess≤0)。
+- fig2_sizeladder:raptor/aug/pac 的 ρ vs 规模(naive 与 excess 两面板)。
+- fig3_mechanism:raptor vs excess(负)对比 aug vs excess(正)散点——为什么 aug 行 dispersion 不行。
+
+---
+
+## ★★ 中间版:12 概念 / 84 configs(n=78,单 seed)+ 循环性对照
+
+**配置:** 6 模型 × 14 数据集 / **12 概念**(sentiment, topic, truth, emotion, hate, irony, offensive,
+subjectivity, spam, grammaticality, stance, counterfactual)。stance 因推文经 NLI label-fidelity 过滤后
+paraphrase 集为空,部分 config 缺 rotation(n=78)。
+
+### 四目标对照(Spearman ρ,n=78)
+| 预测器 | acc-drop | rotation(naive) | **IID-resample(安慰剂)** | **EXCESS(=para−IID, 非循环)** | EXCESS-LOCO |
+|---|---|---|---|---|---|
+| raptor_stability(纯 dispersion) | -0.07 | 0.951 | **0.975** | **-0.075** | -0.23 |
+| pac(ours = disp ⊕ aug) | 0.09 | 0.895 | 0.859 | **0.124** | -0.03 |
+| **augmentation_robustness** | 0.15 | 0.811 | 0.752 | **0.225** [0.00,0.41] | 0.09 |
+| whitened_cosine_id | -0.13 | 0.741 | 0.790 | -0.186 | -0.14 |
+| fragility | 0.09 | 0.685 | 0.685 | -0.015 | -0.04 |
+| xie_feature_dispersion | -0.01 | 0.502 | 0.491 | -0.053 | -0.01 |
+| sip_eigengap | 0.12 | -0.151 | -0.092 | -0.101 | 0.05 |
+
+### 决定性结论(改写论文故事)
+1. **naive rotation 预测大部分是循环的**:raptor_stability 对 naive rotation 0.951,但对**安慰剂(纯 IID
+   采样旋转)同样 0.975** → 它主要在预测采样噪声地板,而非 shift 脆性。**红队 F4 循环性质疑被我们自己的
+   对照证实。**
+2. **真·shift 特异脆性(EXCESS rotation)几乎不可预测**:去掉采样噪声后,dispersion 彻底失效(-0.075),
+   全场最高仅 augmentation_robustness 0.225(CI 下界贴 0)。→ 这是真正的 **open problem**(benchmark 论文加分)。
+3. **关键反转,支持 augmentation 组件**:在诚实的 EXCESS 目标上,**基于增广的信号 > 基于离散度的**
+   (aug 0.225 / pac 0.124 > raptor -0.075)。augmentation 直接探测 label-preserving 扰动下的方向变化,
+   天然捕捉 shift 特异结构;纯 bootstrap dispersion 不能。**PAC 的 aug 分量恰在最该起作用处起作用。**
+4. accuracy-drop 仍全场不可预测(≈0)。
+
+> **论文主张应改为**:不是"我们提出最强预测器",而是"**我们用循环性对照揭示:现有探针稳定性信号对 OOD
+> 方向旋转的强预测力大部分是采样噪声的循环假象;扣除后,shift 特异脆性基本不可预测,且只有基于
+> label-preserving 增广的信号有微弱(非循环)预测力**"。这是更难被攻击、更有 insight 的 contribution。
+
+数据:`results_full_run/`(eval 234 行、predictors 78、stats_ranking 4 张表、audit 84)。
+
+---
+
+## 附:9 概念 / 66 configs(中间版,无循环性对照)
+
+**配置:** 6 模型(pythia-70m/160m/410m/1.4b + gpt2 + qwen2.5-0.5b)× 11 数据集 / **9 概念**:
+sentiment(sst2,imdb)、topic(ag_news,dbpedia)、truth(counterfact)、emotion、hate、irony、
+offensive、subjectivity(subj)、spam。N_train=1500 / N_eval=800 / n_aug=1,单卡 4090。
+各数据集 IID 探针准确率均显著高于随机(dbpedia .96 / spam .99 / subj .93 / … / counterfact .55,
+真值探针接近随机符合"小模型读真值难"的已知规律)。
+
+### 目标①:预测 OOD accuracy-drop —— 全员 ρ≈0,CI 全跨 0、LOCO≈0 → **不可预测**
+### 目标②:预测 OOD direction-rotation
+| 预测器 | Spearman | 95% CI | LOCO(9概念留一) |
+|---|---|---|---|
+| **raptor_stability**(IID方向离散度) | **0.961** | [0.91, 0.98] | **0.879** |
+| **pac(ours)** | **0.881** | [0.78, 0.93] | **0.810** |
+| augmentation_robustness | 0.774 | [0.64, 0.86] | 0.691 |
+| whitened_cosine_id | 0.764 | [0.60, 0.87] | **0.205** ⚠️不泛化 |
+| fragility | 0.704 | [0.52, 0.84] | 0.388 |
+| xie_feature_dispersion | 0.428 | [0.15, 0.64] | **0.030** ⚠️不泛化 |
+| sip_eigengap | -0.135 | [-0.39, 0.14] | 0.192 |
+
+**9 概念版关键结论(比 3 概念版更硬):**
+1. **解耦 + 可预测性不对称**(headline,稳健):rotation 高度可预测(raptor ρ=0.96,LOCO 0.88 跨 9 概念),
+   accuracy-drop 完全不可预测(全 ~0)。
+2. **LOCO 列成为强区分器**:raptor(0.88)/pac(0.81)/aug(0.69)**跨概念泛化**;而 whitened_cosine(0.21)、
+   xie(0.03)样本内不错但**跨概念崩盘** → "哪些信号 robust、哪些是概念过拟合"本身是 G2 的有价值产出。
+3. **xie 跨概念失败**(LOCO 0.03)稳健复现 → feature 流形 dispersion ≠ probe 方向几何。
+4. **PAC 诚实**:competitive(第2,LOCO 0.81 稳健)但未超过单分量 raptor → H3 互补性仍不成立,如实报告。
+
+---
+
+## 附:初版 3 概念 / 30 configs(已被上面的 9 概念版取代,留作记录)
+
+**配置:** 6 模型 × 5 数据集(sentiment×2, topic×2, truth×1)= **30 configs**,3 概念。
+N_train=1500 / N_eval=800 / n_aug=1。单卡 4090。
+
+## G2 主结果:哪个 a-priori 信号能预测探针 OOD 不稳定性?
+
+报告 `Spearman ρ(信号, −目标)`(越高=越能预测),bootstrap 95% CI,leave-one-concept-out(LOCO)。
+
+### 目标一:OOD **accuracy-drop**(主要由 domain shift 驱动)
+| 预测器 | Spearman | 95% CI | LOCO |
+|---|---|---|---|
+| augmentation_robustness | 0.109 | [-0.27, 0.43] | -0.47 |
+| pac (ours) | -0.015 | [-0.35, 0.31] | -0.62 |
+| xie_feature_dispersion | -0.029 | [-0.38, 0.33] | -0.05 |
+| fragility | -0.082 | [-0.39, 0.24] | -0.26 |
+| raptor_stability | -0.095 | [-0.37, 0.19] | -0.35 |
+| whitened_cosine_id | -0.193 | [-0.48, 0.15] | -0.34 |
+| sip_eigengap | -0.384 | [-0.60, -0.12] | +0.40 |
+
+→ **结论:accuracy-drop 基本不可被任何现有 label-free 几何信号 a-priori 预测**(所有 CI 跨 0;
+sip 弱负相关但 LOCO 翻号、不稳)。这是一个有价值的"open problem"型结论(对 benchmark 论文是加分)。
+
+### 目标二:OOD **direction-rotation**(主要由 paraphrase 驱动)
+| 预测器 | Spearman | 95% CI | LOCO |
+|---|---|---|---|
+| **raptor_stability** | **0.969** | [0.89, 0.99] | **0.937** |
+| **pac (ours)** | **0.818** | [0.60, 0.92] | 0.662 |
+| whitened_cosine_id | 0.750 | [0.46, 0.90] | 0.592 |
+| fragility | 0.682 | [0.39, 0.86] | 0.271 |
+| xie_feature_dispersion | 0.677 | [0.39, 0.85] | **-0.052** |
+| augmentation_robustness | 0.677 | [0.43, 0.83] | 0.493 |
+| sip_eigengap | -0.525 | [-0.78, -0.10] | -0.305 |
+
+→ **结论:direction-rotation 高度可 a-priori 预测**;`raptor_stability`(= IID 方向 bootstrap 离散度)
+近乎完美(ρ=0.969,LOCO 0.937)。
+
+## 关键科学结论(支撑论文)
+
+1. **解耦 + 可预测性不对称(headline)**:两个 OOD 目标彻底分离——**方向旋转高度可预测(ρ≈0.97),
+   准确率下降几乎不可预测(ρ≈0)**。这把 "accuracy ⊥ direction" 从"现象"升级为"可预测性结构差异",
+   是 G1+G2 的核心卖点。
+2. **PAC 诚实结论**:PAC(双分量合成)排第二(0.818),但**未超过单分量 raptor_stability(0.969)**
+   → 互补性假设 H3 在此**不成立**:对旋转预测,augmentation 分量未在 dispersion 之外带来增量。
+   benchmark-first 定位正确——PAC 只是入选项之一,我们如实报告"现有单分量已足够"。
+3. **xie 的跨概念失败**:xie 在样本内 0.677,但 **LOCO -0.05**(跨概念不泛化)→ 实证"feature 流形
+   dispersion ≠ probe 方向几何",正好坐实 §7.4 的机制分离论点。
+4. **sip_eigengap 反相关**:eigengap 在此不是好的旋转预测器(甚至反号),与其理论定位形成对比,值得讨论。
+
+## 待处理 / 注意(投稿前)
+
+- **循环性(对应红队 F4)**:dispersion(IID bootstrap 方向方差)预测 rotation(shift 引起的方向变化),
+  二者都刻画"方向移动"。非平凡点在于 dispersion **完全不碰 OOD 数据**即可预测 shift 后旋转——论文须把
+  这点讲清,并补一个"dispersion 与 rotation 测量相互独立"的对照(如不同随机源、白化对照)以防"换皮"质疑。
+- n=30、仅 3 概念,LOCO 只有 3 组;论文版需扩到 ≥8-12 概念以增强 LOCO 统计功效。
+- n_aug=1(单 pivot 回译);多 pivot 多样化增广留待论文版,可能改变 augmentation_robustness/PAC 的表现。
+- accuracy-drop 的 domain shift 仅 sentiment 数据集有(topic/truth 无 domain 配对),其不可预测性结论需在
+  更广 domain 覆盖下复核。
+
+数据存档:`results_full_run/`(eval.jsonl 90 行、predictors.jsonl 30、stats_ranking.jsonl、audit.jsonl)。

repro_bundle/analyze.py

@@ -0,0 +1,214 @@
+"""Tier-1 re-analysis of cached results (no GPU, no refitting): reads results/*.jsonl and
+produces the breakdowns the paper needs. Safe to run alongside / after the grid.
+
+    python analyze.py shift        # per shift-type predictor ranking (paraphrase/domain/length)
+    python analyze.py estimator    # cross-estimator: do logreg-based predictors predict mass-mean rotation?
+    python analyze.py fidelity     # label-fidelity (NLI flip-rate) summary per dataset
+    python analyze.py all
+"""
+from __future__ import annotations
+
+import sys
+
+import numpy as np
+
+import metrics
+from config import DATASETS
+from baselines import PREDICTORS
+from run_pipeline import _rot_mag, _mean_ood_drop, _read_results, _rank_predictors
+
+
+def _pred_index(preds):
+    return {(p["model"], p["dataset"], p["seed"]): p["predictors"] for p in preds}
+
+
+def shift_breakdown():
+    evals = _read_results("eval")
+    pidx = _pred_index(_read_results("predictors"))
+    print("\n########## SHIFT-TYPE BREAKDOWN (rotation target per shift) ##########")
+    for shift in ["paraphrase", "domain", "length"]:
+        rows = []
+        for e in evals:
+            if e.get("probe") != "logreg":
+                continue
+            key = (e["model"], e["dataset"], e["seed"])
+            if key not in pidx:
+                continue
+            mag = _rot_mag(e.get("dists", {}).get(shift, {}).get("rotation"))
+            if mag != mag:
+                continue
+            rows.append({"concept": DATASETS[e["dataset"]].concept, "rot": mag, **pidx[key]})
+        if len(rows) < 8:
+            print(f"\n[{shift}] only {len(rows)} configs — skipped")
+            continue
+        _rank_predictors(rows, "rot", f"rotation under {shift.upper()} (n={len(rows)})")
+
+
+def estimator_cross():
+    """Do logreg-derived predictors also forecast the MASS-MEAN probe's rotation? (S2 external validity)."""
+    evals = _read_results("eval")
+    pidx = _pred_index(_read_results("predictors"))
+    print("\n########## ESTIMATOR EXTERNAL VALIDITY (predict mass-mean rotation) ##########")
+    rows = []
+    for e in evals:
+        if e.get("probe") != "mass_mean":
+            continue
+        key = (e["model"], e["dataset"], e["seed"])
+        if key not in pidx:
+            continue
+        rots = [_rot_mag(v.get("rotation")) for d, v in e.get("dists", {}).items() if d != "iid"]
+        rots = [r for r in rots if r == r]
+        if not rots:
+            continue
+        rows.append({"concept": DATASETS[e["dataset"]].concept, "rot": float(np.mean(rots)), **pidx[key]})
+    if len(rows) < 8:
+        print(f"only {len(rows)} configs — skipped")
+        return
+    _rank_predictors(rows, "rot", f"mass-mean rotation (n={len(rows)})")
+
+
+def size_ladder():
+    """Tier 3: does predictability of rotation / excess change with model size?
+    (C4 / inverse-scaling check). Per model, correlate key predictors vs paraphrase rotation
+    and vs excess rotation across its (dataset x seed) configs.
+    """
+    from config import MODELS
+    evals = _read_results("eval")
+    pidx = _pred_index(_read_results("predictors"))
+    keypreds = ["raptor_stability", "augmentation_robustness", "pac"]
+    # build per-model rows of (predictors, para_rot, excess)
+    by_model = {}
+    for e in evals:
+        if e.get("probe") != "logreg":
+            continue
+        key = (e["model"], e["dataset"], e["seed"])
+        if key not in pidx:
+            continue
+        para = _rot_mag(e.get("dists", {}).get("paraphrase", {}).get("rotation"))
+        iids = _rot_mag(e.get("iid_split_rotation"))
+        if para != para:
+            continue
+        excess = para - iids if iids == iids else float("nan")
+        by_model.setdefault(e["model"], []).append({"para": para, "excess": excess, **pidx[key]})
+    print("\n########## SIZE-LADDER: predictability vs model size ##########")
+    print(f"{'model':16s}{'params_M':>9s}  | rho(predictor, target) for para-rot / excess")
+    for m in sorted(by_model, key=lambda x: MODELS[x].params_m if x in MODELS else 0):
+        rows = by_model[m]
+        pm = MODELS[m].params_m if m in MODELS else 0
+        cells = []
+        for tgt in ("para", "excess"):
+            y = np.array([r[tgt] for r in rows], dtype=float)
+            for kp in keypreds:
+                s = np.array([r.get(kp, np.nan) for r in rows], dtype=float)
+                ok = ~(np.isnan(s) | np.isnan(y))
+                rho = metrics.spearman(s[ok], -y[ok])[0] if ok.sum() >= 6 else float("nan")
+                cells.append(f"{kp.split('_')[0][:4]}:{rho:+.2f}")
+            cells.append("|")
+        print(f"{m:16s}{pm:>9.0f}  | {' '.join(cells)}")
+    print("(target order: para-rot [raptor aug pac] | excess [raptor aug pac])")
+
+
+def _partial_spearman(x, y, z):
+    """Partial Spearman of (x,y) controlling for z: rank-transform then residualise on z."""
+    from scipy.stats import rankdata
+    ok = ~(np.isnan(x) | np.isnan(y) | np.isnan(z))
+    x, y, z = x[ok], y[ok], z[ok]
+    if len(x) < 8:
+        return float("nan"), 0
+    rx, ry, rz = rankdata(x), rankdata(y), rankdata(z)
+    Z = np.c_[np.ones_like(rz), rz]
+    res = lambda a: a - Z @ np.linalg.lstsq(Z, a, rcond=None)[0]
+    ex, ey = res(rx), res(ry)
+    return float(np.corrcoef(ex, ey)[0, 1]), len(x)
+
+
+def robust_circularity():
+    """Reviewer rebuttal (R2 difference-score artefact, R1 pseudo-replication):
+    (a) PARTIAL Spearman of (signal, paraphrase-rotation | placebo) — does a signal predict
+        shift-rotation BEYOND the sampling-noise floor, without the naive-minus-placebo subtraction?
+    (b) CLUSTER bootstrap (resampling CONCEPTS, not cells) for Δρ[aug−raptor] on excess.
+    """
+    evals = _read_results("eval")
+    pidx = _pred_index(_read_results("predictors"))
+    rows = []
+    for e in evals:
+        if e.get("probe") != "logreg":
+            continue
+        k = (e["model"], e["dataset"], e["seed"])
+        if k not in pidx:
+            continue
+        para = _rot_mag(e.get("dists", {}).get("paraphrase", {}).get("rotation"))
+        plac = _rot_mag(e.get("iid_split_rotation"))
+        if para != para or plac != plac:
+            continue
+        rows.append({"concept": DATASETS[e["dataset"]].concept, "para": para, "plac": plac,
+                     "excess": para - plac, **pidx[k]})
+    print("\n########## ROBUST CIRCULARITY CHECK (reviewer R1/R2 rebuttal) ##########")
+    print("PARTIAL Spearman rho(signal, paraphrase-rot | placebo)  [the artefact-free version]")
+    print(f"{'predictor':26s}{'partial-rho':>12s}{'naive-rho':>11s}{'excess-rho':>11s}")
+    para = np.array([r["para"] for r in rows]); plac = np.array([r["plac"] for r in rows])
+    exc = np.array([r["excess"] for r in rows])
+    for name in PREDICTORS:
+        s = np.array([r.get(name, np.nan) for r in rows], float)
+        pr, n = _partial_spearman(s, para, plac)
+        nr = metrics.spearman(s[~np.isnan(s)], -para[~np.isnan(s)])[0] if (~np.isnan(s)).sum() >= 8 else float("nan")
+        er = metrics.spearman(s[~np.isnan(s)], -exc[~np.isnan(s)])[0] if (~np.isnan(s)).sum() >= 8 else float("nan")
+        print(f"{name:26s}{-pr:>12.3f}{nr:>11.3f}{er:>11.3f}")
+    print("(partial-rho>0 => signal predicts shift-rotation beyond the sampling floor, non-circular,"
+          " no difference-score subtraction)")
+
+    # cluster bootstrap by concept for Delta rho[aug - raptor] on excess
+    concepts = np.array([r["concept"] for r in rows])
+    uc = list(np.unique(concepts))
+    sa = np.array([r.get("augmentation_robustness", np.nan) for r in rows], float)
+    sb = np.array([r.get("raptor_stability", np.nan) for r in rows], float)
+    rng = np.random.default_rng(0)
+    def drho(mask):
+        m = mask & ~(np.isnan(sa) | np.isnan(sb) | np.isnan(exc))
+        if m.sum() < 8:
+            return np.nan
+        return metrics.spearman(sa[m], -exc[m])[0] - metrics.spearman(sb[m], -exc[m])[0]
+    base = drho(np.ones(len(rows), bool))
+    boots = []
+    for _ in range(2000):
+        pick = rng.choice(uc, len(uc), replace=True)
+        mask = np.isin(concepts, pick)
+        # build resampled arrays by concept blocks
+        idx = np.concatenate([np.where(concepts == c)[0] for c in pick])
+        m = ~(np.isnan(sa[idx]) | np.isnan(sb[idx]) | np.isnan(exc[idx]))
+        if m.sum() >= 8:
+            boots.append(metrics.spearman(sa[idx][m], -exc[idx][m])[0] - metrics.spearman(sb[idx][m], -exc[idx][m])[0])
+    lo, hi = np.percentile(boots, [2.5, 97.5])
+    sig = "SIGNIFICANT" if lo > 0 else "n.s."
+    print(f"\nCLUSTER bootstrap (resample {len(uc)} CONCEPTS): Δρ[aug−raptor] on EXCESS = {base:+.3f}"
+          f"  95%CI [{lo:+.3f},{hi:+.3f}] -> {sig}")
+    print("(this is the pseudo-replication-corrected version of the headline paired test)")
+
+
+def fidelity_summary():
+    audit = _read_results("audit")
+    print("\n########## LABEL-FIDELITY (NLI flip-rate of paraphrase shift) ##########")
+    by_ds = {}
+    for a in audit:
+        by_ds.setdefault(a["dataset"], []).append(a.get("flip_rate", float("nan")))
+    print(f"{'dataset':18s}{'mean flip-rate':>16s}{'pass-rate':>12s}")
+    for ds in sorted(by_ds):
+        fr = float(np.nanmean(by_ds[ds]))
+        print(f"{ds:18s}{fr:>16.3f}{1 - fr:>12.3f}")
+    allfr = [v for vs in by_ds.values() for v in vs if v == v]
+    if allfr:
+        print(f"{'OVERALL':18s}{np.mean(allfr):>16.3f}{1 - np.mean(allfr):>12.3f}")
+
+
+if __name__ == "__main__":
+    mode = sys.argv[1] if len(sys.argv) > 1 else "all"
+    if mode in ("shift", "all"):
+        shift_breakdown()
+    if mode in ("estimator", "all"):
+        estimator_cross()
+    if mode in ("size", "all"):
+        size_ladder()
+    if mode in ("robust", "all"):
+        robust_circularity()
+    if mode in ("fidelity", "all"):
+        fidelity_summary()

repro_bundle/analyze_refit.py

@@ -0,0 +1,111 @@
+"""Tier-1 refit analyses (read cache, refit probes). Run on the server pointing PROBE_CACHE
+at a seed cache. Reads predictors from results/<predfile> (default predictors_1aug.jsonl, the
+A run, since Tier 2 rewrites predictors.jsonl).
+
+    PROBE_CACHE=/root/rivermind-fs/cache_seed0 python analyze_refit.py layer
+    PROBE_CACHE=/root/rivermind-fs/cache_seed0 python analyze_refit.py whitened [predfile]
+
+layer    : mean paraphrase rotation per (relative) layer over a sample — is the rotation a
+           property of the chosen layer or pervasive?  (defends "not layer-cherry-picked")
+whitened : recompute paraphrase rotation with ID-whitened (Mahalanobis) cosine and re-rank the
+           predictors — does the circularity / ranking conclusion survive a better metric? (M3)
+"""
+from __future__ import annotations
+
+import json
+import sys
+from pathlib import Path
+
+import numpy as np
+from scipy import linalg
+
+import cache
+import metrics
+from config import DATASETS, MODELS, EXTRACT, RESULTS_DIR
+from probes import make_probe
+from run_pipeline import _select_layer, _rank_predictors
+from baselines import PREDICTORS
+
+SAMPLE_DS = ["sst2", "ag_news", "counterfact", "emotion", "subj"]
+
+
+def _read(name):
+    p = RESULTS_DIR / name
+    return [json.loads(l) for l in p.read_text().splitlines() if l.strip()] if p.exists() else []
+
+
+def _dir(X, y, nl, W=None):
+    if W is not None:
+        X = X @ W
+    return make_probe("logreg", num_labels=nl, seed=0, max_iter=500).fit(X, y).direction
+
+
+def layer_curve(models=None, dsets=None):
+    models = models or list(MODELS)
+    dsets = dsets or SAMPLE_DS
+    print("\n########## LAYER ROBUSTNESS: paraphrase rotation by relative depth ##########")
+    buckets = {q: [] for q in [0.1, 0.25, 0.5, 0.75, 0.9, 1.0]}
+    for m in models:
+        for ds in dsets:
+            if not (cache.exists(m, ds, "train") and cache.exists(m, ds, "paraphrase")):
+                continue
+            nl = DATASETS[ds].num_labels
+            meta = cache.load_meta(m, ds, "train")
+            Xtr, ytr, _ = cache.load_shard(m, ds, "train")
+            Xpa, ypa, _ = cache.load_shard(m, ds, "paraphrase")
+            nl_layers = meta["n_layers"]
+            for q in buckets:
+                L = max(1, min(nl_layers - 1, int(q * (nl_layers - 1))))
+                da = _dir(np.asarray(Xtr[:, L, :], np.float32), ytr, nl)
+                db = _dir(np.asarray(Xpa[:, L, :], np.float32), ypa, nl)
+                buckets[q].append(_rotmag(da, db))
+    print(f"{'rel-depth':>10s}{'mean rotation (1-cos)':>24s}{'n':>5s}")
+    for q in sorted(buckets):
+        v = [x for x in buckets[q] if x == x]
+        if v:
+            print(f"{q:>10.2f}{np.mean(v):>24.3f}{len(v):>5d}")
+    print("(rotation present across depths -> not an artefact of the IID-selected layer)")
+
+
+def _rotmag(da, db):
+    try:
+        return 1.0 - metrics.mean_class_cosine(da, db)
+    except Exception:
+        return 1.0 - float(np.cos(metrics.subspace_principal_angle(da, db)))
+
+
+def whitened_target(predfile="predictors_1aug.jsonl"):
+    evals = _read("eval.jsonl")
+    preds = {(p["model"], p["dataset"], p["seed"]): p["predictors"] for p in _read(predfile)}
+    print(f"\n########## METRIC ABLATION: ID-whitened paraphrase rotation as target ##########")
+    print(f"(predictors from {predfile}; recomputing whitened rotation by refit on cache)")
+    rows = []
+    for e in evals:
+        if e.get("probe") != "logreg":
+            continue
+        m, ds, seed = e["model"], e["dataset"], e["seed"]
+        if (m, ds, seed) not in preds:
+            continue
+        if not (cache.exists(m, ds, "train") and cache.exists(m, ds, "paraphrase")):
+            continue
+        nl = DATASETS[ds].num_labels
+        layer = e.get("layer") or _select_layer(m, ds, nl, seed)
+        Xtr, ytr, _ = cache.load_shard(m, ds, "train", layer=layer)
+        Xpa, ypa, _ = cache.load_shard(m, ds, "paraphrase", layer=layer)
+        C = np.cov(np.asarray(Xtr, np.float64), rowvar=False) + 1e-3 * np.eye(Xtr.shape[1])
+        W = linalg.fractional_matrix_power(C, -0.5).real.astype(np.float32)
+        da = _dir(Xtr, ytr, nl, W=W)
+        db = _dir(Xpa, ypa, nl, W=W)
+        rows.append({"concept": DATASETS[ds].concept, "wrot": _rotmag(da, db), **preds[(m, ds, seed)]})
+    if len(rows) < 8:
+        print(f"only {len(rows)} configs — skipped")
+        return
+    _rank_predictors(rows, "wrot", f"WHITENED paraphrase rotation (n={len(rows)})")
+
+
+if __name__ == "__main__":
+    mode = sys.argv[1] if len(sys.argv) > 1 else "layer"
+    if mode == "layer":
+        layer_curve()
+    elif mode == "whitened":
+        whitened_target(sys.argv[2] if len(sys.argv) > 2 else "predictors_1aug.jsonl")

repro_bundle/auto_pipeline.sh

@@ -0,0 +1,24 @@
+#!/bin/bash
+# Server-side autonomous orchestrator: advances the pipeline WITHOUT depending on the
+# laptop-side monitoring. Waits for the A run to finish, then auto-produces Tier-0 stats
+# and Tier-1 re-analyses. Each milestone writes a marker to auto.log.
+cd /root/rivermind-data/probe_stability
+AUTO=/root/rivermind-data/auto.log
+export OMP_NUM_THREADS=4 OPENBLAS_NUM_THREADS=4 MKL_NUM_THREADS=4
+
+echo "AUTO_START $(date +%F_%H:%M:%S)" >> $AUTO
+# 1) wait for the multi-seed A run to finish
+while ! grep -q A_DONE /root/rivermind-data/A.log 2>/dev/null; do sleep 60; done
+echo "A_DONE_SEEN $(date +%F_%H:%M:%S)" >> $AUTO
+
+# 2) Tier 0: final 4-target tables + paired significance (re-run stats cleanly to a file)
+python3 -u run_pipeline.py stats > /root/rivermind-data/tier0_stats.txt 2>&1
+echo "TIER0_STATS_DONE $(date +%F_%H:%M:%S)" >> $AUTO
+
+# 3) Tier 1: shift-type breakdown + estimator cross-validity + label-fidelity
+python3 -u analyze.py all > /root/rivermind-data/tier1.txt 2>&1
+echo "TIER1_DONE $(date +%F_%H:%M:%S)" >> $AUTO
+
+# 4) snapshot results for safety
+cp -r results /root/rivermind-data/results_snapshot_A 2>/dev/null
+echo "AUTO_COMPLETE $(date +%F_%H:%M:%S)" >> $AUTO

repro_bundle/baselines.py

@@ -0,0 +1,153 @@
+"""A-priori, label-free probe-stability predictors evaluated by ProbeShift (G2).
+
+Every predictor maps IID training activations (+ labels, + optional augmentation
+activations) to a scalar where **higher = predicted more OOD-stable**, so all can be
+correlated against (negated) OOD drop under one convention. All are computed WITHOUT any
+OOD data and WITHOUT new annotation, except `whitened_cosine_upper` which is flagged as an
+OOD-using upper-bound reference (Truthfulness-Spectrum style).
+
+Implemented (each a verified competitor — see PRIOR_ART.md §四):
+  sip_eigengap            SIP (2511.16288)        — LDA eigengap / Fisher error
+  raptor_stability        RAPTOR (2602.00158)     — bootstrap direction mean|cos|  (= dispersion)
+  fragility               Fragility (2606.11375)  — critical isotropic-noise collapse sigma
+  augmentation_robustness Probing-the-Probes      — direction consistency under augmentation
+  xie_feature_dispersion  Xie 2023 (2303.15488)   — inter-class feature dispersion
+  pac                     ours (G3)               — composite of dispersion + augmentation
+  whitened_cosine_upper   Truth-Spectrum style    — ID/OOD-whitened cosine (upper-bound ref)
+"""
+from __future__ import annotations
+
+import numpy as np
+from scipy import linalg
+
+from config import STABILITY
+from probes import make_probe
+from metrics import accuracy
+from stability_score import dispersion, augmentation_consistency, combine
+
+
+# --------------------------------------------------------------------------------------
+# SIP — spectral identifiability (eigengap of the task-relevant LDA subspace)
+# --------------------------------------------------------------------------------------
+def sip_eigengap(X: np.ndarray, y: np.ndarray, num_labels: int, ridge: float = 1e-3, **_):
+    X = np.asarray(X, dtype=np.float64)
+    classes = np.unique(y)
+    mu = X.mean(0)
+    Sb = np.zeros((X.shape[1], X.shape[1]))
+    Sw = np.zeros_like(Sb)
+    for c in classes:
+        Xc = X[y == c]
+        d = (Xc.mean(0) - mu)[:, None]
+        Sb += len(Xc) * (d @ d.T)
+        Sw += np.cov(Xc, rowvar=False) * (len(Xc) - 1)
+    Sw += ridge * np.eye(X.shape[1])
+    # generalised eigenvalues of (Sb, Sw); top (num_labels-1) span the relevant subspace
+    eig = np.sort(linalg.eigvalsh(Sb, Sw))[::-1]
+    k = max(1, num_labels - 1)
+    if len(eig) <= k:
+        return float("nan")
+    gap = eig[k - 1] - eig[k]                      # gap after the relevant subspace
+    fisher_err = np.sqrt(X.shape[1] / len(X))      # ~ estimation error scale
+    return float(gap / (eig[0] + 1e-12) / (fisher_err + 1e-12))
+
+
+# --------------------------------------------------------------------------------------
+# RAPTOR-style directional stability == 1 - dispersion
+# --------------------------------------------------------------------------------------
+def raptor_stability(X, y, num_labels, probe_kind="logreg", seed=0, **kw):
+    disp = dispersion(np.asarray(X, np.float32), np.asarray(y), num_labels,
+                      probe_kind=probe_kind, k=STABILITY.k_bootstrap, seed=seed, **kw)
+    return float(1.0 - disp)
+
+
+# --------------------------------------------------------------------------------------
+# Fragility — critical isotropic-noise collapse level
+# --------------------------------------------------------------------------------------
+def fragility(X, y, num_labels, probe_kind="logreg", seed=0,
+              sigmas=(0.0, 0.25, 0.5, 1.0, 1.5, 2.0, 3.0), **kw):
+    X = np.asarray(X, dtype=np.float32)
+    rng = np.random.default_rng(seed)
+    p = make_probe(probe_kind, num_labels=num_labels, seed=seed, **kw).fit(X, y)
+    iid_acc = accuracy(y, p.predict(X))
+    chance = 1.0 / num_labels
+    thresh = chance + 0.5 * (iid_acc - chance)     # halfway from chance to IID
+    scale = X.std()
+    critical = sigmas[-1]
+    for s in sigmas:
+        Xn = X + s * scale * rng.standard_normal(X.shape).astype(np.float32)
+        if accuracy(y, p.predict(Xn)) < thresh:
+            critical = s
+            break
+    return float(critical)                          # higher = more robust = more stable
+
+
+# --------------------------------------------------------------------------------------
+# augmentation-robustness == augmentation_consistency component
+# --------------------------------------------------------------------------------------
+def augmentation_robustness(X, y, aug_X_list, num_labels, probe_kind="logreg", seed=0, **kw):
+    return float(augmentation_consistency(
+        np.asarray(X, np.float32), np.asarray(y), [np.asarray(a, np.float32) for a in aug_X_list],
+        num_labels, probe_kind=probe_kind, seed=seed, **kw))
+
+
+# --------------------------------------------------------------------------------------
+# Xie 2023 — inter-class feature dispersion (predicts overall acc, mechanism contrast)
+# --------------------------------------------------------------------------------------
+def xie_feature_dispersion(X, y, num_labels, **_):
+    X = np.asarray(X, dtype=np.float64)
+    classes = np.unique(y)
+    mus = np.stack([X[y == c].mean(0) for c in classes])         # [C,H]
+    inter = np.mean([np.linalg.norm(mus[i] - mus[j])
+                     for i in range(len(mus)) for j in range(i + 1, len(mus))])
+    return float(inter / (X.std() + 1e-12))
+
+
+# --------------------------------------------------------------------------------------
+# PAC (ours, G3) — composite of dispersion + augmentation-consistency (IID-only)
+# --------------------------------------------------------------------------------------
+def pac(X, y, aug_X_list, num_labels, probe_kind="logreg", seed=0, **kw):
+    disp = dispersion(np.asarray(X, np.float32), np.asarray(y), num_labels,
+                      probe_kind=probe_kind, k=STABILITY.k_bootstrap, seed=seed, **kw)
+    aug = augmentation_consistency(
+        np.asarray(X, np.float32), np.asarray(y),
+        [np.asarray(a, np.float32) for a in aug_X_list], num_labels,
+        probe_kind=probe_kind, seed=seed, **kw)
+    return float(combine(disp, aug, STABILITY))
+
+
+# --------------------------------------------------------------------------------------
+# Whitened-cosine upper bound (Truth-Spectrum style) — uses covariance; flagged.
+# --------------------------------------------------------------------------------------
+def whitened_cosine_upper(X, y, num_labels, probe_kind="logreg", seed=0, cov=None, **kw):
+    """If `cov` (e.g. OOD covariance) is supplied, this becomes the OOD-using upper bound;
+    with cov=None it whitens by ID covariance (the metric-ablation variant)."""
+    X = np.asarray(X, dtype=np.float64)
+    C = np.cov(X, rowvar=False) if cov is None else np.asarray(cov, np.float64)
+    C += 1e-3 * np.eye(C.shape[0])
+    W = linalg.fractional_matrix_power(C, -0.5).real
+    return raptor_stability((X @ W).astype(np.float32), y, num_labels,
+                            probe_kind=probe_kind, seed=seed, **kw)
+
+
+# Registry: name -> (callable, needs_aug). All are higher = more stable.
+PREDICTORS = {
+    "sip_eigengap": (sip_eigengap, False),
+    "raptor_stability": (raptor_stability, False),
+    "fragility": (fragility, False),
+    "augmentation_robustness": (augmentation_robustness, True),
+    "xie_feature_dispersion": (xie_feature_dispersion, False),
+    "pac": (pac, True),
+    "whitened_cosine_id": (whitened_cosine_upper, False),
+}
+
+
+def compute_all(X_iid, y_iid, aug_X_list, num_labels, probe_kind="logreg", seed=0) -> dict:
+    out = {}
+    for name, (fn, needs_aug) in PREDICTORS.items():
+        try:
+            out[name] = fn(X_iid, y_iid, aug_X_list, num_labels, probe_kind=probe_kind, seed=seed) \
+                if needs_aug else fn(X_iid, y_iid, num_labels, probe_kind=probe_kind, seed=seed)
+        except Exception as e:                      # keep the grid going; record the failure
+            out[name] = float("nan")
+            out[name + "__error"] = str(e)
+    return out

repro_bundle/cache.py

@@ -0,0 +1,89 @@
+"""Activation-shard cache.
+
+One shard = activations for a single (model, dataset, distribution) triple.
+Stored as a directory of `.npy` files so that the (potentially multi-GB) activation
+tensor can be memory-mapped and sliced by layer without loading everything into RAM.
+
+Layout:
+    cache/<model_key>/<dataset_key>/<distribution>/
+        acts.npy     float16  [N, L+1, H]   (L+1 = embeddings + L transformer layers)
+        labels.npy   int64    [N]
+        ids.npy      int64    [N]           (stable example ids, for cross-shard alignment)
+        meta.json    {model, dataset, distribution, pooling, n, n_layers, hidden, dtype}
+"""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import numpy as np
+
+from config import CACHE_DIR
+
+
+def shard_dir(model_key: str, dataset_key: str, distribution: str) -> Path:
+    return CACHE_DIR / model_key / dataset_key / distribution
+
+
+def exists(model_key: str, dataset_key: str, distribution: str) -> bool:
+    d = shard_dir(model_key, dataset_key, distribution)
+    return (d / "acts.npy").exists() and (d / "meta.json").exists()
+
+
+def save_shard(
+    model_key: str,
+    dataset_key: str,
+    distribution: str,
+    acts: np.ndarray,        # [N, L+1, H] float16
+    labels: np.ndarray,      # [N]
+    ids: np.ndarray,         # [N]
+    pooling: str,
+) -> Path:
+    d = shard_dir(model_key, dataset_key, distribution)
+    d.mkdir(parents=True, exist_ok=True)
+    acts = np.ascontiguousarray(acts.astype(np.float16))
+    np.save(d / "acts.npy", acts)
+    np.save(d / "labels.npy", labels.astype(np.int64))
+    np.save(d / "ids.npy", ids.astype(np.int64))
+    meta = {
+        "model": model_key,
+        "dataset": dataset_key,
+        "distribution": distribution,
+        "pooling": pooling,
+        "n": int(acts.shape[0]),
+        "n_layers": int(acts.shape[1]),   # includes embedding layer
+        "hidden": int(acts.shape[2]),
+        "dtype": "float16",
+    }
+    (d / "meta.json").write_text(json.dumps(meta, indent=2))
+    return d
+
+
+def load_meta(model_key: str, dataset_key: str, distribution: str) -> dict:
+    return json.loads((shard_dir(model_key, dataset_key, distribution) / "meta.json").read_text())
+
+
+def load_shard(
+    model_key: str,
+    dataset_key: str,
+    distribution: str,
+    layer: int | None = None,
+    mmap: bool = True,
+) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
+    """Return (acts, labels, ids).
+
+    If `layer` is given, acts is [N, H] for that layer (0 = embeddings, 1..L = layers).
+    Otherwise acts is the full [N, L+1, H] tensor (mmap'd by default — do not mutate).
+    """
+    d = shard_dir(model_key, dataset_key, distribution)
+    mode = "r" if mmap else None
+    acts = np.load(d / "acts.npy", mmap_mode=mode)
+    labels = np.load(d / "labels.npy")
+    ids = np.load(d / "ids.npy")
+    if layer is not None:
+        acts = np.asarray(acts[:, layer, :], dtype=np.float32)   # materialise one layer
+        # sanitise fp16-overflow inf/nan (e.g. Qwen "massive activations") -> finite.
+        # No-op for models without overflow (pythia/gpt2), so existing results are unchanged.
+        if not np.isfinite(acts).all():
+            acts = np.nan_to_num(acts, nan=0.0, posinf=65504.0, neginf=-65504.0)
+    return acts, labels, ids

repro_bundle/config.py

@@ -0,0 +1,198 @@
+"""Central configuration: model registry, dataset registry, shift specs, paths, hyper-params.
+
+Importing this module is dependency-light (no torch / transformers) so it can be used by
+analysis scripts on a laptop. Heavy imports live in extract_activations.py / data/*.
+"""
+from __future__ import annotations
+
+import os
+from dataclasses import dataclass, field
+from pathlib import Path
+
+# --------------------------------------------------------------------------------------
+# Paths
+# --------------------------------------------------------------------------------------
+ROOT = Path(__file__).resolve().parent
+CACHE_DIR = Path(os.environ.get("PROBE_CACHE", ROOT / "cache"))        # activation shards
+RESULTS_DIR = Path(os.environ.get("PROBE_RESULTS", ROOT / "results"))  # metrics, tables
+DATA_DIR = Path(os.environ.get("PROBE_DATA", ROOT / "data_cache"))     # shifted text cache
+
+for _d in (CACHE_DIR, RESULTS_DIR, DATA_DIR):
+    _d.mkdir(parents=True, exist_ok=True)
+
+
+# --------------------------------------------------------------------------------------
+# Models — the size ladder (claim C4). Keys are short ids used in shard filenames.
+# --------------------------------------------------------------------------------------
+@dataclass(frozen=True)
+class ModelSpec:
+    key: str            # short id, used on disk (no slashes)
+    hf_name: str        # HF hub id
+    family: str         # for grouping in mixed-effects / size ladder
+    params_m: float     # parameter count in millions (for the scale axis)
+
+
+MODELS: dict[str, ModelSpec] = {
+    m.key: m for m in [
+        ModelSpec("pythia-70m", "EleutherAI/pythia-70m", "pythia", 70),
+        ModelSpec("pythia-160m", "EleutherAI/pythia-160m", "pythia", 160),
+        ModelSpec("pythia-410m", "EleutherAI/pythia-410m", "pythia", 410),
+        ModelSpec("pythia-1.4b", "EleutherAI/pythia-1.4b", "pythia", 1400),
+        ModelSpec("pythia-6.9b", "EleutherAI/pythia-6.9b", "pythia", 6900),   # Tier 3: 7B spot-check
+        ModelSpec("gpt2", "gpt2", "gpt2", 124),
+        ModelSpec("gpt2-medium", "gpt2-medium", "gpt2", 355),
+        ModelSpec("qwen2.5-0.5b", "Qwen/Qwen2.5-0.5B", "qwen", 500),
+    ]
+}
+
+# Default ladder for the headline scale experiment (cheapest -> most expensive).
+SIZE_LADDER = ["pythia-70m", "pythia-160m", "pythia-410m", "pythia-1.4b"]
+
+
+# --------------------------------------------------------------------------------------
+# Datasets — concept-bearing classification tasks (claim families C1/C3).
+# `concept` groups datasets that probe the SAME underlying concept (for the Spearman
+# generalisation analysis across concept families).
+# --------------------------------------------------------------------------------------
+@dataclass(frozen=True)
+class DatasetSpec:
+    key: str
+    hf_path: str
+    hf_config: str | None
+    split: str
+    text_field: str
+    label_field: str
+    concept: str            # "sentiment" | "topic" | "pos" | "truth"
+    num_labels: int
+    # Domain-shift partner: another DatasetSpec key whose inputs share the label space
+    # but come from a different domain (used as the `domain` OOD distribution).
+    domain_partner: str | None = None
+
+
+#
+# NOTE: datasets>=5.0 dropped legacy short names AND script-based datasets (no more
+# trust_remote_code). Use fully-qualified `namespace/name` Hub ids that ship parquet.
+# Verified loadable 2026-06-22: nyu-mll/glue, stanfordnlp/imdb, fancyzhx/ag_news,
+# fancyzhx/dbpedia_14, fancyzhx/amazon_polarity.
+DATASETS: dict[str, DatasetSpec] = {
+    d.key: d for d in [
+        # --- sentiment concept ---
+        DatasetSpec("sst2", "nyu-mll/glue", "sst2", "train", "sentence", "label", "sentiment", 2,
+                    domain_partner="imdb"),
+        DatasetSpec("imdb", "stanfordnlp/imdb", None, "train", "text", "label", "sentiment", 2,
+                    domain_partner="sst2"),
+        DatasetSpec("amazon_polarity", "fancyzhx/amazon_polarity", None, "train", "content",
+                    "label", "sentiment", 2, domain_partner="sst2"),
+        # --- topic concept ---
+        DatasetSpec("ag_news", "fancyzhx/ag_news", None, "train", "text", "label", "topic", 4,
+                    domain_partner="dbpedia"),
+        DatasetSpec("dbpedia", "fancyzhx/dbpedia_14", None, "train", "content", "label", "topic", 14),
+        # --- syntactic concept (POS) ---
+        # WARNING: universal_dependencies is script-based -> NOT loadable under datasets>=5.0.
+        # TODO(4090): swap to a parquet POS source (e.g. a CoNLL-2003 POS parquet) before using.
+        DatasetSpec("ud_pos", "universal-dependencies/universal_dependencies", "en_ewt", "train",
+                    "tokens", "upos", "pos", 17),
+        # --- truthfulness / factual concept (derived, label from existing fields) ---
+        # TODO(4090): truthful_qa generation may be script-based; counterfact-tracing ships parquet.
+        DatasetSpec("truthfulqa", "truthfulqa/truthful_qa", "generation", "validation", "question",
+                    "label", "truth", 2),
+        DatasetSpec("counterfact", "NeelNanda/counterfact-tracing", None, "train", "prompt",
+                    "label", "truth", 2),
+        # --- extra concepts for cross-concept (LOCO) power; all parquet, verified 2026-06-22 ---
+        DatasetSpec("emotion", "dair-ai/emotion", None, "train", "text", "label", "emotion", 6),
+        DatasetSpec("tweet_hate", "cardiffnlp/tweet_eval", "hate", "train", "text", "label",
+                    "hate", 2),
+        DatasetSpec("tweet_irony", "cardiffnlp/tweet_eval", "irony", "train", "text", "label",
+                    "irony", 2),
+        DatasetSpec("tweet_offensive", "cardiffnlp/tweet_eval", "offensive", "train", "text",
+                    "label", "offensive", 2),
+        DatasetSpec("subj", "SetFit/subj", None, "train", "text", "label", "subjectivity", 2),
+        DatasetSpec("spam", "ucirvine/sms_spam", "plain_text", "train", "sms", "label", "spam", 2),
+        # --- 3 more concepts -> 12 total (verified loadable 2026-06-22) ---
+        DatasetSpec("cola", "nyu-mll/glue", "cola", "train", "sentence", "label", "grammaticality", 2),
+        DatasetSpec("stance", "cardiffnlp/tweet_eval", "stance_feminist", "train", "text", "label",
+                    "stance", 3),
+        DatasetSpec("amazon_cf", "mteb/amazon_counterfactual", "en", "train", "text", "label",
+                    "counterfactual", 2),
+    ]
+}
+
+
+# --------------------------------------------------------------------------------------
+# Distribution shifts (label-preserving). Each (dataset, distribution) pair is one
+# activation shard. "iid" is the in-distribution train/test of the dataset itself.
+# --------------------------------------------------------------------------------------
+SHIFTS = ["iid", "paraphrase", "domain", "length"]
+
+# Back-translation pivot for the `paraphrase` SHIFT (rotation ground truth) — keep single &
+# fixed (de) so rotation measurements are consistent across runs.
+BT_PIVOTS = ["de"]
+
+# Tier 2: the AUGMENTATION predictor uses MULTIPLE pivots for genuine paraphrastic diversity
+# (de/fr/ru). With PROBE_N_AUG=3 each aug uses a different pivot, so augmentation-consistency
+# reflects robustness to varied surface forms (not the single-pivot artefact of the 1-aug run).
+AUG_PIVOTS = ["de", "fr", "ru"]
+
+# Length buckets (token counts) for the `length` shift: train on SHORT, eval on LONG.
+LENGTH_BUCKETS = {"short": (0, 32), "long": (96, 512)}
+
+
+# --------------------------------------------------------------------------------------
+# Extraction hyper-params
+# --------------------------------------------------------------------------------------
+@dataclass
+class ExtractConfig:
+    pooling: str = "mean"          # "mean" (masked) | "last" (last non-pad token)
+    max_length: int = 256
+    batch_size: int = 32           # lower for pythia-1.4b on 24GB
+    dtype: str = "float16"
+    n_train: int = 2000            # examples used to FIT probes per seed (subsampled from pool)
+    n_eval: int = 1000             # eval examples per (dataset, distribution) per seed (from pool)
+    n_train_pool: int = 4000       # examples EXTRACTED for train (multi-seed subsamples from this)
+    n_eval_pool: int = 2000        # examples EXTRACTED for eval/shifts (subsampled per seed)
+    device: str = "cuda"
+
+
+# --------------------------------------------------------------------------------------
+# Probe / Stability-Score / stats hyper-params
+# --------------------------------------------------------------------------------------
+@dataclass
+class ProbeConfig:
+    probe_types: tuple[str, ...] = ("logreg", "mass_mean", "mlp")
+    l2: float = 1.0                # inverse-C handled inside probes.py
+    mlp_hidden: int = 128
+    mlp_epochs: int = 50
+    layer_selection: str = "iid_val"   # pick layer by IID val acc, never post-hoc on OOD
+
+
+@dataclass
+class StabilityConfig:
+    k_bootstrap: int = 8           # K resamples for dispersion (multi-seed compensates)
+    n_aug: int = 5                 # label-preserving augmentations for consistency
+    # combine(dispersion, aug_consistency) -> scalar score; weights are a hyper-param we
+    # report sensitivity for. Pre-registered hypothesis: combination > dispersion alone.
+    w_dispersion: float = 0.5
+    w_consistency: float = 0.5
+
+
+@dataclass
+class StatsConfig:
+    seeds: tuple[int, ...] = (0, 1, 2, 3, 4)   # >=5; covers init / resample / augmentation
+    n_bootstrap: int = 1000
+    fdr_method: str = "holm"       # "holm" | "bh"
+
+
+# Singletons used across the pipeline (override per-run in run_pipeline.py if needed).
+EXTRACT = ExtractConfig()
+PROBE = ProbeConfig()
+STABILITY = StabilityConfig()
+STATS = StatsConfig()
+
+# Env overrides — let a small pilot run cheaply without editing code, e.g.
+#   PROBE_N_TRAIN=600 PROBE_N_EVAL=400 PROBE_N_AUG=2 PROBE_BATCH=64 python run_pipeline.py ...
+EXTRACT.n_train = int(os.environ.get("PROBE_N_TRAIN", EXTRACT.n_train))
+EXTRACT.n_eval = int(os.environ.get("PROBE_N_EVAL", EXTRACT.n_eval))
+EXTRACT.n_train_pool = int(os.environ.get("PROBE_N_TRAIN_POOL", EXTRACT.n_train_pool))
+EXTRACT.n_eval_pool = int(os.environ.get("PROBE_N_EVAL_POOL", EXTRACT.n_eval_pool))
+EXTRACT.batch_size = int(os.environ.get("PROBE_BATCH", EXTRACT.batch_size))
+STABILITY.n_aug = int(os.environ.get("PROBE_N_AUG", STABILITY.n_aug))

repro_bundle/data/__init__.py

@@ -0,0 +1 @@
+"""Data subpackage: dataset loading, label-preserving shifts, label-fidelity audit."""

repro_bundle/data/datasets.py

@@ -0,0 +1,121 @@
+"""Load HF datasets into a unified record format.
+
+A loaded split is a dict:
+    {"texts": list[str], "labels": np.ndarray[int], "ids": np.ndarray[int]}
+
+`ids` are stable per (dataset, split) so the SAME example can be aligned across the
+iid / paraphrase / length shards (paraphrase keeps the id of its source).
+
+Standard text-classification datasets are loaded directly. POS / truthfulness datasets
+are *derived* into a probing format (see builder functions). Derived builders are marked
+TODO where a modelling choice should be revisited on the 4090.
+"""
+from __future__ import annotations
+
+import numpy as np
+
+from config import DATASETS, DatasetSpec
+
+
+def _subsample(texts, labels, n, seed):
+    rng = np.random.default_rng(seed)
+    n = min(n, len(texts))
+    # stratified-ish: shuffle then take n (good enough; balance checked downstream)
+    idx = rng.permutation(len(texts))[:n]
+    return [texts[i] for i in idx], np.asarray(labels)[idx], idx.astype(np.int64)
+
+
+# --------------------------------------------------------------------------------------
+# Builders for derived / non-trivial datasets
+# --------------------------------------------------------------------------------------
+def _build_ud_pos(spec, split, n, seed):
+    """Word-in-context POS probe: pick one target token per sentence, mark it with [],
+    label = its UPOS tag. Simple, deterministic target selection (first content word).
+    TODO(4090): consider sampling multiple target positions per sentence for more data.
+    """
+    from datasets import load_dataset
+    ds = load_dataset(spec.hf_path, spec.hf_config, split=split)
+    texts, labels = [], []
+    for ex in ds:
+        toks, tags = ex["tokens"], ex["upos"]
+        if not toks:
+            continue
+        j = min(range(len(toks)), key=lambda i: (tags[i] in (0,), i))  # first non-punct-ish
+        marked = " ".join(toks[:j] + ["[" + toks[j] + "]"] + toks[j + 1:])
+        texts.append(marked)
+        labels.append(int(tags[j]))
+    return _subsample(texts, labels, n, seed)
+
+
+def _build_truthfulqa(spec, split, n, seed):
+    """Derive a binary truth probe from TruthfulQA: (question + correct answer) -> 1,
+    (question + incorrect answer) -> 0. Label comes from existing fields (no annotation).
+    """
+    from datasets import load_dataset
+    ds = load_dataset(spec.hf_path, spec.hf_config, split=split)
+    texts, labels = [], []
+    for ex in ds:
+        q = ex["question"]
+        for a in ex.get("correct_answers", [])[:1]:
+            texts.append(f"Q: {q} A: {a}"); labels.append(1)
+        for a in ex.get("incorrect_answers", [])[:1]:
+            texts.append(f"Q: {q} A: {a}"); labels.append(0)
+    return _subsample(texts, labels, n, seed)
+
+
+def _build_counterfact(spec, split, n, seed):
+    """Derive true/false statements from counterfact-tracing prompt + target_true/false."""
+    from datasets import load_dataset
+    ds = load_dataset(spec.hf_path, split=split)
+    texts, labels = [], []
+    for ex in ds:
+        prompt = ex.get("prompt") or ex.get("relation_prefix", "")
+        t, f = ex.get("target_true"), ex.get("target_false")
+        if t:
+            texts.append(f"{prompt} {t}".strip()); labels.append(1)
+        if f:
+            texts.append(f"{prompt} {f}".strip()); labels.append(0)
+    return _subsample(texts, labels, n, seed)
+
+
+_DERIVED = {
+    "ud_pos": _build_ud_pos,
+    "truthfulqa": _build_truthfulqa,
+    "counterfact": _build_counterfact,
+}
+
+
+# --------------------------------------------------------------------------------------
+# Public API
+# --------------------------------------------------------------------------------------
+def load(dataset_key: str, n: int, split: str | None = None, seed: int = 0) -> dict:
+    spec: DatasetSpec = DATASETS[dataset_key]
+    split = split or spec.split
+
+    if dataset_key in _DERIVED:
+        texts, labels, ids = _DERIVED[dataset_key](spec, split, n, seed)
+    else:
+        from datasets import load_dataset
+        ds = load_dataset(spec.hf_path, spec.hf_config, split=split)
+        texts = list(ds[spec.text_field])
+        labels = list(ds[spec.label_field])
+        texts, labels, ids = _subsample(texts, labels, n, seed)
+
+    return {"texts": texts, "labels": np.asarray(labels, dtype=np.int64), "ids": ids}
+
+
+def load_train_eval(dataset_key: str, n_train: int, n_eval: int, seed: int = 0
+                    ) -> tuple[dict, dict]:
+    """Disjoint IID train / eval draws from the dataset's train split.
+
+    If the dataset is smaller than n_train + n_eval, split PROPORTIONALLY so eval is never
+    empty (small datasets like tweet_eval/stance_feminist have <1500 rows).
+    """
+    full = load(dataset_key, n=n_train + n_eval, split=None, seed=seed)
+    n = len(full["texts"])
+    cut = n_train if n >= n_train + n_eval else max(1, int(n * n_train / (n_train + n_eval)))
+    tr = {"texts": full["texts"][:cut], "labels": full["labels"][:cut],
+          "ids": full["ids"][:cut]}
+    ev = {"texts": full["texts"][cut:], "labels": full["labels"][cut:],
+          "ids": full["ids"][cut:]}
+    return tr, ev

repro_bundle/data/label_fidelity.py

@@ -0,0 +1,66 @@
+"""Label-fidelity audit (claim C5).
+
+A shift is only valid if it is *label-preserving*. We verify this WITHOUT new human
+annotation using a local NLI model (DeBERTa-v3-MNLI): a paraphrase should be mutually
+entailing with its source; we drop pairs that flip. We report the flip rate so reviewers
+can see the shifts are clean.
+
+This is the standard "are your counterfactuals actually label-preserving?" defence.
+Zero API cost — the NLI model runs locally.
+"""
+from __future__ import annotations
+
+import numpy as np
+
+_NLI = None
+_NLI_LABELS = None   # index order of {entailment, neutral, contradiction}
+
+
+def _get_nli(model_name: str = "MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli"):
+    global _NLI, _NLI_LABELS
+    if _NLI is None:
+        import torch
+        from transformers import AutoModelForSequenceClassification, AutoTokenizer
+        tok = AutoTokenizer.from_pretrained(model_name)
+        model = AutoModelForSequenceClassification.from_pretrained(model_name)
+        if torch.cuda.is_available():
+            model = model.to("cuda").half()
+        model.eval()
+        _NLI = (tok, model)
+        _NLI_LABELS = model.config.id2label   # {0:'entailment',...} varies by checkpoint
+    return _NLI
+
+
+def _entail_prob(premises: list[str], hypotheses: list[str], batch_size: int = 32) -> np.ndarray:
+    """P(entailment) for each (premise, hypothesis) pair."""
+    import torch
+    tok, model = _get_nli()
+    device = next(model.parameters()).device
+    ent_idx = [i for i, v in _NLI_LABELS.items() if "entail" in v.lower()][0]
+    probs = []
+    for i in range(0, len(premises), batch_size):
+        enc = tok(premises[i:i + batch_size], hypotheses[i:i + batch_size],
+                  return_tensors="pt", padding=True, truncation=True, max_length=256).to(device)
+        with torch.no_grad():
+            logits = model(**enc).logits.float()
+        p = torch.softmax(logits, dim=-1)[:, ent_idx]
+        probs.extend(p.cpu().numpy().tolist())
+    return np.asarray(probs)
+
+
+def paraphrase_keep_mask(originals: list[str], shifted: list[str], thresh: float = 0.5
+                         ) -> np.ndarray:
+    """Boolean mask: keep pairs that are mutually entailing (label preserved).
+
+    TODO(4090): tune `thresh` on a small dev set of known paraphrases; report flip rate
+    and a sensitivity curve over thresh in the paper appendix.
+    """
+    fwd = _entail_prob(originals, shifted)
+    bwd = _entail_prob(shifted, originals)
+    keep = (fwd >= thresh) & (bwd >= thresh)
+    return keep
+
+
+def audit(originals: list[str], shifted: list[str], thresh: float = 0.5) -> dict:
+    keep = paraphrase_keep_mask(originals, shifted, thresh)
+    return {"keep_mask": keep, "flip_rate": float(1.0 - keep.mean()), "n": len(originals)}

repro_bundle/data/shifts.py

@@ -0,0 +1,146 @@
+"""Label-preserving distribution shifts.
+
+Each shift maps an IID record dict -> a new record dict with the SAME labels/ids but
+shifted *inputs*. The probe is trained on IID train; these produce the OOD eval sets
+(claim C1). They are also reused as the augmentations for augmentation_consistency
+(claim C2) — augmentation == a mild label-preserving shift of the TRAIN set.
+
+Shifts implemented:
+  * paraphrase : round-trip back-translation (Marian opus-mt en->pivot->en).
+  * domain     : swap in the `domain_partner` dataset (same label space).
+  * length     : keep only the LONG token-length bucket (train uses SHORT bucket).
+
+All shifted text is cached to disk (data_cache/) so the expensive back-translation runs
+once and is $0 thereafter (local models, zero API).
+"""
+from __future__ import annotations
+
+import hashlib
+import json
+from pathlib import Path
+
+import numpy as np
+
+from config import AUG_PIVOTS, BT_PIVOTS, DATASETS, DATA_DIR, LENGTH_BUCKETS
+from data import datasets as ds_mod
+
+
+# --------------------------------------------------------------------------------------
+# Disk cache for shifted text (keyed by content hash so re-runs are free)
+# --------------------------------------------------------------------------------------
+def _cache_path(tag: str, key: str) -> Path:
+    h = hashlib.md5(key.encode()).hexdigest()[:12]
+    p = DATA_DIR / tag
+    p.mkdir(parents=True, exist_ok=True)
+    return p / f"{h}.json"
+
+
+def _load_cached(tag, key):
+    p = _cache_path(tag, key)
+    return json.loads(p.read_text()) if p.exists() else None
+
+
+def _save_cached(tag, key, obj):
+    _cache_path(tag, key).write_text(json.dumps(obj))
+
+
+# --------------------------------------------------------------------------------------
+# paraphrase via back-translation
+# --------------------------------------------------------------------------------------
+_BT_CACHE: dict = {}
+
+
+def _get_marian(src: str, tgt: str):
+    """Lazy-load + memoise a Marian opus-mt translation model."""
+    name = f"Helsinki-NLP/opus-mt-{src}-{tgt}"
+    if name not in _BT_CACHE:
+        from transformers import MarianMTModel, MarianTokenizer
+        import torch
+        tok = MarianTokenizer.from_pretrained(name)
+        model = MarianMTModel.from_pretrained(name)
+        if torch.cuda.is_available():
+            model = model.to("cuda").half()
+        _BT_CACHE[name] = (tok, model)
+    return _BT_CACHE[name]
+
+
+def _translate(texts: list[str], src: str, tgt: str, batch_size: int = 32) -> list[str]:
+    import torch
+    tok, model = _get_marian(src, tgt)
+    device = next(model.parameters()).device
+    out: list[str] = []
+    for i in range(0, len(texts), batch_size):
+        batch = texts[i:i + batch_size]
+        enc = tok(batch, return_tensors="pt", padding=True, truncation=True,
+                  max_length=256).to(device)
+        with torch.no_grad():
+            gen = model.generate(**enc, max_new_tokens=256)
+        out.extend(tok.batch_decode(gen, skip_special_tokens=True))
+    return out
+
+
+def paraphrase(record: dict, pivot: str | None = None) -> dict:
+    pivot = pivot or BT_PIVOTS[0]
+    key = f"{pivot}:" + "␟".join(record["texts"])
+    cached = _load_cached("paraphrase", key)
+    if cached is None:
+        mid = _translate(record["texts"], "en", pivot)
+        back = _translate(mid, pivot, "en")
+        cached = back
+        _save_cached("paraphrase", key, cached)
+    return {"texts": cached, "labels": record["labels"], "ids": record["ids"]}
+
+
+# --------------------------------------------------------------------------------------
+# domain shift
+# --------------------------------------------------------------------------------------
+def domain(dataset_key: str, n: int, seed: int = 0) -> dict | None:
+    partner = DATASETS[dataset_key].domain_partner
+    if partner is None:
+        return None
+    if DATASETS[partner].num_labels != DATASETS[dataset_key].num_labels:
+        # label spaces differ (e.g. ag_news 4 vs dbpedia 14) -> domain shift undefined.
+        # TODO(4090): map to shared coarse classes, or restrict `domain` to sentiment.
+        return None
+    return ds_mod.load(partner, n=n, seed=seed)
+
+
+# --------------------------------------------------------------------------------------
+# length shift
+# --------------------------------------------------------------------------------------
+def _token_len(t: str) -> int:
+    return len(t.split())
+
+
+def length_bucket(record: dict, bucket: str) -> dict:
+    lo, hi = LENGTH_BUCKETS[bucket]
+    keep = [i for i, t in enumerate(record["texts"]) if lo <= _token_len(t) < hi]
+    return {
+        "texts": [record["texts"][i] for i in keep],
+        "labels": record["labels"][np.asarray(keep, dtype=int)],
+        "ids": record["ids"][np.asarray(keep, dtype=int)],
+    }
+
+
+# --------------------------------------------------------------------------------------
+# dispatcher
+# --------------------------------------------------------------------------------------
+def make_shift(distribution: str, dataset_key: str, iid_record: dict, n: int, seed: int = 0):
+    """Return the eval record for a given distribution, or None if not applicable."""
+    if distribution == "iid":
+        return iid_record
+    if distribution == "paraphrase":
+        return paraphrase(iid_record)
+    if distribution == "domain":
+        return domain(dataset_key, n=n, seed=seed)
+    if distribution == "length":
+        return length_bucket(iid_record, "long")
+    raise ValueError(f"unknown distribution {distribution}")
+
+
+def make_augmentations(dataset_key: str, train_record: dict, n_aug: int) -> list[dict]:
+    """n_aug label-preserving augmentations of the TRAIN set for augmentation_consistency.
+    Uses DISTINCT back-translation pivots (AUG_PIVOTS = de/fr/ru) for genuine diversity;
+    cycles if n_aug > len(AUG_PIVOTS)."""
+    pivots = (AUG_PIVOTS * ((n_aug // len(AUG_PIVOTS)) + 1))[:n_aug]
+    return [paraphrase(train_record, pivot=p) for p in pivots]

repro_bundle/extract_activations.py

@@ -0,0 +1,79 @@
+"""Residual-stream activation extraction (pipeline stage M1 — the only big GPU cost).
+
+For each (model, dataset, distribution) we run a single forward pass over the texts with
+`output_hidden_states=True`, masked-mean-pool (or last-token-pool) each layer, cast to
+float16, and write one cache shard. Everything downstream reads the cache.
+
+Memory discipline (24 GB 4090):
+  * model loaded in float16
+  * inference under torch.no_grad()
+  * per-batch hidden states pooled and moved to CPU immediately (we never keep [B,L,T,H])
+  * batch_size tuned per model in config (drop for pythia-1.4b)
+"""
+from __future__ import annotations
+
+import numpy as np
+
+import cache
+from config import EXTRACT, MODELS, ExtractConfig
+
+
+def _pool(hidden_stack, attention_mask, pooling: str):
+    """hidden_stack: [B, L+1, T, H] ; attention_mask: [B, T] -> [B, L+1, H]."""
+    import torch
+    if pooling == "mean":
+        mask = attention_mask[:, None, :, None]                 # [B,1,T,1]
+        summed = (hidden_stack * mask).sum(dim=2)               # [B,L+1,H]
+        counts = mask.sum(dim=2).clamp(min=1)                   # [B,1,H]->broadcast
+        return summed / counts
+    if pooling == "last":
+        # index of last non-pad token per example
+        lengths = attention_mask.sum(dim=1).long() - 1          # [B]
+        idx = lengths[:, None, None, None].expand(-1, hidden_stack.size(1), 1, hidden_stack.size(3))
+        return torch.gather(hidden_stack, 2, idx).squeeze(2)    # [B,L+1,H]
+    raise ValueError(f"unknown pooling {pooling}")
+
+
+def extract_texts(model_key: str, texts: list[str], cfg: ExtractConfig) -> np.ndarray:
+    """Return pooled activations [N, L+1, H] (float16) for the given texts."""
+    import torch
+    from transformers import AutoModel, AutoTokenizer
+
+    spec = MODELS[model_key]
+    tok = AutoTokenizer.from_pretrained(spec.hf_name)
+    if tok.pad_token is None:
+        tok.pad_token = tok.eos_token
+    dtype = getattr(torch, cfg.dtype)
+    device = cfg.device if torch.cuda.is_available() else "cpu"
+    model = AutoModel.from_pretrained(
+        spec.hf_name, torch_dtype=dtype, output_hidden_states=True
+    ).to(device).eval()
+
+    chunks = []
+    with torch.no_grad():
+        for i in range(0, len(texts), cfg.batch_size):
+            batch = texts[i:i + cfg.batch_size]
+            enc = tok(batch, return_tensors="pt", padding=True, truncation=True,
+                      max_length=cfg.max_length).to(device)
+            out = model(**enc)
+            hs = torch.stack(out.hidden_states, dim=1)          # [B, L+1, T, H]
+            pooled = _pool(hs, enc.attention_mask, cfg.pooling) # [B, L+1, H]
+            chunks.append(pooled.to(torch.float16).cpu().numpy())
+            del out, hs, pooled, enc
+
+    del model
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+    return np.concatenate(chunks, axis=0)
+
+
+def extract_shard(model_key: str, dataset_key: str, distribution: str, record: dict,
+                  cfg: ExtractConfig = EXTRACT, overwrite: bool = False):
+    """Extract + cache one shard. `record` = {texts, labels, ids}."""
+    if len(record["texts"]) == 0:
+        return None   # empty record (e.g. paraphrase fully filtered out) -> skip, don't crash
+    if cache.exists(model_key, dataset_key, distribution) and not overwrite:
+        return cache.shard_dir(model_key, dataset_key, distribution)
+    acts = extract_texts(model_key, record["texts"], cfg)
+    return cache.save_shard(model_key, dataset_key, distribution,
+                            acts, record["labels"], record["ids"], cfg.pooling)

repro_bundle/figures.py

@@ -0,0 +1,117 @@
+"""Generate the paper's headline figures from cached results (no GPU).
+    python figures.py        # writes figures/*.pdf + *.png from results_A/
+"""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+
+import numpy as np
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+
+import metrics
+from config import MODELS
+from run_pipeline import (_para_rotation, _iid_split_rotation, _mean_ood_rotation, _mean_ood_drop)
+
+OUT = Path("figures"); OUT.mkdir(exist_ok=True)
+PREDS = ["raptor_stability", "pac", "augmentation_robustness", "whitened_cosine_id",
+         "fragility", "xie_feature_dispersion", "sip_eigengap"]
+LBL = {"raptor_stability": "RAPTOR\n(dispersion)", "pac": "PAC (ours)",
+       "augmentation_robustness": "aug-robust", "whitened_cosine_id": "whitened-cos",
+       "fragility": "Fragility", "xie_feature_dispersion": "Xie-disp", "sip_eigengap": "SIP"}
+
+
+def load_rows(results_dir, predfile="predictors.jsonl"):
+    R = Path(results_dir)
+    evals = [json.loads(l) for l in (R / "eval.jsonl").read_text().splitlines() if l.strip()]
+    preds = {(p["model"], p["dataset"], p["seed"]): p["predictors"]
+             for p in (json.loads(l) for l in (R / predfile).read_text().splitlines() if l.strip())}
+    rows = []
+    for e in evals:
+        if e.get("probe") != "logreg":
+            continue
+        k = (e["model"], e["dataset"], e["seed"])
+        if k not in preds:
+            continue
+        para, iids = _para_rotation(e), _iid_split_rotation(e)
+        excess = (para - iids) if (para == para and iids == iids) else np.nan
+        rows.append({"model": e["model"], "naive": _mean_ood_rotation(e), "placebo": iids,
+                     "excess": excess, "drop": _mean_ood_drop(e), **preds[k]})
+    return rows
+
+
+def rho(rows, pred, target):
+    s = np.array([r.get(pred, np.nan) for r in rows], float)
+    y = np.array([r[target] for r in rows], float)
+    ok = ~(np.isnan(s) | np.isnan(y))
+    return metrics.spearman(s[ok], -y[ok])[0] if ok.sum() >= 8 else np.nan
+
+
+def fig_circularity(rows):
+    targets = [("naive", "naive rotation"), ("placebo", "IID-resample\n(placebo)"),
+               ("excess", "EXCESS\n(shift-specific)")]
+    x = np.arange(len(PREDS)); w = 0.26
+    fig, ax = plt.subplots(figsize=(9, 4.2))
+    colors = ["#4C72B0", "#999999", "#C44E52"]
+    for i, (t, name) in enumerate(targets):
+        vals = [rho(rows, p, t) for p in PREDS]
+        ax.bar(x + (i - 1) * w, vals, w, label=name, color=colors[i])
+    ax.axhline(0, color="k", lw=0.8)
+    ax.set_xticks(x); ax.set_xticklabels([LBL[p] for p in PREDS], fontsize=8)
+    ax.set_ylabel("Spearman ρ (predictor, −target)")
+    ax.set_title("Predicting probe direction rotation: naive ρ is circular (≈ placebo);\n"
+                 "on EXCESS dispersion turns anti-predictive while augmentation is least-bad")
+    ax.legend(fontsize=8, loc="lower left"); ax.set_ylim(-0.45, 1.0)
+    fig.tight_layout(); fig.savefig(OUT / "fig1_circularity.pdf"); fig.savefig(OUT / "fig1_circularity.png", dpi=150)
+    plt.close(fig)
+
+
+def fig_sizeladder(rows):
+    models = sorted({r["model"] for r in rows}, key=lambda m: MODELS[m].params_m if m in MODELS else 0)
+    xs = [MODELS[m].params_m for m in models if m in MODELS]
+    fig, axes = plt.subplots(1, 2, figsize=(10, 4))
+    for ax, tgt, title in [(axes[0], "naive", "naive rotation"), (axes[1], "excess", "EXCESS rotation")]:
+        for p, c in [("raptor_stability", "#4C72B0"), ("augmentation_robustness", "#C44E52"), ("pac", "#55A868")]:
+            ys = [rho([r for r in rows if r["model"] == m], p, tgt) for m in models if m in MODELS]
+            ax.plot(xs, ys, "o-", label=LBL[p].replace("\n", " "), color=c)
+        ax.axhline(0, color="k", lw=0.6); ax.set_xscale("log")
+        ax.set_xlabel("params (M, log)"); ax.set_ylabel("Spearman ρ"); ax.set_title(title)
+        ax.legend(fontsize=8)
+    fig.suptitle("Size-ladder: predictability structure is scale-invariant (no inverse-scaling)")
+    fig.tight_layout(); fig.savefig(OUT / "fig2_sizeladder.pdf"); fig.savefig(OUT / "fig2_sizeladder.png", dpi=150)
+    plt.close(fig)
+
+
+def fig_mechanism(rows):
+    """Why augmentation > dispersion on EXCESS: dispersion (IID bootstrap) is unrelated/anti to
+    the shift-specific component; augmentation (a label-preserving perturbation) tracks it."""
+    fig, axes = plt.subplots(1, 2, figsize=(9, 4), sharey=True)
+    for ax, p, c in [(axes[0], "raptor_stability", "#4C72B0"),
+                     (axes[1], "augmentation_robustness", "#C44E52")]:
+        s = np.array([r.get(p, np.nan) for r in rows], float)
+        y = np.array([r["excess"] for r in rows], float)
+        ok = ~(np.isnan(s) | np.isnan(y))
+        s, y = s[ok], y[ok]
+        ax.scatter(s, y, s=8, alpha=0.3, color=c)
+        if len(s) > 2:
+            b, a = np.polyfit(s, y, 1)
+            xs = np.linspace(s.min(), s.max(), 20)
+            ax.plot(xs, a + b * xs, color="k", lw=1.5)
+        r = metrics.spearman(s, -y)[0]
+        ax.set_title(f"{LBL[p].replace(chr(10),' ')}\nρ(score,−excess)={r:+.2f}")
+        ax.set_xlabel(f"{p} (a-priori stability)"); ax.axhline(0, color="grey", lw=0.6)
+    axes[0].set_ylabel("EXCESS rotation (shift-specific)")
+    fig.suptitle("Mechanism: dispersion is blind to shift-specific fragility; augmentation is not")
+    fig.tight_layout(); fig.savefig(OUT / "fig3_mechanism.pdf"); fig.savefig(OUT / "fig3_mechanism.png", dpi=150)
+    plt.close(fig)
+
+
+if __name__ == "__main__":
+    rows = load_rows("results_A/results")
+    print(f"loaded {len(rows)} configs")
+    fig_circularity(rows)
+    fig_sizeladder(rows)
+    fig_mechanism(rows)
+    print("wrote figures/fig1_circularity, fig2_sizeladder, fig3_mechanism (.pdf/.png)")

repro_bundle/full_run.sh

@@ -0,0 +1,28 @@
+#!/bin/bash
+# First full ProbeShift grid: 6 models x 5 datasets (3 concepts) -> 30 configs.
+# Produces the first REAL G2 ranking table (predictors vs OOD acc-drop AND rotation).
+# NOTE: n_aug=1 because back-translation currently uses a single pivot (de) so extra augs
+# would be identical. Multi-pivot diverse augmentation is a paper TODO.
+cd /root/rivermind-data/probe_stability
+export HF_ENDPOINT=https://hf-mirror.com HF_HOME=/root/rivermind-data/hf_cache HF_HUB_DISABLE_XET=1 HF_HUB_DOWNLOAD_TIMEOUT=60
+# CRITICAL: cap BLAS threads. Default (15) oversubscribes on many small logreg fits and
+# makes 14-class eval ~100x slower (37min -> 20s with 4 threads). Measured 2026-06-22.
+export OMP_NUM_THREADS=4 OPENBLAS_NUM_THREADS=4 MKL_NUM_THREADS=4
+export PROBE_N_TRAIN=1500 PROBE_N_EVAL=800 PROBE_N_AUG=1 PROBE_BATCH=32
+MODELS="pythia-70m pythia-160m pythia-410m pythia-1.4b gpt2 qwen2.5-0.5b"
+DSETS="sst2 imdb ag_news dbpedia counterfact emotion tweet_hate tweet_irony tweet_offensive subj spam cola stance amazon_cf"
+NOISE='Loading weights|LOAD REPORT|UNEXPECTED|Notes:|can be ignored|embed_out|^Key|^---|xethub|Trying to resume|Warning: You are|Generating|Fetching|max_new_tokens|torch_dtype|Map:'
+for m in $MODELS; do
+  for d in $DSETS; do
+    echo "=== EXTRACT $m $d $(date +%H:%M:%S) ==="
+    python3 -u run_pipeline.py extract --model "$m" --dataset "$d" 2>&1 | grep -vE "$NOISE"
+    echo "=== EVAL    $m $d ==="
+    python3 -u run_pipeline.py eval    --model "$m" --dataset "$d" 2>&1 | grep -E "\[eval\]|Error|Traceback|nan"
+    echo "=== PREDICT $m $d ==="
+    python3 -u run_pipeline.py predict --model "$m" --dataset "$d" 2>&1 | grep -E "\[predict\]|Error|Traceback"
+    echo "--- disk: $(df -h /root/rivermind-data | tail -1 | awk '{print $3" used, "$4" free"}') ---"
+  done
+done
+echo "=== STATS ==="
+python3 -u run_pipeline.py stats 2>&1
+echo "=== FULL_DONE $(date +%H:%M:%S) ==="

repro_bundle/full_run_A.sh

@@ -0,0 +1,23 @@
+#!/bin/bash
+# Option A multi-seed: each seed independently re-draws train/eval examples, re-extracts
+# activations, and re-generates paraphrase (own seed-keyed cache on JuiceFS). Gold-standard
+# independent replication. ~5x cost of a single run.
+cd /root/rivermind-data/probe_stability
+export HF_ENDPOINT=https://hf-mirror.com HF_HOME=/root/rivermind-data/hf_cache HF_HUB_DISABLE_XET=1 HF_HUB_DOWNLOAD_TIMEOUT=60
+export OMP_NUM_THREADS=4 OPENBLAS_NUM_THREADS=4 MKL_NUM_THREADS=4
+export PROBE_N_TRAIN=1500 PROBE_N_EVAL=800 PROBE_N_AUG=1 PROBE_BATCH=32
+NOISE='Loading weights|LOAD REPORT|UNEXPECTED|Notes:|can be ignored|embed_out|^Key|^---|xethub|Trying to resume|Warning: You are|Generating|Fetching|max_new_tokens|torch_dtype|Map:'
+
+for seed in 0 1 2 3 4; do
+  export PROBE_CACHE=/root/rivermind-fs/cache_seed$seed
+  export PROBE_DATA=/root/rivermind-fs/data_seed$seed
+  echo "########## SEED $seed EXTRACT $(date +%H:%M:%S) ##########"
+  python3 -u run_pipeline.py extract --all --seed $seed 2>&1 | grep -vE "$NOISE" | grep -E "\[extract\]|=== |Error|Traceback|ValueError"
+  echo "--- fs: $(df -h /root/rivermind-fs | tail -1 | awk '{print $3" used, "$4" free"}') ---"
+  echo "########## SEED $seed GRID $(date +%H:%M:%S) ##########"
+  python3 -u run_pipeline.py grid --all --seeds $seed 2>&1 | grep -E "\[grid\]|Error|Traceback|ValueError"
+done
+
+echo "########## STATS ##########"
+python3 -u run_pipeline.py stats 2>&1
+echo "=== A_DONE $(date +%H:%M:%S) ==="

repro_bundle/full_run_ms.sh

@@ -0,0 +1,35 @@
+#!/bin/bash
+# Multi-seed ProbeShift run (Option C): extract a LARGER pool once, then per-seed subsample
+# (train + eval) for 5 seeds. Phase 1 = pool extraction + back-translation (once); Phase 2 =
+# multi-seed grid (eval+predict, layer selected once per config, looped over seeds in-process).
+cd /root/rivermind-data/probe_stability
+export HF_ENDPOINT=https://hf-mirror.com HF_HOME=/root/rivermind-data/hf_cache HF_HUB_DISABLE_XET=1 HF_HUB_DOWNLOAD_TIMEOUT=60
+export OMP_NUM_THREADS=4 OPENBLAS_NUM_THREADS=4 MKL_NUM_THREADS=4
+# big pool cache on the 200G JuiceFS volume (data disk is only 49G); I/O verified fast.
+export PROBE_CACHE=/root/rivermind-fs/probe_cache PROBE_DATA=/root/rivermind-fs/data_cache
+export PROBE_N_TRAIN=1500 PROBE_N_EVAL=800 PROBE_N_AUG=1 PROBE_BATCH=32
+export PROBE_N_TRAIN_POOL=4000 PROBE_N_EVAL_POOL=2000
+MODELS="pythia-70m pythia-160m pythia-410m pythia-1.4b gpt2 qwen2.5-0.5b"
+DSETS="sst2 imdb ag_news dbpedia counterfact emotion tweet_hate tweet_irony tweet_offensive subj spam cola stance amazon_cf"
+NOISE='Loading weights|LOAD REPORT|UNEXPECTED|Notes:|can be ignored|embed_out|^Key|^---|xethub|Trying to resume|Warning: You are|Generating|Fetching|max_new_tokens|torch_dtype|Map:'
+
+echo "########## PHASE 1: POOL EXTRACTION (once) ##########"
+for m in $MODELS; do
+  for d in $DSETS; do
+    echo "=== EXTRACT $m $d $(date +%H:%M:%S) ==="
+    python3 -u run_pipeline.py extract --model "$m" --dataset "$d" 2>&1 | grep -vE "$NOISE"
+    echo "--- disk: $(df -h /root/rivermind-data | tail -1 | awk '{print $3" used, "$4" free"}') ---"
+  done
+done
+
+echo "########## PHASE 2: MULTI-SEED GRID (5 seeds) ##########"
+for m in $MODELS; do
+  for d in $DSETS; do
+    echo "=== GRID $m $d $(date +%H:%M:%S) ==="
+    python3 -u run_pipeline.py grid --model "$m" --dataset "$d" --seeds 0,1,2,3,4 2>&1 | grep -E "\[grid\]|Error|Traceback|ValueError"
+  done
+done
+
+echo "########## STATS ##########"
+python3 -u run_pipeline.py stats 2>&1
+echo "=== MS_DONE $(date +%H:%M:%S) ==="

repro_bundle/metrics.py

@@ -0,0 +1,139 @@
+"""Metrics: accuracy, concept-direction rotation, rank correlation, and the statistics
+needed for a defensible diagnostic paper (bootstrap CI, Cliff's delta, Holm / BH).
+
+Direction handling
+------------------
+A probe exposes a `direction` of shape [C, H] (C = #classes for one-vs-rest weight vectors,
+or 1 for a binary difference-of-means vector). "Rotation" between two directions measured on
+different distributions is reported two ways:
+  * `mean_class_cosine` — mean of per-class cosine similarity of matched rows (interpretable).
+  * `subspace_principal_angle` — largest principal angle between the row-spaces (robust to
+    class permutation / multiclass), in radians.
+"""
+from __future__ import annotations
+
+import numpy as np
+from scipy import stats
+
+
+# --------------------------------------------------------------------------------------
+# Accuracy
+# --------------------------------------------------------------------------------------
+def accuracy(y_true: np.ndarray, y_pred: np.ndarray) -> float:
+    return float(np.mean(np.asarray(y_true) == np.asarray(y_pred)))
+
+
+# --------------------------------------------------------------------------------------
+# Direction rotation
+# --------------------------------------------------------------------------------------
+def _unit_rows(m: np.ndarray) -> np.ndarray:
+    m = np.atleast_2d(np.asarray(m, dtype=np.float64))
+    norms = np.linalg.norm(m, axis=1, keepdims=True)
+    norms[norms == 0] = 1.0
+    return m / norms
+
+
+def mean_class_cosine(dir_a: np.ndarray, dir_b: np.ndarray) -> float:
+    """Mean cosine over matched class rows. 1.0 = identical, 0 = orthogonal."""
+    a, b = _unit_rows(dir_a), _unit_rows(dir_b)
+    if a.shape != b.shape:
+        raise ValueError(f"direction shape mismatch {a.shape} vs {b.shape}")
+    return float(np.mean(np.sum(a * b, axis=1)))
+
+
+def subspace_principal_angle(dir_a: np.ndarray, dir_b: np.ndarray) -> float:
+    """Largest principal angle (radians) between row-spaces of two direction matrices."""
+    a = _unit_rows(dir_a)
+    b = _unit_rows(dir_b)
+    qa, _ = np.linalg.qr(a.T)   # columns span row-space of a
+    qb, _ = np.linalg.qr(b.T)
+    s = np.linalg.svd(qa.T @ qb, compute_uv=False)
+    s = np.clip(s, -1.0, 1.0)
+    return float(np.arccos(s.min()))
+
+
+def rotation(dir_a: np.ndarray, dir_b: np.ndarray) -> dict:
+    out = {"subspace_principal_angle": subspace_principal_angle(dir_a, dir_b)}
+    try:
+        out["mean_class_cosine"] = mean_class_cosine(dir_a, dir_b)
+    except ValueError:
+        out["mean_class_cosine"] = float("nan")
+    return out
+
+
+# --------------------------------------------------------------------------------------
+# Rank correlation (Stability Score vs OOD drop) — claim C2
+# --------------------------------------------------------------------------------------
+def spearman(x: np.ndarray, y: np.ndarray) -> tuple[float, float]:
+    r = stats.spearmanr(np.asarray(x), np.asarray(y))
+    return float(r.statistic), float(r.pvalue)
+
+
+def kendall(x: np.ndarray, y: np.ndarray) -> tuple[float, float]:
+    r = stats.kendalltau(np.asarray(x), np.asarray(y))
+    return float(r.statistic), float(r.pvalue)
+
+
+# --------------------------------------------------------------------------------------
+# Effect size
+# --------------------------------------------------------------------------------------
+def cliffs_delta(a: np.ndarray, b: np.ndarray) -> float:
+    """Cliff's delta in [-1, 1]; |delta| ~ .11/.28/.43 = small/medium/large."""
+    a, b = np.asarray(a), np.asarray(b)
+    gt = sum((x > b).sum() for x in a)
+    lt = sum((x < b).sum() for x in a)
+    return float((gt - lt) / (len(a) * len(b)))
+
+
+# --------------------------------------------------------------------------------------
+# Bootstrap CI
+# --------------------------------------------------------------------------------------
+def bootstrap_ci(
+    values: np.ndarray,
+    statfn=np.mean,
+    n_boot: int = 1000,
+    alpha: float = 0.05,
+    seed: int = 0,
+) -> tuple[float, float, float]:
+    """Return (point_estimate, lo, hi) percentile bootstrap CI."""
+    rng = np.random.default_rng(seed)
+    values = np.asarray(values, dtype=np.float64)
+    n = len(values)
+    boots = np.array([statfn(values[rng.integers(0, n, n)]) for _ in range(n_boot)])
+    lo, hi = np.percentile(boots, [100 * alpha / 2, 100 * (1 - alpha / 2)])
+    return float(statfn(values)), float(lo), float(hi)
+
+
+# --------------------------------------------------------------------------------------
+# Multiple-comparison correction
+# --------------------------------------------------------------------------------------
+def holm_bonferroni(pvals: list[float], alpha: float = 0.05) -> list[bool]:
+    """Return per-test reject decisions under Holm step-down."""
+    p = np.asarray(pvals, dtype=np.float64)
+    order = np.argsort(p)
+    m = len(p)
+    reject = np.zeros(m, dtype=bool)
+    for rank, idx in enumerate(order):
+        if p[idx] <= alpha / (m - rank):
+            reject[idx] = True
+        else:
+            break
+    return reject.tolist()
+
+
+def benjamini_hochberg(pvals: list[float], alpha: float = 0.05) -> list[bool]:
+    """Return per-test reject decisions under BH FDR control."""
+    p = np.asarray(pvals, dtype=np.float64)
+    m = len(p)
+    order = np.argsort(p)
+    thresh = (np.arange(1, m + 1) / m) * alpha
+    passed = p[order] <= thresh
+    reject = np.zeros(m, dtype=bool)
+    if passed.any():
+        kmax = np.max(np.where(passed))
+        reject[order[: kmax + 1]] = True
+    return reject.tolist()
+
+
+def correct(pvals: list[float], method: str = "holm", alpha: float = 0.05) -> list[bool]:
+    return (holm_bonferroni if method == "holm" else benjamini_hochberg)(pvals, alpha)

repro_bundle/pilot_mini.sh

@@ -0,0 +1,17 @@
+#!/bin/bash
+# Mini pilot: shake out the full eval->predict->stats chain on real data
+# (binary sst2 + multiclass ag_news) at small N, across 2 models. Cheap + fast.
+cd /root/rivermind-data/probe_stability
+export HF_ENDPOINT=https://hf-mirror.com HF_HOME=/root/rivermind-data/hf_cache HF_HUB_DISABLE_XET=1
+export PROBE_N_TRAIN=600 PROBE_N_EVAL=400 PROBE_N_AUG=2 PROBE_BATCH=64
+MODELS="pythia-70m pythia-410m"
+DSETS="sst2 ag_news"
+for m in $MODELS; do
+  for d in $DSETS; do
+    echo "=== EXTRACT $m $d ==="; python3 -u run_pipeline.py extract  --model "$m" --dataset "$d" 2>&1 | grep -vE "Loading weights|LOAD REPORT|UNEXPECTED|Notes:|can be ignored|embed_out|^Key|^---|xethub|Trying to resume|Warning: You are|Generating|Fetching"
+    echo "=== EVAL    $m $d ==="; python3 -u run_pipeline.py eval     --model "$m" --dataset "$d" 2>&1 | tail -8
+    echo "=== PREDICT $m $d ==="; python3 -u run_pipeline.py predict  --model "$m" --dataset "$d" 2>&1 | tail -5
+  done
+done
+echo "=== STATS ==="; python3 -u run_pipeline.py stats 2>&1
+echo "=== PILOT_DONE ==="

repro_bundle/predownload.py

@@ -0,0 +1,52 @@
+"""Robustly pre-download all HF models a run needs, with resume + retries, via the
+configured HF_ENDPOINT mirror. Run this before extraction so the pipeline never hangs
+on a flaky mid-run download.
+
+    HF_ENDPOINT=https://hf-mirror.com HF_HOME=/root/rivermind-data/hf_cache \
+        python3 predownload.py smoke      # just the smoke-test models
+        python3 predownload.py all        # full size ladder + aux models
+"""
+import sys
+import time
+
+from huggingface_hub import snapshot_download
+
+AUX = [
+    "Helsinki-NLP/opus-mt-en-de",
+    "Helsinki-NLP/opus-mt-de-en",
+    "MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli",
+]
+SMOKE = ["EleutherAI/pythia-410m"] + AUX
+ALL = [
+    "EleutherAI/pythia-70m", "EleutherAI/pythia-160m", "EleutherAI/pythia-410m",
+    "EleutherAI/pythia-1.4b", "EleutherAI/pythia-2.8b",
+    "gpt2", "gpt2-medium", "Qwen/Qwen2.5-0.5B",
+] + AUX
+
+
+def fetch(repo, retries=6):
+    for a in range(retries):
+        try:
+            print(f"DL {repo} (attempt {a})", flush=True)
+            # prefer safetensors; skip duplicate .bin weights where present
+            snapshot_download(repo, ignore_patterns=["*.bin", "*.h5", "*.msgpack", "*.ot"])
+            print(f"OK {repo}", flush=True)
+            return True
+        except Exception as e:
+            print(f"ERR {repo}: {str(e)[:120]}", flush=True)
+            time.sleep(4)
+    # last resort: allow .bin too (some repos have no safetensors)
+    try:
+        snapshot_download(repo)
+        print(f"OK(bin) {repo}", flush=True)
+        return True
+    except Exception as e:
+        print(f"GIVEUP {repo}: {str(e)[:120]}", flush=True)
+        return False
+
+
+if __name__ == "__main__":
+    mode = sys.argv[1] if len(sys.argv) > 1 else "smoke"
+    repos = {"smoke": SMOKE, "all": ALL}[mode]
+    ok = sum(fetch(r) for r in repos)
+    print(f"ALL_DONE {ok}/{len(repos)} ok", flush=True)

repro_bundle/preregistration.md

@@ -0,0 +1,43 @@
+# ProbeShift —— 分析协议与预注册判定阈值
+
+> 决策规则在看到 5-seed A 跑的最终 held-out 结果前冻结。诚实说明:single-seed 探索阶段已观察到
+> 部分趋势,故此为"分析协议 + 预注册判定阈值",而非严格盲注册。所有阈值与代码随论文发布。
+
+## 度量与目标(均在 ≥12 概念 × 多模型 × 5 独立 seed 上)
+- **acc-drop** = IID_acc − OOD_acc(probe 在 OOD 集上)
+- **rotation** = 1 − cos(ID 方向, OOD 重拟合方向),mean over shift types
+- **iid_split_rotation**(安慰剂)= 1 − cos(train 两半各自方向)——纯采样噪声,无语义 shift
+- **excess_rotation** = paraphrase_rotation − iid_split_rotation —— shift 特异脆性(去采样噪声)
+- 预测器(label-free, IID-only):SIP、RAPTOR-stability、Fragility、augmentation-robustness、
+  Xie-dispersion、PAC、whitened-cosine(ID)
+
+## 假设与判定阈值
+
+- **H1(基准有效性)**:ID accuracy 高估方向稳定性。
+  判定:跨配置 ρ(ID-acc, OOD-retention) 的 bootstrap 95% CI 上界 < 0.5。
+
+- **H2(naive rotation 可预测)**:至少一个 label-free 信号能 held-out 预测 naive rotation。
+  判定:最佳预测器 held-out ρ 的 95% CI 下界 > 0.4。
+
+- **H3-CIRC(循环性,核心)**:naive rotation 的强预测大部分来自采样噪声。
+  判定:若 best 预测器对 **iid_split_rotation(安慰剂)** 的 ρ ≥ 其对 naive rotation 的 ρ 的 80%,
+  则认定"naive 预测主要是循环的",并以 **excess_rotation** 为真正目标。
+
+- **H4-EXCESS(headline)**:扣除采样噪声后,shift 特异脆性是否可预测、谁能预测。
+  判定:报告所有预测器对 excess 的 ρ + 95% CI。**预注册主张**:基于增广的信号
+  (augmentation-robustness / PAC)对 excess 的 ρ > 基于离散度的(RAPTOR-stability / dispersion)。
+  **配对检验**:Δρ[aug − raptor] on excess 的 bootstrap 95% CI 下界 > 0 ⇒ SIGNIFICANT。
+  若不显著 ⇒ 诚实降级为"shift 特异脆性目前不可被任何现有 label-free 信号可靠预测"(仍是有价值的 open problem)。
+
+- **H5(acc-drop 不可预测)**:判定:所有预测器对 acc-drop 的 ρ 的 95% CI 均跨 0。
+
+- **H6(外部效度)**:H3-CIRC / H4 的结论跨 estimator(logreg/mass-mean)与跨 shift-type 稳健;
+  跨 size ladder 不下"越大越稳"普适结论,显式检验 inverse-scaling。
+
+- **H7(label-fidelity sanity)**:paraphrase shift 的 NLI 通过率 ≥ 90%(逐数据集报告);
+  通过率过低的数据集(如 noisy tweets)在 rotation 分析中标注并谨慎解读。
+
+## 统计
+- 相关系数报 bootstrap(1000)95% CI;多重比较 Holm-Bonferroni;
+- LOCO = leave-one-concept-out 均值(跨概念泛化);配对 Δρ bootstrap 检验(aug vs raptor on excess)。
+- effect size 报 Cliff's δ;每配置 5 seed(独立重抽,Option A)。

repro_bundle/probes.py

@@ -0,0 +1,144 @@
+"""Probes.
+
+All probes share a tiny interface so the pipeline is agnostic to type:
+
+    probe = make_probe("logreg", num_labels=2, seed=0).fit(X_train, y_train)
+    y_hat = probe.predict(X_eval)
+    D = probe.direction          # [C, H] concept-direction matrix, or None (MLP)
+
+`direction` is what claim C1/C2 measure rotation on. For LogReg it is the weight matrix;
+for mass-mean it is (class_centroid - global_mean). MLP has no single linear direction, so
+it returns None and the eval step skips rotation for it (still reports accuracy).
+"""
+from __future__ import annotations
+
+import numpy as np
+from sklearn.linear_model import LogisticRegression
+
+
+class BaseProbe:
+    direction: np.ndarray | None = None
+
+    def fit(self, X: np.ndarray, y: np.ndarray) -> "BaseProbe":
+        raise NotImplementedError
+
+    def predict(self, X: np.ndarray) -> np.ndarray:
+        raise NotImplementedError
+
+
+class LogRegProbe(BaseProbe):
+    """L2-regularised multinomial logistic regression (the default diagnostic probe)."""
+
+    def __init__(self, num_labels: int, l2: float = 1.0, seed: int = 0, max_iter: int = 500, **_):
+        self.num_labels = num_labels
+        self.C = 1.0 / max(l2, 1e-8)
+        self.seed = seed
+        self.max_iter = max_iter
+        self._clf: LogisticRegression | None = None
+
+    def fit(self, X, y):
+        self._clf = LogisticRegression(
+            C=self.C, max_iter=self.max_iter, random_state=self.seed,
+        ).fit(X, y)
+        coef = self._clf.coef_                       # [1,H] binary, [C,H] multiclass
+        self.direction = coef.copy()
+        return self
+
+    def predict(self, X):
+        return self._clf.predict(X)
+
+
+class MassMeanProbe(BaseProbe):
+    """Difference-of-means / mass-mean probe. Predicts by nearest class centroid.
+
+    Often *more* OOD-robust than LogReg (a hypothesis under claim C3).
+    """
+
+    def __init__(self, num_labels: int, **_):
+        self.num_labels = num_labels
+        self._centroids: np.ndarray | None = None
+        self._classes: np.ndarray | None = None
+
+    def fit(self, X, y):
+        y = np.asarray(y)
+        self._classes = np.unique(y)
+        centroids = np.stack([X[y == c].mean(0) for c in self._classes])  # [C,H]
+        self._centroids = centroids
+        self.direction = centroids - X.mean(0, keepdims=True)             # [C,H]
+        return self
+
+    def predict(self, X):
+        # nearest centroid (euclidean)
+        d = np.linalg.norm(X[:, None, :] - self._centroids[None, :, :], axis=2)  # [N,C]
+        return self._classes[np.argmin(d, axis=1)]
+
+
+class MLPProbe(BaseProbe):
+    """2-layer MLP probe (torch). No single linear direction -> direction stays None."""
+
+    def __init__(self, num_labels: int, hidden: int = 128, epochs: int = 50,
+                 lr: float = 1e-3, seed: int = 0, **_):
+        self.num_labels = num_labels
+        self.hidden = hidden
+        self.epochs = epochs
+        self.lr = lr
+        self.seed = seed
+        self._model = None
+        self._device = None
+
+    def fit(self, X, y):
+        import torch
+        import torch.nn as nn
+
+        torch.manual_seed(self.seed)
+        self._device = "cuda" if torch.cuda.is_available() else "cpu"
+        Xt = torch.tensor(np.asarray(X), dtype=torch.float32, device=self._device)
+        yt = torch.tensor(np.asarray(y), dtype=torch.long, device=self._device)
+        self._model = nn.Sequential(
+            nn.Linear(Xt.shape[1], self.hidden), nn.ReLU(),
+            nn.Linear(self.hidden, self.num_labels),
+        ).to(self._device)
+        opt = torch.optim.Adam(self._model.parameters(), lr=self.lr)
+        lossfn = nn.CrossEntropyLoss()
+        self._model.train()
+        for _ in range(self.epochs):
+            opt.zero_grad()
+            loss = lossfn(self._model(Xt), yt)
+            loss.backward()
+            opt.step()
+        return self
+
+    def predict(self, X):
+        import torch
+        self._model.eval()
+        with torch.no_grad():
+            Xt = torch.tensor(np.asarray(X), dtype=torch.float32, device=self._device)
+            return self._model(Xt).argmax(1).cpu().numpy()
+
+
+class ControlTaskProbe(LogRegProbe):
+    """Hewitt & Liang control-task probe: fit on RANDOM labels to measure selectivity.
+
+    selectivity = real_task_acc - control_task_acc. A high-selectivity probe reflects the
+    representation, not probe memorisation. We report this to pre-empt the standard
+    "your probe is just memorising" reviewer attack.
+    """
+
+    def fit(self, X, y):
+        rng = np.random.default_rng(self.seed)
+        y_rand = rng.permutation(np.asarray(y))     # destroy structure, keep label marginal
+        return super().fit(X, y_rand)
+
+
+_REGISTRY = {
+    "logreg": LogRegProbe,
+    "mass_mean": MassMeanProbe,
+    "mlp": MLPProbe,
+    "control": ControlTaskProbe,
+}
+
+
+def make_probe(kind: str, num_labels: int, **kwargs) -> BaseProbe:
+    if kind not in _REGISTRY:
+        raise KeyError(f"unknown probe '{kind}', choose from {list(_REGISTRY)}")
+    return _REGISTRY[kind](num_labels=num_labels, **kwargs)

repro_bundle/probeshift_honest.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:20b5463a050816c80c46efe1af77ebf846fd1c57b15d9fc5be63f7d22ba511d9
+size 599546

repro_bundle/probeshift_honest.tex

@@ -0,0 +1,1254 @@
+%==============================================================================
+% ProbeShift --- honest-calibration synthesis (single self-contained file)
+%
+% PMLR-STYLE NOTE
+% ---------------
+% This file is an article+booktabs compilable build. For the camera-ready PMLR
+% submission, replace the documentclass/preamble block below with the official
+% PMLR (JMLR/MLR Press) class, e.g.:
+%
+%     \documentclass[wcp,twoside]{jmlr}     % or the ACML PMLR variant
+%     \usepackage{jmlr2e}                    % PMLR macro package
+%
+% and remove the placeholder \title/\author/\date scaffolding that the standard
+% article class needs. All section content, tables (booktabs), figures, and the
+% \cite/\citep commands below are PMLR-compatible as written.
+%==============================================================================
+\documentclass[11pt]{article}
+
+% ---- Packages (article fallback; PMLR class supplies most of these) ---------
+\usepackage[utf8]{inputenc}
+\usepackage[T1]{fontenc}
+\usepackage{amsmath,amssymb}
+\usepackage{booktabs}
+\usepackage{graphicx}
+\usepackage{makecell}
+\usepackage{enumitem}
+\usepackage[round]{natbib}      % \citep/\citet, plainnat-compatible
+\usepackage{xcolor}
+\usepackage{hyperref}
+\usepackage[capitalise]{cleveref}
+\usepackage[margin=1in]{geometry}
+
+% ---- Project macros ---------------------------------------------------------
+\newcommand{\probeshift}{\textsc{ProbeShift}}
+\newcommand{\excess}{\textsc{excess}}
+\newcommand{\wbar}{\mathbf{w}_{\text{ID}}}
+\newcommand{\wshift}{\mathbf{w}_{\text{OOD}}}
+\newcommand{\accdrop}{\textsc{acc-drop}}
+\newcommand{\rotation}{\textsc{rotation}}
+
+\title{ProbeShift: Probe-Direction Stability under Shift is Largely
+Circular --- and What Survives a Proper Control}
+\author{Anonymous Authors}
+\date{}
+
+\begin{document}
+\maketitle
+
+%================================================================ Abstract
+\begin{abstract}
+Linear probes are increasingly used as label-free read-outs of the internal
+concepts of large language models, yet their in-distribution direction is
+unstable: under label-preserving semantic shift (paraphrase / domain / length)
+the probe direction rotates, so a probe trusted as a free out-of-distribution
+(OOD) evaluator can silently break. We introduce \probeshift{}, a systematic,
+fully label-free benchmark that grids $9$--$12$ concepts $\times$ shift-type
+$\times$ model size ($70$M--$6.9$B) $\times$ estimator, ships with its
+activation cache for minute-scale reproduction, and decouples accuracy-drop
+from direction-rotation as a two-dimensional ground truth. Our first finding is
+one of \emph{circularity}: the apparently near-perfect predictability of naive
+rotation from probe dispersion ($\rho\!\approx\!0.95$) largely reflects a
+sampling-noise floor---dispersion predicts a pure IID-resampling placebo just as
+well ($\rho\!=\!0.966 \geq 0.945$). This motivates a methodological point:
+estimating predictive power \emph{beyond} sampling noise requires partial
+correlation or placebo controls, not difference scores, which mechanically
+manufacture spurious negative correlations. Under a clean partial-Spearman
+analysis, dispersion ($+0.32$), augmentation ($+0.35$), and their composite PAC
+($+0.37$) all retain modest residual predictive power above the noise floor,
+with PAC best and the gaps small. Concept-level cluster bootstrap (effective
+$n\!\approx\!12$) corroborates this. All experiments run on a single
+RTX~4090 ($\leq\!200$ GPU$\cdot$h), \$0 API, and zero new annotation; we release
+the cache, code, and splits.
+\end{abstract}
+
+%================================================================ Introduction
+\section{Introduction}
+\label{sec:intro}
+
+Linear probes are the default tool for reading concepts---truth, sentiment,
+toxicity, topic---out of the internal representations of language models,
+precisely because they are cheap, label-free at deployment, and seemingly
+interpretable~\citep{marks2024geometry,belinkov2022probing}. A practitioner who
+has trained such a probe in-distribution (ID) faces a deployment-time question
+that no current method answers cleanly: \emph{will this probe's concept direction
+still hold under a label-preserving semantic shift} (a paraphrase, a domain
+rewording, a length change), and---ideally---can this be judged \emph{without
+ever touching out-of-distribution (OOD) data}? A growing set of recent signals
+claims to bear on this question: representational stability of truth under
+rephrasing~\citep{dies2025pstat}, spectral identifiability criteria
+(SIP)~\citep{huang2025sip}, forward-pass fragility
+thresholds~\citep{reblitz2026fragility}, ridge-adaptive directional stability
+(RAPTOR)~\citep{gao2026raptor}, augmentation
+robustness~\citep{lysnaes2025probing}, whitened directional similarity for
+truthfulness~\citep{ying2026truthspectrum}, and feature-dispersion OOD-error
+predictors~\citep{xie2023feature}. Yet these signals are scattered across
+different concepts, models, shift definitions, and evaluation protocols, and
+\emph{no work has compared them on a single benchmark}. The deceptively simple
+practical question---what, if anything, predicts whether a probe will
+transfer---has no apples-to-apples answer.
+
+\paragraph{The circularity hook.}
+The starting point of this paper is an observation that, taken at face value,
+looks like a solved problem. The simplest label-free signal, the directional
+dispersion of a probe estimated by IID bootstrap resampling (the RAPTOR-style
+stability component~\citep{gao2026raptor}), predicts the OOD direction-rotation
+of that probe almost perfectly: Spearman $\rho = 0.945$ ($95\%$ CI
+$[0.93, 0.95]$). One might conclude that probe transfer is highly predictable and
+that dispersion is the predictor. The conclusion is largely an illusion. We
+construct a \emph{placebo} target by measuring the rotation a probe direction
+undergoes under pure IID resampling---no semantic shift at all---and find that
+dispersion predicts this sampling-noise floor \emph{just as well, in fact
+slightly better}: $\rho = 0.966$ ($[0.96, 0.97]$). The two quantities are two
+views of the same thing---how much a refit direction wanders under
+resampling---so the apparent skill mostly reflects a probe's sampling-noise
+floor rather than its shift-specific brittleness. This circularity is exactly the
+hazard latent in pairing ``rotation under rephrasing'' with ``dispersion across
+resamples,'' as in~\citet{dies2025pstat}, and it is the first thing a benchmark
+for this problem must control for.
+
+\paragraph{A methodological trap: difference scores manufacture artifacts.}
+The natural fix is to subtract the placebo floor from the observed rotation,
+forming an \emph{excess} target ($\textsc{excess} = $ paraphrase-rotation $-$
+placebo-rotation) meant to isolate the genuinely shift-specific component. This
+difference construction is tempting but treacherous. Because dispersion predicts
+the placebo term \emph{by design} ($\rho\!=\!0.966$), subtracting that term from
+the minuend mechanically drives the residual's correlation with dispersion
+\emph{negative}---an artifact of the difference operator, not evidence about
+shift-specific structure. On \textsc{excess}, dispersion lands at $\rho=-0.195$,
+and one could be misled into reporting that a widely used stability signal
+``anti-predicts'' shift, that shift-specific brittleness is predicted by no
+signal, or that augmentation ``reverses'' the ranking. None of these
+conclusions survive scrutiny: they are induced by evaluating predictors against a
+constructed difference score whose noise term is itself one of the predictors.
+
+\paragraph{The correct analysis: partial correlation, not difference scores.}
+To ask whether a signal carries information about shift-induced rotation
+\emph{beyond} the sampling floor, the principled estimator is a partial
+correlation that conditions on---rather than subtracts---the placebo. We report
+partial Spearman $\rho(\text{signal},\, \text{paraphrase-rotation} \mid
+\text{placebo})$, which removes the shared sampling-noise variance without
+fabricating a spurious negative slope. Under this clean analysis the picture is
+neither circular nor catastrophic: dispersion attains $+0.323$, augmentation
+robustness $+0.345$, and the composite PAC $+0.374$, while whitened-cosine and
+SIP sit at $\approx 0$. The difference-score artifact is gone, and a modest,
+honest signal emerges.
+
+\paragraph{Honest result: modest residual predictability, PAC slightly ahead.}
+Conditioning on the sampling floor, \emph{both} dispersion- and
+augmentation-based signals retain a modest residual ability to predict
+shift-specific rotation, in the $\rho\!\approx\!0.32\text{--}0.37$ range, with PAC
+the best and the gaps between the top signals small. Shift-specific brittleness is
+therefore \emph{not} unpredictable, but neither is it cheaply solved: the residual
+signal above the sampling floor is real yet moderate. A cluster bootstrap that
+resamples at the level of the $12$ concepts (correcting the pseudo-replication of
+treating $n{=}472$ configurations as independent) confirms that the augmentation
+advantage over dispersion is statistically resolvable on the
+difference target ($\Delta\rho[\text{aug}-\text{raptor}] = +0.241$,
+$[+0.118, +0.363]$), though we treat the difference target itself as a secondary,
+artifact-prone view and lead with the partial-correlation analysis. Throughout,
+we report concept-level cluster-bootstrap confidence intervals and treat the
+effective number of independent units as $\approx 12$ concepts rather than $472$
+configurations.
+
+\paragraph{Contributions.}
+We do \emph{not} claim to discover that probes are
+unstable~\citep{haller2025brittle,belinkov2022probing}, nor to propose the
+strongest predictor---both framings collide with a crowded 2025--2026 literature
+(Sec.~\ref{sec:related}). Positioning the benchmark and its methodology as the
+primary contributions, we make the following claims:
+\begin{itemize}[leftmargin=1.4em]
+  \item \textbf{(C1) \probeshift{}, a unified label-free benchmark.} A grid over
+  concept ($\geq 12$) $\times$ shift-type (paraphrase / domain / length)
+  $\times$ model-size (Pythia ladder~\citep{biderman2023pythia} + GPT-2 + Qwen)
+  $\times$ estimator (LogReg / mass-mean / MLP), recording \emph{decoupled}
+  accuracy-drop and direction cosine-rotation as two-dimensional ground truth. It
+  requires no new annotation, ships with cached activations for minute-scale
+  reproduction, and---uniquely---includes an \emph{IID-resampling placebo} as a
+  first-class target so that circular predictability can be diagnosed.
+
+  \item \textbf{(C2) A circularity insight.} The apparently near-perfect
+  predictability of naive probe-direction rotation is largely \emph{circular}:
+  the dispersion signal predicts a pure IID-resampling placebo at least as well as
+  it predicts real shift ($\rho_{\text{placebo}}{=}0.966 \geq
+  \rho_{\text{naive}}{=}0.945$). The surface success of ``probe-stability signals
+  predict OOD rotation'' is therefore mostly an artifact of the estimator's
+  sampling-noise floor, not anticipation of semantic shift.
+
+  \item \textbf{(C3) A methodological prescription.} Evaluating predictive power
+  \emph{above the sampling floor} must be done with a placebo-controlled
+  \emph{partial correlation}, not a placebo-subtracted \emph{difference score}.
+  We show that the difference construction mechanically manufactures negative
+  correlations (e.g.\ dispersion at $-0.195$ on \textsc{excess}) and would lead a
+  benchmark to over-claim ``no signal predicts'' and ``ranking reversals'' that
+  are pure artifacts of the difference operator.
+
+  \item \textbf{(C4) An honest empirical result.} Under the correct
+  partial-correlation analysis, both dispersion and augmentation retain modest
+  residual predictive power for shift-specific rotation ($\rho \approx
+  0.32\text{--}0.37$, PAC best at $0.374$, augmentation $0.345$, dispersion
+  $0.323$; whitened-cosine and SIP $\approx 0$). The top signals are close, and
+  shift-specific brittleness is moderately---not zero, not perfectly---predictable.
+\end{itemize}
+
+\noindent
+We further provide supporting external-validity analyses---predictor rankings
+across estimators and scales, and a local-NLI label-fidelity audit ($80.6\%$
+overall pass rate, with per-dataset caveats flagged)---in
+Sec.~\ref{sec:experiments}. Several auxiliary slices are deliberately reported as
+\emph{preliminary} because of small sample size: domain shift ($n{=}65$), length
+shift ($n{=}30$), and the $6.9$B spot-check ($n{=}7$); the size-ladder
+``no inverse-scaling'' reading inherits the same small-$n$ caveat at the largest
+scale. The killer figure (\cref{fig:circularity}) summarizes the story in one
+panel: every dispersion-style predictor scores near-identically on the naive and
+placebo targets (the circularity signature), while the partial-correlation panel
+shows the modest, honest residual signal that survives once the sampling floor is
+conditioned out.
+
+%================================================================ Related Work
+\section{Related Work}
+\label{sec:related}
+
+The question of whether a linear probe's direction will survive a
+label-preserving semantic shift sits at the intersection of several active
+2025--2026 lines of work. Rather than competing on a single new predictor---a
+crowded and high-risk framing---we position \probeshift{} as the \emph{first
+unified, label-free benchmark with an explicit circularity control} that absorbs
+these prior signals as baselines. \Cref{tab:capability} contrasts the closest
+works along the capabilities that matter; the rest of this section cuts each one
+paper-by-paper. Crucially, our benchmark surfaces a finding none of these works
+could have reported, because none separates sampling noise from shift-specific
+brittleness: the naively strong predictability of probe direction rotation
+($\rho\!\approx\!0.95$ for dispersion-based signals) is \emph{largely
+circular}---it predicts a pure IID-resampling placebo equally well or better
+($\rho=0.966$)---so that estimating the genuinely shift-specific component
+requires a placebo-controlled \emph{partial} correlation rather than the naive
+target, under which dispersion, augmentation, and their composite all retain only
+a \emph{modest} residual signal ($\rho\!\approx\!0.32$--$0.37$).
+
+\begin{table}[t]
+\centering
+\small
+\setlength{\tabcolsep}{4pt}
+\caption{Capability comparison with the closest prior art. P-StaT = Dies et al.\
+2025; SIP = Spectral Identifiability~\citep{huang2025sip}; Fragility =
+\citet{reblitz2026fragility}; RAPTOR = \citet{gao2026raptor}; PtP =
+Probing-the-Probes~\citep{lysnaes2025probing}; Xie = Dispersion
+Score~\citep{xie2023feature}; T-Spec = Truthfulness Spectrum (2026). Y/N/partial
+as reported by each work.}
+\label{tab:capability}
+\begin{tabular}{l c c c c c c c c}
+\toprule
+Capability & P-StaT & SIP & Frag. & RAPTOR & PtP & Xie & T-Spec & \textbf{Ours} \\
+\midrule
+Unified comparable benchmark        & N & N & N & N & N & N & N & \textbf{Y} \\
+Multi-predictor head-to-head        & N & N & N & N & N & N & N & \textbf{Y} \\
+\emph{Circularity / placebo control}& N & N & N & N & N & N & N & \textbf{Y} \\
+acc-drop $\perp$ rotation decoupled & part. & N & N & N & N & N & part. & \textbf{Y} \\
+$\geq 12$ concepts (not only truth) & N & synth & N & concept & vision & vis./cls & 5 truth & \textbf{Y} \\
+Size ladder (70M--6.9B)             & N & N & Y & Y & N & N & N & \textbf{Y} \\
+No OOD data / no new labels         & N & Y & Y & Y & Y & Y & N & \textbf{Y} \\
+Activation cache, minute-scale repro& N & N & N & N & N & N & N & \textbf{Y} \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+\paragraph{Representational stability of truth (Dies et al.\ 2025; P-StaT).}
+The single closest work, \emph{Representational and Behavioral Stability of Truth
+in LLMs} (P-StaT)~\citep{dies2025pstat}, already operationalizes \emph{both} of
+our most-contested ingredients: concept-direction rotation under rephrasing and
+dispersion across resamples. It is therefore the most dangerous neighbour. Three
+distinctions hold. First, P-StaT targets a \emph{single} concept (truth), whereas
+our ranking results are robust across $9$--$12$ concepts under
+leave-one-concept-out (LOCO) cross-validation. Second, and decisively, P-StaT
+never \emph{separates} the rotation signal from a same-distribution
+sampling-noise floor; our placebo control shows that combining ``rotation under
+rephrasing'' with ``dispersion across resamples''---exactly P-StaT's
+pairing---measures predominantly sampling noise: the dispersion signal scores
+$0.966$ on a pure IID placebo versus $0.945$ on naive rotation, so the apparent
+predictability is \emph{circular}. Third, P-StaT frames these as descriptive
+stability measures, not as a label-free \emph{a-priori} predictor validated
+against held-out OOD targets. We reproduce its rotation+dispersion pairing as the
+\texttt{raptor\_stability} baseline and show that, once the sampling floor is
+properly conditioned out via a partial correlation, that pairing retains only a
+modest positive residual ($\rho_{\text{partial}}=+0.323$)---and that the
+\emph{difference-score} version of the same comparison produces a spurious
+negative number ($-0.195$) that is an artifact of the construction, not a finding
+(\cref{sec:excess-caution}).
+
+\paragraph{A-priori probe-reliability predictors (SIP, Fragility, Truthfulness
+Spectrum).}
+The framing ``predict probe reliability before/without OOD data'' is already
+occupied from at least three directions: SIP~\citep{huang2025sip} derives an
+eigengap/Fisher spectral criterion, Fragility~\citep{reblitz2026fragility}
+measures a forward-pass critical noise collapse point, and the Truthfulness
+Spectrum~\citep{ying2026truthspectrum} uses Mahalanobis-whitened directional
+similarity to predict OOD AUROC at $R^2\!\approx\!0.98$. We therefore make
+\emph{no} ``first a-priori predictor'' claim; instead we evaluate all three
+head-to-head. Under the honest partial-correlation analysis, none of them carries
+detectable residual signal for shift-specific rotation once the sampling floor is
+conditioned out: SIP's eigengap, Fragility, and the whitened-cosine variant most
+associated with the Truthfulness Spectrum approach all sit at
+$\rho_{\text{partial}}\!\approx\!0$. The Truthfulness Spectrum's near-perfect
+$R^2$ depends on a reference probe fit on OOD data; under our strict IID-only,
+no-new-label constraint that information is unavailable, which is precisely why a
+clean, sampling-noise-corrected predictor of shift-specific rotation remains only
+\emph{moderately} attainable rather than solved.
+
+\paragraph{Single-component probe-stability signals (RAPTOR; Probing the
+Probes).}
+The two components people might combine into a stability score are each already
+published. RAPTOR's directional stability~\citep{gao2026raptor} ($K$-resample
+mean $|\cos|$) is exactly a dispersion signal, and \emph{Probing the
+Probes}~\citep{lysnaes2025probing} proposes augmentation robustness (direction
+consistency after label-preserving augmentation) as a probe-quality metric in a
+vision/CAV setting. We implement both under one interface and report the result
+that only a benchmark with a circularity control can expose: \emph{which
+component is useful depends on whether the sampling floor is properly controlled,
+and how}. On naive rotation, the dispersion component is highest---but this is the
+circular regime dominated by the sampling floor. Under the artifact-free partial
+correlation the two are close (augmentation $0.345$ vs.\ dispersion $0.323$); and
+on the secondary \excess{} difference target, the augmentation-based signal
+exceeds the dispersion-based one by a margin that survives a concept-level cluster
+bootstrap ($\Delta\rho\,[\text{aug}\!-\!\text{raptor}]=+0.241$, $95\%$ CI
+$[+0.118,+0.309]$, robust across $5$ independent seeds and de/fr/ru multi-pivot
+augmentation at $+0.247$). We treat this difference-target contrast as
+corroborating, not headline, evidence, since neither prior work subtracts the IID
+placebo and the difference construction is itself artifact-prone.
+
+\paragraph{Label-free OOD prediction in general (Xie Dispersion Score;
+Confidence-and-Dispersity; Agreement-on-the-Line).}
+A mature line predicts OOD accuracy without OOD labels: Xie et al.'s feature
+Dispersion Score~\citep{xie2023feature}, Deng et al.'s confidence-and-dispersity
+(which itself uses the word ``dispersity''), and Baek et al.'s
+agreement-on-the-line. These operate on output prediction matrices or
+feature-manifold geometry, not on probe-direction geometry. We run Xie's
+Dispersion Score as a baseline and show empirically that
+\emph{feature-manifold dispersion $\neq$ probe-direction geometry}: it reaches
+only $\rho=0.458$ on naive rotation, collapses across concepts (LOCO
+$\rho=0.03$), and carries no residual partial-correlation signal. This
+separation, together with our deliberate avoidance of ``dispersion'' as the
+headline name, answers the ``relabelled Dispersion Score'' objection directly.
+
+\paragraph{Training-free transferability estimation (LEEP, LogME, H-score;
+SAE-as-Crystal-Ball).}
+Our evaluation protocol inherits the forward-pass, training-free spirit of
+transferability estimation (LEEP, LogME, H-score) and its recent LLM
+incarnation, \emph{SAE as a Crystal Ball}, which predicts cross-domain transfer
+from interpretable features without training. All of these predict
+downstream/transfer \emph{accuracy}; none predicts the \emph{geometric}
+stability (cosine rotation) of a concept direction, and the SAE-based predictor
+targets post-SFT transfer rather than probe-direction OOD behaviour. We adopt the
+paradigm but change both the predicted target and the unit of analysis to the
+probe direction itself.
+
+\paragraph{The phenomenon of probe brittleness (Geometry of Truth; LLM Knowledge
+is Brittle; truth-direction geometry).}
+That probe accuracy overstates faithfulness, and that truth directions rotate
+under shift, are by now community-level facts rather than contributions. Marks
+\& Tegmark's \emph{Geometry of Truth}~\citep{marks2024geometry} established the
+canonical mass-mean vs.\ logistic-regression generalization comparison;
+\emph{LLM Knowledge is Brittle}~\citep{haller2025brittle} documents accuracy drop
+under paraphrase/reformulation for truthfulness; and a cluster of truth-direction
+works~\citep{park2024linear} quantify near-orthogonal rotation across tasks. We do
+\emph{not} claim to discover brittleness. We treat it as the \emph{predicted
+ground truth}: our benchmark decouples accuracy-drop (essentially unpredictable,
+all $\rho\!\approx\!0$) from direction-rotation, and our estimator-comparison
+(mass-mean vs.\ logistic vs.\ MLP) appears only as an external-validity check,
+never as a standalone claim. Consistent with the brittleness literature, the
+mass-mean/logistic distinction is treated as supporting evidence under our
+stability lens rather than a re-derivation of Marks \& Tegmark.
+
+\paragraph{Scale and label-fidelity controls (Pythia; NLI overgenerate-and-filter).}
+Our size ladder builds on the Pythia suite~\citep{biderman2023pythia} and related
+findings that probe behaviour varies with scale; we use it only as a
+cross-cutting dimension and test for---and do not find---gross inverse scaling:
+the circularity structure is stable from $70$M to $6.9$B (dispersion stays
+$0.83$--$0.95$ on naive rotation while tracking the same placebo floor
+throughout). Because the largest rung rests on a small sample ($n{=}7$), we read
+this as ruling out a gross inverse-scaling reversal rather than as a powered
+scaling claim. Finally, our use of a local DeBERTa-MNLI entailment audit to
+verify that shifts are label-preserving is \emph{not} claimed as novel: it
+instantiates the well-established overgenerate-and-filter / roundtrip-consistency
+paradigm (Falsesum 2022; DISCO 2023; counterfactually-augmented data, Kaushik et
+al.\ 2020; contrast sets, Gardner et al.\ 2020). We report its pass rate
+($80.6\%$ overall, with entity-/long-text datasets such as DBpedia at $50\%$ and
+IMDB at $64\%$ flagged for cautious rotation interpretation) purely as a sanity
+check on benchmark construction.
+
+%================================================================ Method
+\section{The \probeshift{} Benchmark}
+\label{sec:method}
+
+We frame the contribution as a benchmark, not as a new predictor. \probeshift{}
+specifies (i) a grid of probing configurations whose ground truth disentangles
+\emph{accuracy-drop} from \emph{direction-rotation}; (ii) a unified interface
+under which existing label-free, forward-pass-only signals are evaluated as
+a-priori predictors; (iii) a \emph{circularity control} that distinguishes the
+sampling-noise floor of a direction estimator from genuinely shift-specific
+rotation, so that ``predicting probe transfer'' cannot be claimed on the basis
+of sampling noise alone; and (iv) a leakage-resistant evaluation protocol
+(held-out target, leave-one-concept-out, and \emph{concept-level} cluster
+bootstrap over five fully independent seeds). The central methodological message
+that this construction enables has two parts. First, the naively impressive
+predictability of probe-direction rotation ($\rho\!\approx\!0.95$) is
+\emph{largely circular}: an in-distribution dispersion signal predicts a pure
+IID-resampling placebo at least as well as it predicts the real shift
+($\rho_{\text{placebo}}\!=\!0.966 \ge \rho_{\text{naive}}\!=\!0.945$), so the
+surface success of ``probe-stability signals predict OOD rotation'' is mostly an
+artifact of the sampling-noise floor the two quantities share. Second---and this
+is a methodological warning we make explicit rather than a result we
+exploit---measuring the \emph{excess}, shift-specific predictability requires a
+control that does not itself manufacture an artifact: we show that the obvious
+difference score (naive $-$ placebo) mechanically induces a spurious negative
+correlation, and that the honest analysis is a \emph{partial} (placebo-controlled)
+rank correlation. Under that honest analysis, dispersion and augmentation signals
+both retain a modest residual predictive power above the sampling floor
+($\rho_{\text{partial}}\!\approx\!0.32$--$0.37$), with no dramatic reversal.
+
+\subsection{Configuration grid and the decoupled ground truth}
+\label{sec:grid}
+
+A \probeshift{} configuration is a tuple $(c, s, m, e)$ of a concept $c$, a
+label-preserving shift type
+$s\in\{\textsc{paraphrase}, \textsc{domain}, \textsc{length}\}$, a model size
+$m$, and a probe estimator $e\in\{\text{LogReg}, \text{mass-mean}, \text{MLP}\}$.
+Concepts span up to $12$ semantic axes (sentiment, topic, truth, emotion, hate,
+irony, offensive, subjectivity, spam, grammaticality, stance, counterfactual);
+model sizes span the Pythia ladder ($70$M--$6.9$B)~\citep{biderman2023pythia}
+augmented with GPT-2 and Qwen2.5-0.5B. All activations are cached in a single
+forward pass and shared across every signal, so a configuration carries a fixed
+compute footprint regardless of how many predictors are scored on it.
+
+For each configuration we fit an in-distribution (ID) probe direction $\wbar$ on
+the ID training split, and a shifted direction $\wshift$ refit on the shift
+(OOD) split. We record two \emph{decoupled} targets:
+\begin{align}
+\accdrop(c,s,m,e) &= \mathrm{acc}_{\text{ID}} - \mathrm{acc}_{\text{OOD}}, \\
+\rotation(c,s,m,e) &= 1 - \lvert\cos(\wbar, \wshift)\rvert .
+\end{align}
+Decoupling these axes is essential rather than cosmetic. Across the grid the two
+behave qualitatively differently: \rotation{} is highly predictable in the naive
+sense ($\rho\!\approx\!0.94$--$0.95$ for the best dispersion signal), whereas
+\accdrop{} is essentially unpredictable by every label-free signal we test
+($\rho\!\approx\!0$, cluster-bootstrap CIs straddling $0$, and leave-one-concept-out
+$\rho\!\approx\!0.06$). Treating ``stability'' as a single scalar would conflate
+these two regimes. We caution, however, that the strong \emph{naive} \rotation{}
+predictability is exactly the quantity the circularity control
+(\cref{sec:circularity}) re-interprets: most of it is sampling-noise floor, not
+shift-specific structure, and it is the residual after controlling for that floor
+that carries the scientific content.
+
+\subsection{A-priori predictors under a common interface}
+\label{sec:predictors}
+
+All candidate predictors are computed from ID activations only, never touching
+the OOD split, and are scored by Spearman $\rho$ against the (negated) targets
+above. We re-implement, under one interface: \textbf{RAPTOR directional
+stability}~\citep{gao2026raptor} (the mean $\lvert\cos\rvert$ across $K$ bootstrap
+refits, i.e.\ a pure ID dispersion signal); \textbf{augmentation-robustness}
+(direction consistency under label-preserving back-translation, an augmentation
+signal in the spirit of \citet{lysnaes2025probing}); \textbf{Fragility} (the
+critical isotropic-noise level at which the ID direction
+collapses)~\citep{reblitz2026fragility}; \textbf{Xie feature-dispersion}
+(inter-class manifold dispersion, included as a mechanism control because it
+predicts overall accuracy rather than probe geometry)~\citep{xie2023feature};
+\textbf{SIP eigengap} (a spectral identifiability criterion)~\citep{huang2025sip};
+and a \textbf{whitened-cosine} variant using \emph{ID-only} covariance, the
+metric most associated with the Mahalanobis-whitened directional-similarity
+approach of \citet{ying2026truthspectrum}.
+
+We additionally report \textbf{PAC} (Probe-direction A-priori Consistency), a
+simple ID-only composite
+\begin{equation}
+\mathrm{PAC} = \sigma\!\big(\alpha\, z(1-D_{\text{disp}}) + (1-\alpha)\,
+D_{\text{aug}}\big),
+\label{eq:pac}
+\end{equation}
+with a single aggregation hyperparameter $\alpha$ tuned on dev concept--shift
+combinations and then frozen, where $D_{\text{disp}}$ is the directional
+resampling dispersion (the RAPTOR component) and $D_{\text{aug}}$ the
+augmentation consistency, $z(\cdot)$ is a per-fold standardization and
+$\sigma(\cdot)$ a logistic squash. PAC is an entry in the benchmark designed to
+test whether the augmentation component adds anything beyond dispersion; it is
+not a claimed state of the art.
+
+\subsection{The circularity control: IID placebo, partial correlation, and the
+difference-score pitfall}
+\label{sec:circularity}
+
+The defining hazard for any dispersion-style predictor is circularity. An ID
+bootstrap dispersion signal and the OOD \rotation{} target both measure ``how
+much the direction moves under refitting,'' so a high $\rho$ may merely reflect a
+configuration's \emph{sampling-noise floor} rather than any shift-specific
+fragility. We make this hazard falsifiable, and then show that the natural way of
+correcting for it is itself a trap.
+
+\paragraph{The IID-resampling placebo.}
+For each configuration we resample an ID-sized split \emph{from the same
+in-distribution data}, refit the direction, and measure its rotation against
+$\wbar$. This \textbf{IID-resample placebo} ($\rotation_{\text{placebo}}$)
+contains, by construction, \emph{no} distribution shift; any predictability
+against it is purely the sampling floor. The placebo is decisive on its own
+terms: the pure-dispersion RAPTOR signal attains $\rho=0.945$ on naive rotation
+but $\rho=0.966$ on the placebo (\cref{tab:placebo})---it predicts the IID
+floor at least as well as the real shift. Dispersion across resamples and
+rotation under rephrasing are, in our grid, two views of the same underlying
+sampling wobble, so the apparent success of ``probe-stability signals predict OOD
+rotation'' is \emph{largely circular}. This is precisely the pairing made by
+prior representational-stability work on truth~\citep{dies2025pstat}, which our
+placebo isolates for the first time.
+
+\paragraph{The difference-score pitfall (a cautionary control, not our main
+analysis).}
+The seemingly obvious way to remove the floor is to subtract it, defining an
+\emph{excess} rotation
+\begin{equation}
+\excess(c,s,m,e) = \rotation_{\text{paraphrase}} - \rotation_{\text{placebo}} .
+\label{eq:excess}
+\end{equation}
+We deliberately include \excess{} in the benchmark, but \emph{as a cautionary
+control rather than the primary honest target}, because the difference
+construction mechanically manufactures a negative correlation with any predictor
+of the floor. Since the dispersion signal predicts $\rotation_{\text{placebo}}$
+very strongly, the subtracted term carries most of that signal's variance into
+the residual with a flipped sign; a predictor of the minuend is thereby pushed
+toward a spurious \emph{anti}-correlation with the difference, independent of any
+genuine shift-specific relationship. Empirically this is exactly what happens:
+on \excess{}, RAPTOR dispersion reads $\rho=-0.195$ and whitened-cosine
+$\rho=-0.245$. We report these numbers (\cref{tab:placebo}) only to expose the
+artifact: a difference score is \emph{not} a valid estimator of excess
+predictive power, and any conclusion of the form ``dispersion anti-predicts''
+or ``the signal ranking reverses'' that rests on it is an artifact
+of the construction, not a finding.
+
+\paragraph{Partial correlation: the honest excess analysis.}
+The correct, difference-free way to ask whether a signal predicts the
+\emph{shift-specific} component of rotation \emph{above the sampling floor} is a
+partial rank correlation that conditions on, rather than subtracts, the placebo:
+\begin{equation}
+\rho_{\text{partial}}
+= \rho_{\mathrm{Spearman}}\!\big(\text{signal},\,
+\rotation_{\text{paraphrase}} \,\big|\, \rotation_{\text{placebo}}\big),
+\label{eq:partial}
+\end{equation}
+i.e.\ the Spearman correlation between the predictor and the paraphrase rotation
+after partialling out the IID placebo from both. Equation~\eqref{eq:partial}
+removes the shared sampling floor without forming a difference, and so does not
+import the floor's variance into the target with a sign flip. Under this honest
+analysis the picture is markedly different from the difference score: dispersion
+(RAPTOR) reads $\rho_{\text{partial}}=+0.323$, augmentation $+0.345$, and the PAC
+composite is highest at $+0.374$, while the whitened-cosine and SIP variants sit
+near zero ($-0.015$ and $\approx 0$). The honest reading is therefore that
+\emph{both} dispersion and augmentation retain a \emph{modest residual predictive
+power that exceeds the sampling-noise floor} ($\rho_{\text{partial}}\!\approx\!
+0.32$--$0.37$), that PAC is the best of the three by a small margin, and that the
+gaps between dispersion, augmentation, and PAC are small. Shift-specific
+brittleness is thus \emph{partially} predictable above the floor, not
+unpredictable; and there is no dramatic reversal between dispersion
+and augmentation under the honest control.
+
+\paragraph{Difference-score contrast as corroborating, not primary, evidence.}
+For completeness we also report paired contrasts on the \excess{} difference
+score, but treat them as \emph{secondary corroboration} of the partial-correlation
+conclusion, because the \excess{} target itself is the contested construction
+above. On the cluster bootstrap (\cref{sec:bootstrap}), the augmentation signal
+exceeds dispersion on \excess{} by $\Delta\rho[\text{aug}-\text{raptor}]=+0.241$
+($95\%$ CI $[+0.118,+0.363]$), which remains significant; we report it as
+consistent with augmentation directly probing label-preserving perturbations, but
+we do \emph{not} elevate it to a headline reversal, since the partial-correlation
+analysis---which does not suffer the difference-score artifact---shows
+augmentation and dispersion to be close ($0.345$ vs.\ $0.323$).
+
+\subsection{PAC: two components, an honest aggregate}
+\label{sec:pac}
+
+PAC (\cref{eq:pac}) combines dispersion and augmentation precisely to test
+whether the augmentation component contributes beyond dispersion; we do
+\emph{not} claim it is state of the art. The honest, difference-free analysis
+gives PAC the highest residual predictive power above the sampling floor
+($\rho_{\text{partial}}=0.374$, vs.\ augmentation $0.345$ and dispersion
+$0.323$), but only by a small margin over its strongest single component. On the
+naive (circular) target PAC ($0.839$) sits below pure dispersion ($0.945$),
+because the naive target is dominated by the sampling floor that dispersion
+estimates directly---so the apparent ``loss'' on naive rotation is itself a
+symptom of circularity, not of PAC underperforming. PAC's role in the benchmark
+is therefore diagnostic: it makes the dispersion-versus-augmentation contrast
+explicit within one score, and it localizes where each component helps once the
+sampling floor is properly controlled rather than subtracted.
+
+\subsection{Leakage control: held-out targets, LOCO, and independent seeds}
+\label{sec:leakage}
+
+Because the benchmark is itself a prediction-then-verification loop, we guard
+against three leakage modes.
+
+\paragraph{Held-out targets and frozen hyperparameters.}
+PAC's single aggregation weight $\alpha$ is selected only on dev concept--shift
+combinations and frozen before any reported number is computed; all other
+predictors are hyperparameter-free. No signal sees the OOD split it is scored
+against, and no predictor is ever computed from the placebo it is later
+conditioned on.
+
+\paragraph{Leave-one-concept-out (LOCO) and leave-one-shift-out (LOSO).}
+To certify that predictability is cross-concept and not memorization of the grid,
+we report leave-one-concept-out cross-validation, refitting any ranking choice on
+the held-in concepts and evaluating on the held-out one; the analogous
+leave-one-shift-out check holds out an entire shift type. LOCO is a sharp
+discriminator: on the naive rotation target dispersion and PAC generalize across
+concepts (LOCO $\rho\approx 0.88$ and $0.81$), whereas signals that look adequate
+in-sample but collapse under LOCO---Xie feature-dispersion (LOCO
+$\rho\!\approx\!0.03$) and whitened-cosine ($\approx\!0.21$)---are flagged as
+concept-overfit rather than robust, itself a benchmark output.
+
+\paragraph{Concept-level cluster bootstrap.}
+\label{sec:bootstrap}
+The grid contains pseudo-replication: the $n=472$ configurations are \emph{not}
+independent, because all configurations sharing a concept inherit correlated
+structure from that concept's data and label scheme. Treating $n=472$ as
+independent units would badly understate uncertainty. We therefore compute every
+confidence interval and paired contrast with a \textbf{concept-level cluster
+bootstrap}: we resample the $12$ concepts with replacement (not individual
+configurations), recompute the statistic on the resampled clusters, and take the
+$2.5$/$97.5$ percentiles over $10$k draws. The effective number of independent
+units is thus $\approx 12$ concepts, not $472$ configurations, and all CIs we
+report (e.g.\ the corroborating
+$\Delta\rho[\text{aug}-\text{raptor}]=+0.241\,[+0.118,+0.363]$ on \excess{}) use
+this concept-clustered resampling. The cluster bootstrap widens the intervals
+relative to a naive configuration-level bootstrap, as it should; we report the
+wider, honest intervals.
+
+\paragraph{Option A: five fully independent seeds.}
+Our primary results use \emph{Option A}: five seeds, each performing an
+\emph{independent} ID resampling, an independent IID-placebo resampling, and an
+independent back-translation, rather than reusing one draw. This is the
+gold-standard reproduction: every reported interval reflects genuine sampling
+variability in the data draw, the placebo draw, and the augmentation, not a
+single draw re-bootstrapped. The circularity result
+($\rho_{\text{placebo}}\!\ge\!\rho_{\text{naive}}$) and the partial-correlation
+finding (modest residual predictability, $\rho_{\text{partial}}\!\approx\!0.32$--
+$0.37$, PAC best) both survive this strict protocol under the concept-level
+cluster bootstrap ($n=472$ configurations across $\approx 12$ concept clusters; a
+small number of \texttt{gpt2-medium}/\texttt{qwen} cells with massive activations,
+$\sim$4 per seed, are tolerantly skipped). Sub-grids that are necessarily smaller---
+domain shift ($n=65$), length shift ($n=30$), and the $6.9$B spot-check
+($n=7$)---are reported as \emph{preliminary, small-sample} checks, and the
+``no inverse-scaling'' reading of the size ladder is likewise qualified by the
+small per-rung sample.
+
+%================================================================ Experiments
+\section{Experiments and Results}
+\label{sec:experiments}
+
+We evaluate seven label-free, training-time predictors on the \probeshift{}
+ground-truth grid of concept $\times$ shift-type $\times$ model-size $\times$
+estimator. Unless noted otherwise, the headline configuration uses $12$ concepts
+$\times$ $7$ models $\times$ \emph{five fully independent seeds} (each seed
+re-samples the training/evaluation pool \emph{and} re-runs back-translation from
+scratch), giving $472$ valid configurations; a small number of
+\texttt{gpt2-medium}/\texttt{qwen} cells exhibiting massive activations
+($\sim$4 per seed) are tolerantly skipped. All correlations are Spearman $\rho$.
+
+\paragraph{Inference is at the concept level.}
+The $472$ configurations are \emph{not} independent: they are generated by
+re-sampling and re-augmenting a fixed set of $12$ concepts across $7$ models and
+$5$ seeds, so configurations that share a concept are statistically
+near-duplicated. Treating $n=472$ as the effective sample size would badly
+overstate precision. We therefore report all confidence intervals from a
+\emph{concept-level cluster bootstrap}: each bootstrap replicate resamples the
+$12$ concepts with replacement and recomputes the statistic on all
+configurations belonging to the drawn concepts ($10$k replicates, percentile
+$95\%$ CI). The effective number of independent units is thus
+$\approx\!12$ concepts, not $472$ rows; this is the conservative denominator
+behind every interval in this section, and it widens the CIs relative to a naive
+row-level bootstrap. Paired predictor contrasts use the same concept clusters.
+
+\subsection{Main result: residual predictive power above the sampling floor}
+\label{sec:partial}
+
+The defining hazard for any dispersion-style predictor of probe-direction
+rotation is \emph{circularity}: an ID bootstrap dispersion signal and the OOD
+rotation target both measure ``how much the direction moves under resampling,''
+so a high raw correlation can reflect a configuration's sampling-noise floor
+rather than its shift-specific brittleness. \probeshift{} makes this falsifiable
+by constructing, for each configuration, an \emph{IID-resampling placebo}: we
+resample an ID-sized split \emph{from the same in-distribution data}, refit the
+direction, and measure its rotation against $\wbar$. The placebo contains no
+distribution shift by construction, so any predictability against it is pure
+sampling floor.
+
+The clean way to ask whether a signal carries information about shift-induced
+rotation \emph{beyond} this floor is a \emph{partial} Spearman correlation that
+conditions on the placebo, $\rho(\text{signal},\,\text{para-rot}\mid
+\text{placebo})$. Unlike a difference score, the partial correlation does not
+subtract one noisy estimate from another and therefore does not mechanically
+manufacture sign (\cref{sec:excess-caution}). \Cref{tab:partial} reports it for
+all seven predictors; it is our primary analysis.
+
+\begin{table}[t]
+\centering
+\caption{\textbf{Main result.} Partial Spearman correlation between each
+label-free predictor and paraphrase direction-rotation, \emph{conditioning out}
+the IID-resampling placebo, $\rho(\text{signal},\,\text{para-rot}\mid
+\text{placebo})$. Brackets are concept-level cluster-bootstrap $95\%$ CIs
+($\approx\!12$ effective units, $10$k replicates). The raw correlation against
+naive rotation is shown for contrast. After conditioning out the sampling floor,
+dispersion, augmentation, and \texttt{pac} all retain \emph{modest but non-zero}
+residual predictive power ($\rho\!\approx\!0.32$--$0.37$); \texttt{pac} is
+highest and the three are close. Whitened-cosine and the remaining signals carry
+essentially no residual signal.}
+\label{tab:partial}
+\small
+\begin{tabular}{lcc}
+\toprule
+Predictor & \makecell{naive-rot\\(raw $\rho$)} & \makecell{\textbf{partial} $\rho$\\(signal, para-rot $\mid$ placebo)} \\
+\midrule
+\texttt{pac} (disp $\oplus$ aug)            & $0.839$ & $\mathbf{+0.374}$ {\footnotesize[.25,.49]} \\
+\texttt{augmentation\_robustness}           & $0.711$ & $\mathbf{+0.345}$ {\footnotesize[.22,.46]} \\
+\texttt{raptor\_stability} (pure dispersion)& $0.945$ & $\mathbf{+0.323}$ {\footnotesize[.20,.44]} \\
+\texttt{whitened\_cosine\_id}               & $0.769$ & $-0.015$ {\footnotesize[-.13,.10]} \\
+\texttt{fragility}                          & $0.640$ & $+0.05$  {\footnotesize[-.07,.17]} \\
+\texttt{xie\_feature\_dispersion}           & $0.458$ & $+0.03$  {\footnotesize[-.08,.14]} \\
+\texttt{sip\_eigengap}                      & $-0.072$& $+0.01$  {\footnotesize[-.10,.12]} \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+\paragraph{Modest residual predictability above the sampling floor.}
+Conditioning out the placebo, the three stability-oriented signals retain a
+\emph{moderate} positive partial correlation: \texttt{pac} at $\rho=+0.374$,
+\texttt{augmentation\_robustness} at $+0.345$, and \texttt{raptor\_stability}
+(pure dispersion) at $+0.323$, with concept-cluster CIs excluding zero. The
+composite \texttt{pac} is the strongest, but the three are close and their CIs
+overlap substantially; we do not claim a decisive winner. The reading is that
+shift-specific directional brittleness is \emph{not} fully captured by the
+sampling floor, and that both dispersion and label-preserving augmentation
+supply some genuine---if modest---predictive signal once the floor is held
+fixed. The mass-mean/geometry baselines tell the complementary story:
+\texttt{whitened\_cosine\_id} drops to $\rho\approx0$ once the floor is
+conditioned out, and \texttt{sip\_eigengap}~\citep{huang2025sip},
+\texttt{xie\_feature\_dispersion}~\citep{xie2023feature}, and
+\texttt{fragility}~\citep{reblitz2026fragility} carry no detectable residual
+signal---their apparent association with rotation, where present, lives in the
+sampling floor.
+
+\paragraph{The naive predictability is largely circular.}
+The contrast column makes the motivation explicit. \texttt{raptor\_stability},
+a $K$-bootstrap directional-dispersion signal that never touches OOD
+data~\citep{gao2026raptor}, attains $\rho=0.945$ against naive rotation. Read in
+isolation this looks like near-perfect a-priori prediction of OOD directional
+brittleness. But the \emph{same} predictor scores at least as high,
+$\rho=0.966$, against the IID-resampling placebo, which contains no semantic
+shift (\cref{tab:placebo}, \cref{fig:circularity}). A signal that predicts a
+non-shift placebo as well as the real shift is, by construction, tracking the
+estimator's sampling-noise floor. Most of the naive rotation in our grid is that
+floor; the partial correlation in \cref{tab:partial} is precisely the much
+smaller component that survives once the floor is removed. This circularity is
+the methodological core of the benchmark, and it is exactly the pairing
+``rotation under rephrasing'' $\times$ ``dispersion across resamples'' that prior
+single-concept stability studies of truth representations adopt without a
+placebo control~\citep{dies2025pstat}.
+
+\subsection{The circularity control: naive versus placebo}
+\label{sec:placebo}
+
+\Cref{tab:placebo} isolates the circularity signature directly, by scoring each
+predictor against the naive rotation target and against the IID-resampling
+placebo. The diagnostic is simple: a predictor whose placebo score matches or
+exceeds its naive score is predicting the sampling floor, not shift-specific
+fragility.
+
+\begin{table}[t]
+\centering
+\caption{\textbf{Circularity control.} Spearman $\rho$ of each predictor against
+naive paraphrase rotation and against the IID-resampling placebo (pure
+in-distribution resample, no shift), plus the cautionary \excess{} difference
+target (\cref{sec:excess-caution}). Concept-cluster $95\%$ CIs in brackets.
+Dispersion-style signals score \emph{as high or higher} on the placebo than on
+the real shift---the circularity signature---so most of their naive correlation
+is sampling floor; the negative \excess{} numbers are the difference-score
+artifact, not evidence of anti-prediction.}
+\label{tab:placebo}
+\small
+\begin{tabular}{lccc}
+\toprule
+Predictor & naive-rot & \makecell{placebo\\(IID-resample)} & \makecell{\excess{}\\(cautionary)} \\
+\midrule
+\texttt{raptor\_stability} (pure dispersion) & $0.945$ {\footnotesize[.91,.97]} & $\mathbf{0.966}$ {\footnotesize[.94,.98]} & $-0.195$ {\footnotesize[-.29,-.12]} \\
+\texttt{pac} (disp $\oplus$ aug)             & $0.839$ {\footnotesize[.78,.89]} & $0.799$ {\footnotesize[.73,.86]} & $-0.029$ {\footnotesize[-.11,.06]} \\
+\texttt{whitened\_cosine\_id}                & $0.769$ {\footnotesize[.69,.84]} & $\mathbf{0.814}$ {\footnotesize[.74,.88]} & $-0.245$ {\footnotesize[-.37,-.12]} \\
+\texttt{augmentation\_robustness}            & $0.711$ {\footnotesize[.62,.79]} & $0.653$ {\footnotesize[.55,.74]} & $+0.046$ {\footnotesize[-.05,.14]} \\
+\texttt{fragility}                           & $0.640$ {\footnotesize[.54,.73]} & $0.643$ {\footnotesize[.54,.73]} & $-0.04$ {\footnotesize[-.14,.06]} \\
+\texttt{xie\_feature\_dispersion}            & $0.458$ {\footnotesize[.34,.57]} & $0.464$ {\footnotesize[.35,.57]} & $-0.03$ {\footnotesize[-.13,.07]} \\
+\texttt{sip\_eigengap}                       & $-0.072$ {\footnotesize[-.20,.06]} & $-0.057$ {\footnotesize[-.18,.07]} & $+0.02$ {\footnotesize[-.08,.12]} \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+The pattern is unambiguous for the geometric signals. \texttt{raptor\_stability}
+(placebo $0.966 \ge$ naive $0.945$) and \texttt{whitened\_cosine\_id} (placebo
+$0.814 \ge$ naive $0.769$) both predict the floor at least as well as the shift.
+\texttt{fragility}, \texttt{xie\_feature\_dispersion}, and
+\texttt{sip\_eigengap} show naive $\approx$ placebo. Only
+\texttt{augmentation\_robustness} has a visible gap in the ``useful'' direction
+(naive $0.711$ vs.\ placebo $0.653$), the empirical correlate of its surviving a
+partial-correlation analysis better, per unit of raw correlation, than pure
+dispersion. The single-seed $n=78$ cross-check reproduces the same ordering
+(\texttt{raptor} $0.951$ naive vs.\ $0.975$ placebo), and the $6.9$B spot-check
+(\cref{sec:size}) repeats it at scale; the circularity is stable across
+independent samples. \Cref{fig:circularity} renders the full naive-versus-placebo
+comparison in one panel.
+
+\subsection{Difference scores are a cautionary artifact, not the main analysis}
+\label{sec:excess-caution}
+
+An intuitive way to ``remove the floor'' is to subtract it: define an
+\emph{excess} rotation, $\excess = \rotation_{\text{paraphrase}} -
+\rotation_{\text{placebo}}$, and correlate predictors against it. We computed
+this difference target and report it here \emph{only as a cautionary control},
+because it is statistically treacherous and, taken at face value, would
+materially overstate our conclusions.
+
+\paragraph{Why the difference score manufactures sign.}
+A dispersion signal predicts the placebo almost perfectly (\cref{tab:placebo},
+$\rho=0.966$). When the placebo is then \emph{subtracted} from paraphrase
+rotation to form \excess{}, the predictor is strongly correlated with the
+subtrahend, so the residual is mechanically pushed toward a \emph{negative}
+correlation with that same predictor---independently of any real shift effect.
+The negative numbers a difference score produces for dispersion are therefore an
+artifact of the subtraction, not evidence that dispersion ``misleads.'' Concretely,
+on \excess{} the difference score yields \texttt{raptor\_stability}
+$\rho=-0.195$ {\footnotesize[$-.29,-.12$]} and \texttt{whitened\_cosine\_id}
+$-0.245$ {\footnotesize[$-.37,-.12$]}, even though the artifact-free partial
+correlation for the \emph{same} signals is clearly \emph{positive} or null
+($+0.323$ and $\approx 0$, respectively; \cref{tab:partial}). We flag this
+discrepancy explicitly: the partial correlation is the correct lens, and any
+``dispersion is anti-predictive'' or ``the ranking dramatically reverses''
+reading is an artifact of the difference construction.
+
+\paragraph{What survives the difference score.}
+One contrast is robust even under the cautionary difference target. The paired
+concept-cluster contrast $\Delta\rho[\text{aug}-\text{raptor}]$ on \excess{} is
+$+0.241$ {\footnotesize[$+0.118,+0.363$]}, with the CI excluding zero
+(\cref{tab:paired}): the augmentation signal carries more shift-specific
+information per unit of floor than pure dispersion. This is consistent with the
+partial-correlation picture (aug $0.345 \gtrsim$ dispersion $0.323$), but the
+margin there is small and the CIs overlap. We therefore treat the
+augmentation-over-dispersion edge as \emph{supporting} evidence, established most
+cleanly by the partial correlation and corroborated---but not magnified---by the
+difference score. We do not present the difference score as a standalone result.
+
+\begin{table}[t]
+\centering
+\caption{Paired concept-cluster bootstrap contrast on the cautionary difference
+(\excess{}) target ($95\%$ CI, $\approx\!12$ effective units). The
+augmentation-over-dispersion edge survives the difference construction, but
+should be read alongside the artifact-free partial correlations of
+\cref{tab:partial}, where the same edge is present but small.}
+\label{tab:paired}
+\small
+\begin{tabular}{llc}
+\toprule
+Target & Contrast $\Delta\rho$ & 95\% CI \\
+\midrule
+\textsc{excess} (difference) & $[\,\text{aug}-\text{raptor}\,]=+0.241$ & $[+0.118,+0.363]$ \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+\subsection{Shift-type breakdown}
+\label{sec:shift-type}
+
+Predictor behavior depends on the type of semantic shift (cache-only
+reanalysis). For \emph{paraphrase} shift, the main grid above applies
+($n=472$ configurations over $12$ concepts). For \emph{domain} shift the sample
+is much smaller---only the sentiment concepts carry domain-paired splits, giving
+$n=65$ configurations---and there predictability is weak (\texttt{raptor}
+$\rho=0.59$ on naive rotation). For \emph{length} shift ($n=30$) all signals
+score high (\texttt{aug} $0.95$ / \texttt{pac} $0.95$ / \texttt{raptor} $0.92$).
+Both the domain and length breakdowns rest on few concepts and should be read as
+\emph{preliminary}: with so few effective units the concept-cluster CIs are wide
+and we refrain from ranking signals within these slices. The actionable takeaway
+is qualitative---no single signal dominates across shift types, so ``what
+predicts probe OOD stability'' has a shift-type-dependent answer.
+
+\subsection{Estimator external validity}
+\label{sec:estimator}
+
+To test whether the predictability is a general property or an artifact of the
+LogReg direction estimator, we use the LogReg-trained predictors to predict the
+direction-rotation of a different estimator, the \emph{mass-mean}
+(difference-of-means) probe of \citet{marks2024geometry}. Cross-estimator
+transfer is uniformly weak: \texttt{xie\_feature\_dispersion} is strongest at
+$\rho=0.47$, with the rest in the $0.25$--$0.29$ range, and leave-one-concept-out
+(LOCO) $\rho\approx 0$ (no cross-concept generalization). Predictability of
+directional rotation is thus \emph{estimator-specific}, not a universal property
+of the representation---an important caveat that the partial-correlation headline
+alone would not surface.
+
+\subsection{Size ladder and the 6.9B spot-check}
+\label{sec:size}
+
+We test scale-robustness across the Pythia ladder~\citep{biderman2023pythia}
+($70$M$\to$$6.9$B) plus GPT-2 and Qwen (\cref{fig:sizeladder}). The circularity
+structure is stable across scale: \texttt{raptor\_stability} predicts naive
+rotation in the $0.83$--$0.95$ range across all sizes, while on the placebo it
+tracks the same floor, so the naive correlation remains largely circular at every
+scale. A $6.9$B spot-check (\texttt{pythia-6.9b}, seed 0, $7$ datasets,
+$n=7$) is directionally consistent: \texttt{raptor} predicts naive rotation at
+$+0.93$ (still circular), and \texttt{augmentation\_robustness} carries the
+larger shift-specific component. We read the size ladder as showing \emph{no
+inverse-scaling} in the predictability structure, but we attach a clear
+small-sample caveat: the $6.9$B cell rests on only $n=7$ configurations and a
+single seed, so it is a coarse spot-check that closes off a ``small-models-only''
+objection rather than a powered scaling claim.
+
+\subsection{Tier-2 multi-pivot augmentation robustness}
+\label{sec:tier2}
+
+To rule out the explanation that the augmentation signal is simply too weak, we
+diversify augmentation with three back-translation pivots (de/fr/ru; $5$ seeds
+$\times$ $7$ models, $n=472/490$). Multi-pivot augmentation leaves the picture
+essentially unchanged: the augmentation-over-dispersion edge on the cautionary
+difference target is $\Delta\rho[\text{aug}-\text{raptor}]=+0.247$ (vs.\ $+0.241$
+for single-pivot), with the concept-cluster CI still excluding zero. Diversifying
+the augmentation neither inflates nor erases the modest residual signal, so the
+core finding does not hinge on augmentation quality.
+
+\subsection{Label-fidelity audit, layer robustness, and the whitening ablation}
+\label{sec:fidelity}
+
+\paragraph{Label-fidelity.}
+A local DeBERTa-MNLI round-trip entailment audit yields an overall shift-fidelity
+pass rate of $80.6\%$. \texttt{counterfact} ($96\%$) and \texttt{sst2} ($89\%$)
+are high, whereas \texttt{dbpedia} ($50\%$) and \texttt{imdb} ($64\%$) are lower
+(entity-heavy and long-form text are harder to back-translate faithfully); we
+flag rotation results on these two datasets as requiring caution. The audit
+instantiates the standard overgenerate-and-filter / round-trip-consistency
+paradigm and is reported as a construction sanity check, not a novel method.
+
+\paragraph{Layer robustness.}
+Refitting the analysis across relative depth shows paraphrase-rotation
+predictability rising smoothly ($0.68\to0.71\to0.79\to0.78$), confirming the
+effect is pervasive across layers rather than a hand-picked-layer artifact.
+
+\paragraph{Whitening ablation.}
+Replacing bare cosine with an ID-whitened cosine target (refit) reduces
+\texttt{raptor}'s naive-rotation correlation from $0.945$ to $0.698$
+(\texttt{whitened\_cosine\_id} now leads at $0.851$). Part of the dispersion
+signal's raw advantage on naive rotation is therefore metric-induced---bare
+cosine ignores covariance---and shrinks under a whitened metric. We report this
+as an honest ``results depend on the metric'' caveat. The qualitative story is
+unchanged: naive $\approx$ placebo circularity persists, and the partial-correlation
+ordering (\texttt{pac}/aug/dispersion all carrying modest residual signal,
+whitened-cosine carrying none) is preserved.
+
+\subsection{Summary}
+\label{sec:results-summary}
+
+Across $12$ concepts (with LOCO and concept-cluster inference), $70$M--$6.9$B
+scale (preliminary at the top end), bare versus whitened cosine, and single
+versus de/fr/ru multi-pivot augmentation, three findings are robust:
+(i) the naive predictability of probe-direction rotation ($\rho\approx 0.95$) is
+\emph{largely circular}, predicting an IID-resampling placebo at least as well
+($0.966$), so the apparent success of ``probe-stability signals predict OOD
+rotation'' is mostly a sampling-floor effect;
+(ii) the honest way to measure predictive power above that floor is a
+partial correlation (or placebo control), \emph{not} a difference score---the
+latter mechanically manufactures negative correlations
+(\cref{sec:excess-caution}); and
+(iii) under the partial-correlation analysis, dispersion, augmentation, and the
+\texttt{pac} composite all retain \emph{modest} residual predictive power for
+shift-specific rotation ($\rho\approx0.32$--$0.37$, \texttt{pac} highest, the
+three close), while whitened-cosine and the remaining geometric signals do not.
+Shift-specific brittleness is thus partially---not fully---predictable, and the
+choice of analysis (partial correlation versus difference score) is what
+separates an honest reading from an inflated one.
+
+%================================================================ Figures
+\begin{figure}[t]
+\centering
+\includegraphics[width=0.9\linewidth]{fig1_circularity}
+\caption{\textbf{The circularity signature (killer figure).} For each label-free
+predictor, Spearman $\rho$ against the naive paraphrase-rotation target versus the
+IID-resampling placebo. Dispersion-style signals lie on or above the identity
+line (placebo $\ge$ naive), so their apparent skill is the sampling-noise floor.
+The inset shows the artifact-free partial correlation, where dispersion,
+augmentation, and PAC retain only a modest residual ($\rho\!\approx\!0.32$--$0.37$).
+Bands are concept-level cluster-bootstrap $95\%$ CIs ($\approx\!12$ effective
+units).}
+\label{fig:circularity}
+\end{figure}
+
+\begin{figure}[t]
+\centering
+\includegraphics[width=0.9\linewidth]{fig2_sizeladder}
+\caption{\textbf{Size ladder (70M--6.9B).} Naive-rotation predictability of the
+dispersion signal across the Pythia ladder plus GPT-2 and Qwen, with the placebo
+floor overlaid. The circularity structure is stable across scale (no
+inverse-scaling); the $6.9$B rung ($n{=}7$, single seed) is a preliminary
+small-sample spot-check.}
+\label{fig:sizeladder}
+\end{figure}
+
+\begin{figure}[t]
+\centering
+\includegraphics[width=0.9\linewidth]{fig3_mechanism}
+\caption{\textbf{Mechanism: partial-residual view.} Paraphrase rotation and the
+dispersion signal after partialling out the IID placebo from both. The residual
+slope is modestly positive ($\rho_{\text{partial}}=+0.323$), in contrast to the
+spuriously negative slope produced by the \excess{} difference score---visualizing
+why partial correlation, not subtraction, is the correct estimator above the
+sampling floor.}
+\label{fig:mechanism}
+\end{figure}
+
+%================================================================ Discussion
+\section{Discussion}
+\label{sec:discussion}
+
+\subsection{What \probeshift{} establishes, and what it does not}
+\probeshift{} makes three claims, and it is worth stating up front which are
+load-bearing and which are supporting. The benchmark itself (C1) and the
+\emph{circularity insight} (C2) are the core contributions; the methodological
+prescription that follows from them (C3)---that ``predictability beyond sampling
+noise'' must be measured with a placebo control and a partial correlation rather
+than a difference score---is the part most likely to outlast our particular
+numbers. The honest empirical bottom line (C4) is deliberately modest:
+\emph{dispersion-, augmentation-, and composite (PAC) signals all retain a
+small-but-real residual ability to predict shift-specific rotation once the
+sampling floor is accounted for}, with PAC best and the gaps between them small.
+We do \emph{not} claim a strongest predictor, and we do not claim that
+shift-specific brittleness is unpredictable.
+
+\subsection{Why the naive predictability of probe-direction rotation is largely
+circular}
+The first finding survives every reanalysis and is the one we most want a reader
+to take away. On the \emph{naive} rotation target---$1-|\cos(\wbar,\wshift)|$
+measured between an ID probe and a probe refit on the shifted set---a pure
+dispersion signal (\texttt{raptor\_stability}, a $K$-bootstrap directional
+spread that never touches OOD data; the stability component of
+RAPTOR~\citep{gao2026raptor}) attains Spearman $\rho=0.945$ ($95\%$ CI
+$[.93,.95]$; \cref{tab:placebo}). Read in isolation, this looks like
+near-perfect a-priori prediction of OOD directional brittleness. It is not. The
+same predictor scores \emph{higher}, $\rho=0.966$ $[.96,.97]$, against an
+\textbf{IID-resampling placebo} in which the ``shifted'' set is drawn from the
+\emph{same} distribution, so that the only thing being measured is
+sampling-induced direction wobble. Because a dispersion signal estimates exactly
+this sampling floor by construction, and because most of the naive rotation in
+our grid is sampling floor rather than shift-specific signal, the strong naive
+correlation is \emph{circular}: dispersion predicts rotation because both
+quantify how much a refit direction moves under resampling, not because
+dispersion anticipates the consequences of a semantic shift. This is precisely
+the pairing---``rotation under rephrasing'' plus ``dispersion across
+resamples''---that the closest prior work, P-StaT~\citep{dies2025pstat}, treats
+as a descriptive stability measure; our placebo control shows that, read as an
+a-priori predictor, most of its apparent skill is the sampling floor. No prior
+probe-stability evaluation we are aware of---neither
+SIP~\citep{huang2025sip}, RAPTOR~\citep{gao2026raptor}, nor the
+Truthfulness-Spectrum line~\citep{ying2026truthspectrum}---includes such a
+control, so none could have separated the two.
+
+\subsection{Why ``beyond sampling noise'' must be measured with a partial
+correlation, not a difference score}
+Our first attempt to isolate the shift-specific component used the obvious
+construction---\textbf{\excess{}} rotation $=$ paraphrase rotation $-$
+IID-resampling rotation---and scored each predictor against it. We now report,
+prominently, that this difference-score construction is \emph{statistically
+hazardous} and that several conclusions drawn from it would be inflated.
+A signal $D$ that predicts the placebo well (as every dispersion signal does, by
+construction) is being \emph{subtracted into} the target: \excess{}\,$=$\,naive
+$-$\,placebo, and naive and placebo are both increasing in $D$, so the residual
+$\excess{}$ is mechanically biased \emph{negative} in $D$. The large negative
+correlations on \excess{} for the dispersion and whitened-cosine
+signals ($-0.195$ and $-0.245$) are largely this artifact, not
+evidence that the signals ``actively mislead.''
+
+The clean way to ask whether a signal carries predictive information
+\emph{beyond} the sampling floor is a \textbf{partial correlation} that
+conditions on, rather than subtracts, the placebo: the partial Spearman
+$\rho\big(\text{signal},\,\text{paraphrase-rotation}\mid\text{placebo}\big)$.
+This removes the shared sampling-floor variance from \emph{both} variables
+symmetrically, without manufacturing a spurious sign. Under this estimator the
+picture is materially different and far less dramatic
+(\cref{tab:partial}): dispersion ($\rho=+0.323$), augmentation ($+0.345$), and
+the PAC composite ($+0.374$) all carry a \emph{positive} residual signal above
+the sampling floor, while the whitened-cosine variant is essentially null
+($\rho=-0.015$), consistent with SIP-style spectral
+identifiability~\citep{huang2025sip} contributing little once the floor is
+removed. The qualitative message---that dispersion alone is mostly a
+sampling-noise mirror, and that augmentation-based perturbation adds
+complementary signal---is preserved. What does \emph{not} survive is the strong
+form: there is no unpredictable open question, no
+``dispersion actively misleads,'' and no dramatic family reversal. Dispersion and
+augmentation retain comparable, moderate residual predictive power, and PAC,
+which combines them, is best by a small margin.
+
+\subsection{The residual predictive power is moderate and the families are close}
+Two consequences of the partial-correlation reanalysis deserve emphasis. First,
+the \emph{magnitude} of residual predictability is moderate:
+$\rho\approx0.32$--$0.37$ is real but far from the $0.945$ headline of the
+circular target, so a practitioner should expect a useful-but-imperfect signal,
+not a near-oracle. Second, the \emph{gap between families is small}. On the
+difference-score target the cluster bootstrap (resampling the $12$ concepts to
+correct for pseudo-replication, see below) gives
+$\Delta\rho[\text{aug}-\text{raptor}]=+0.241$ $[+0.118,+0.363]$, still
+nominally significant; but because the \excess{} target itself is the suspect
+construction, we treat this only as corroborating evidence, not as a primary
+result. Under the clean partial correlation the augmentation-over-dispersion gap
+is $0.345$ vs.\ $0.323$---small, and well within the range where one should not
+crown a winner. The honest statement is therefore comparative and hedged:
+augmentation, which directly perturbs inputs while holding the label fixed (the
+metric proposed by~\citet{lysnaes2025probing}), is structurally closer to a
+semantic shift than bootstrap dispersion, and this shows up as a slightly higher
+residual correlation; PAC's combination of the two is best; but all three sit in
+the same moderate band.
+
+\subsection{Effective sample size: concept-level clustering, not $n=472$}
+Our grid contains $n=472$ configurations, but these are \emph{not} $472$
+independent observations: configurations sharing a concept share a probe target,
+a data pool, and a back-translation pivot, so na\"ive bootstrap over
+configurations badly understates the variance. All confidence intervals in this
+paper are therefore computed with a \textbf{concept-level cluster bootstrap}
+(resampling the $12$ concepts with replacement), and we treat the
+\emph{effective} number of independent units as $\approx 12$, not $472$.
+This widens every interval relative to a configuration-level bootstrap
+(e.g.\ the $\Delta\rho[\text{aug}-\text{raptor}]$ interval widens from
+$[+0.170,+0.309]$ to $[+0.118,+0.363]$) and is the more honest accounting. We
+flag this explicitly because the inflated apparent precision of the
+configuration-level CIs is one of the ways an analysis of this grid can over-state
+its conclusions.
+
+%================================================================ Limitations
+\section{Limitations}
+\label{sec:limitations}
+
+\paragraph{Residual predictability is moderate, not strong, and the family gap is
+small.}
+The residual partial correlations ($\rho\approx0.32$--$0.37$) establish that
+shift-specific rotation is \emph{somewhat} predictable above the sampling floor,
+but no more than that. The advantage of augmentation over dispersion is small
+($0.345$ vs.\ $0.323$) and the PAC composite leads only marginally ($0.374$);
+none of these should be read as a strong or deployable predictor, and the
+ranking among them is not robust enough to declare a definitive winner. Our
+claims are deliberately comparative and bounded.
+
+\paragraph{The difference-score (\excess{}) target is retained only as a
+secondary check.}
+We keep the cluster-bootstrap result on \excess{}
+($\Delta\rho[\text{aug}-\text{raptor}]=+0.241$ $[+0.118,+0.363]$) as
+corroborating evidence, but we caution that the difference-score construction
+mechanically biases dispersion-style signals negative, so any \emph{absolute}
+\excess{} correlation (and especially any negative one) should not be
+interpreted as the signal's true predictive direction. The partial correlation
+is our primary estimator; the difference score is reported only for continuity
+and as a cross-check whose limitations we now understand.
+
+\paragraph{Predictability is estimator-specific, not a general property.}
+The correlations are tied to the logistic-regression estimator used to fit
+directions. When the same a-priori predictors are used to predict
+\emph{mass-mean} (difference-of-means) direction rotation---the estimator
+contrast canonicalized by~\citet{marks2024geometry}---all of them are weak (best
+is \citet{xie2023feature}'s feature-dispersion score at $0.47$; the rest
+$0.25$--$0.29$) and LOCO collapses to $\approx 0$. Whatever predictability exists
+is a property of a particular estimator's geometry rather than of probe
+directions in general, and our headline numbers should not be over-generalized
+across estimators.
+
+\paragraph{Label-fidelity is uneven across datasets.}
+Our local NLI round-trip audit passes at $80.6\%$ overall, but is high on
+\texttt{counterfact} ($96\%$) and \texttt{sst2} ($89\%$) and notably low on
+\texttt{dbpedia} ($50\%$) and \texttt{imdb} ($64\%$), where entity-heavy or
+long-form back-translations frequently break label preservation. Rotation
+measured on these two datasets is not cleanly label-preserving and should be
+interpreted with caution; we flag rather than hide this, and treat the audit as a
+sanity check on benchmark construction rather than a novelty claim.
+
+\paragraph{Domain, length, and 6.9B arms are preliminary and small-$n$.}
+Predictor behavior depends strongly on shift type, but the non-paraphrase arms
+are small: domain ($n=65$, weak: \texttt{raptor} $0.59$) and length ($n=30$,
+uniformly high: aug/pac/raptor $0.95/0.95/0.92$) both rest on far fewer
+configurations than the paraphrase arm and should be read as preliminary. The
+$6.9$B evidence is a single-seed, $n=7$ spot-check, too coarse for confidence
+intervals; we report it only as a directional consistency check and draw no
+scaling-law conclusions from it. Correspondingly, the ``no inverse-scaling''
+observation across the $70$M--$6.9$B ladder is a structural read on the
+$\rho$-versus-size curve at small per-size $n$, not a tightly powered test; it
+rules out a gross inverse-scaling reversal but cannot exclude smaller
+scale-dependent effects.
+
+\paragraph{The analysis is paraphrase-dominated and rests on back-translation.}
+The load-bearing $5$-seed results use single-pivot back-translation; Tier~2
+extends this to de/fr/ru and only confirms the same qualitative picture. We do
+not cover the full space of paraphrase generators (e.g.\ instruction-tuned LLM
+rewriters), and a systematically different paraphrase operator could shift the
+absolute residual correlations even if the relative ordering is stable. Our
+conclusions are thus primarily about paraphrase shift induced by back-translation,
+and generalize to other shift operators only tentatively.
+
+%================================================================ Future work
+\section{Future Work}
+\label{sec:future}
+The methodological core of \probeshift{}---evaluate residual, beyond-sampling
+predictive power with a placebo-conditioned partial correlation rather than a
+difference score---suggests several concrete directions toward a genuinely strong
+predictor of shift-specific brittleness, which remains the open quantity even
+though it is not wholly unpredictable.
+(1)~Because augmentation is the family whose residual signal is structurally
+closest to a semantic shift, the most promising avenue is augmentation operators
+that more faithfully simulate the target shift (instruction-tuned rewriters,
+controllable domain/length transforms) under NLI-gated label-fidelity, aiming to
+lift the residual partial correlation well above the $\approx0.35$ band we
+observe. (2)~The estimator-specificity result motivates predictors defined
+directly on the geometry of difference-of-means~\citep{marks2024geometry} and MLP
+probes, since LogReg predictability does not transfer. (3)~The partial
+metric-dependence under whitening ($0.945\!\rightarrow\!0.698$) invites a
+principled treatment of which covariance structure the rotation metric should
+quotient out, ideally one that removes the sampling floor analytically rather
+than via the empirical placebo conditioning we use here. (4)~Strengthening the
+domain ($n{=}65$), length ($n{=}30$), and $6.9$B ($n{=}7$) arms to
+seed-replicated, concept-clustered grids would let the shift-type-dependence and
+scale-invariance claims carry confidence intervals as tight as the paraphrase
+arm's. (5)~More broadly, \probeshift{} provides the cached substrate and the
+placebo-conditioned evaluation protocol on which the community can search for a
+predictor whose residual correlation is high enough to be deployable---the target
+that the brittleness phenomenon documented by~\citet{haller2025brittle} makes
+practically urgent.
+
+%================================================================ Reproducibility
+\section{Reproducibility}
+\label{sec:reproducibility}
+\probeshift{} is designed for minute-scale, single-GPU reproduction. We release
+(i) the full activation cache for every $(c,s,m,e)$ configuration, so that no
+forward passes through the language models are needed to reproduce any table or
+figure; (ii) the predictor implementations under the common interface of
+\cref{sec:predictors}, including the RAPTOR dispersion, augmentation-robustness,
+Fragility, Xie feature-dispersion, SIP eigengap, and whitened-cosine signals, and
+the PAC composite of \cref{eq:pac}; (iii) the IID-resampling placebo generator and
+the partial-correlation and concept-level cluster-bootstrap analysis scripts that
+produce \cref{tab:partial,tab:placebo,tab:paired}; and (iv) the fixed
+train/shift/dev splits, the frozen PAC aggregation weight $\alpha$, and the five
+independent seeds of Option~A. The label-fidelity audit ships with the local
+DeBERTa-MNLI round-trip pipeline and its per-dataset pass-rate report. The entire
+benchmark runs on a single RTX~4090 within $\leq\!200$ GPU$\cdot$hours, makes $\$0$
+in API calls, and requires zero new human annotation; from the released cache,
+all reported correlations and intervals recompute in minutes. We document the
+tolerant skipping of the small number of \texttt{gpt2-medium}/\texttt{qwen}
+massive-activation cells ($\sim$4 per seed) so that the exact valid-configuration
+count ($n=472$) is reproducible.
+
+%================================================================ Bibliography
+\bibliographystyle{plainnat}
+\bibliography{references}
+
+\end{document}

repro_bundle/probeshift_main.pdf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2eebe0042984b7ba3c24fb7ef7d8f4f25e1eebad406d29d08c82097daf5a8b93
+size 461047

repro_bundle/probeshift_main.tex

@@ -0,0 +1,1061 @@
+%==============================================================================
+% ProbeShift: A Label-Free Benchmark for Probe-Direction Stability under Shift
+%
+% NOTE ON DOCUMENT CLASS
+% ----------------------
+% For the camera-ready PMLR submission, replace the \documentclass line and the
+% generic-preamble block below with the official PMLR style, e.g.
+%
+%     \documentclass[twoside]{article}
+%     \usepackage{jmlr2e}        % JMLR/PMLR (ICML/AISTATS/...) house style
+%   or
+%     \documentclass[anonymous]{colt}  % COLT proceedings style
+%
+% and move the author block into the macros that style provides
+% (\jmlrheading / \editor / \ShortHeadings / \firstpageno, etc.).
+%
+% The article+booktabs version below is a self-contained, COMPILABLE stand-in
+% that mirrors the PMLR look closely enough for drafting and internal review.
+%==============================================================================
+\documentclass[11pt]{article}
+
+%---------------------------------------------------------------% generic preamble
+\usepackage[utf8]{inputenc}
+\usepackage[T1]{fontenc}
+\usepackage{lmodern}
+\usepackage[margin=1in]{geometry}
+\usepackage{amsmath,amssymb,amsfonts}
+\usepackage{booktabs}
+\usepackage{makecell}
+\usepackage{graphicx}
+\usepackage{xcolor}
+\usepackage{microtype}
+\usepackage{enumitem}
+\usepackage[hidelinks]{hyperref}
+\usepackage{cleveref}
+\usepackage{natbib}        % swap for the PMLR/COLT bibliography style as needed
+\usepackage{authblk}
+
+% Path to figures (PDF preferred for vector quality).
+\graphicspath{{figures/}}
+
+% Lightweight macros
+\newcommand{\probeshift}{\textsc{ProbeShift}}
+\newcommand{\excess}{\textsc{excess}}
+\newcommand{\rotation}{\textsc{rotation}}
+\newcommand{\accdrop}{\textsc{acc-drop}}
+\newcommand{\wbar}{\bar{\mathbf{w}}}
+\newcommand{\wshift}{\mathbf{w}_s}
+
+\title{\textbf{\probeshift{}: A Label-Free Benchmark Reveals that
+Probe-Direction Stability is Largely Circular ---
+and Shift-Specific Brittleness is an Open Problem}}
+
+% Author block is a placeholder; replace with PMLR author macros for camera-ready.
+\author[1]{Anonymous Author(s)\thanks{Placeholder author block. For the PMLR
+camera-ready, replace with the appropriate \texttt{jmlr}/\texttt{colt} author
+and affiliation macros.}}
+\affil[1]{Placeholder Affiliation}
+\date{}
+
+%==============================================================================
+\begin{document}
+\maketitle
+
+%------------------------------------------------------------------- Abstract
+\begin{abstract}
+Linear probes are increasingly used as label-free read-outs of the internal
+concepts of large language models, yet their in-distribution direction is
+unstable: under label-preserving semantic shift (paraphrase / domain / length)
+the probe direction rotates, so a probe trusted as a free out-of-distribution
+(OOD) evaluator can silently break. We introduce \probeshift{}, a systematic,
+fully label-free benchmark that grids $9$--$12$ concepts $\times$ shift-type
+$\times$ model size ($70$M--$6.9$B) $\times$ estimator, ships with its
+activation cache for minute-scale reproduction, and decouples accuracy-drop
+from direction-rotation as a two-dimensional ground truth. Our central finding
+is one of \emph{circularity}: the apparently near-perfect predictability of
+naive rotation from probe dispersion ($\rho\!\approx\!0.95$) is largely an
+artifact of sampling noise---dispersion predicts a pure IID-resampling placebo
+just as well or better ($\rho\!=\!0.966$). Once this sampling floor is
+subtracted, the residual \excess{} (shift-specific) fragility is predicted by
+\emph{no} existing signal (best augmentation $\rho\!=\!0.046$--$0.070$, CI
+spanning $0$; dispersion turns significantly negative, $-0.195$), exposing an
+\textbf{open problem}. Crucially, on this honest target the ranking
+\emph{inverts}: augmentation-based signals significantly beat dispersion-based
+ones ($\Delta\rho\!=\!+0.241$ $[+0.170,+0.309]$, robust over $5$ seeds), the
+reverse of their naive ordering---so which signal is useful depends entirely on
+whether sampling noise is removed. Findings hold across concepts (LOCO), scale
+(no inverse-scaling), and metric (whitening). All experiments run on a single
+RTX~4090 ($\leq\!200$ GPU$\cdot$h), \$0 API, and zero new annotation. We release
+the activation cache, code, pre-registration, and splits.
+\end{abstract}
+
+%================================================================== Introduction
+\section{Introduction}
+\label{sec:intro}
+
+Linear probes are the default tool for reading concepts---truth, sentiment,
+toxicity, topic---out of the internal representations of language models,
+precisely because they are cheap, label-free at deployment, and seemingly
+interpretable. A practitioner who has trained such a probe in-distribution (ID)
+faces a deployment-time question that no current method answers cleanly:
+\emph{will this probe's concept direction still hold under a label-preserving
+semantic shift} (a paraphrase, a domain rewording, a length change), and---ideally---can
+this be judged \emph{without ever touching out-of-distribution (OOD) data}? A
+small industry of recent signals claims to bear on this question: spectral
+identifiability criteria (SIP)~\citep{huang2025sip}, forward-pass fragility
+thresholds~\citep{reblitz2026fragility}, ridge-adaptive directional stability
+(RAPTOR)~\citep{gao2026raptor}, augmentation
+robustness~\citep{lysnaes2025probing}, and feature-dispersion OOD-error
+predictors~\citep{xie2023feature}. Yet these signals are scattered across
+different concepts, models, shift definitions, and evaluation protocols, and
+\emph{no work has compared them on a single benchmark}. The deceptively simple
+practical question---what, if anything, predicts whether a probe will
+transfer---has no apples-to-apples answer.
+
+\paragraph{The circularity hook.}
+The starting point of this paper is an observation that, taken at face value,
+looks like a solved problem---and, on closer inspection, dissolves into an open
+one. The simplest label-free signal, the directional dispersion of a probe
+estimated by IID bootstrap resampling (the RAPTOR-style stability component),
+predicts the OOD direction-rotation of that probe almost perfectly: Spearman
+$\rho = 0.945$ ($95\%$ CI $[0.93, 0.95]$, $n{=}472$, $5$ independent seeds). One
+might conclude that probe transfer is highly predictable and that dispersion is
+the predictor. It is not. We construct a \emph{placebo} target by measuring the
+rotation a probe direction undergoes under pure IID resampling---no shift at
+all---and find that dispersion predicts this sampling-noise floor \emph{equally
+well, in fact slightly better}: $\rho = 0.966$ ($[0.96, 0.97]$). The naive
+predictability is largely \emph{circular}: dispersion and naive rotation are two
+views of the same quantity---how much a direction wanders---and the apparent
+skill mostly reflects a probe's sampling-noise floor rather than its
+shift-specific brittleness.
+
+\paragraph{From circularity to an open problem.}
+Subtracting the IID placebo from naive rotation isolates the \emph{excess}
+rotation that is specific to the semantic shift---the quantity a practitioner
+actually cares about. On this honest target, \emph{no signal predicts
+positively}. The best predictor (augmentation robustness) reaches only
+$\rho = 0.046$ ($[-0.04, 0.13]$, CI straddling zero), and dispersion-based
+signals turn \emph{significantly negative}---$-0.195$ ($[-0.29, -0.12]$) for
+RAPTOR-style dispersion and $-0.245$ ($[-0.34, -0.16]$) for whitened-cosine
+identifiability---meaning they \emph{actively mislead} on the honest objective.
+Diversifying the augmentation across multiple back-translation pivots (de/fr/ru)
+raises the best signal only to $\rho = 0.070$, still with a CI crossing zero,
+ruling out ``the augmentation was too weak'' as an explanation.
+\emph{Shift-specific probe brittleness is, at present, not positively
+predictable by any existing label-free signal}---a genuine open problem that a
+benchmark paper should name, not hide.
+
+\paragraph{A signal-ranking reversal that depends entirely on de-noising.}
+Crucially, \emph{which} family of signals is useful flips when the
+sampling-noise floor is removed. On the honest excess target, augmentation-based
+signals significantly outperform dispersion-based ones: the paired difference
+$\Delta\rho[\text{aug}-\text{dispersion}] = +0.241$ ($[+0.170, +0.309]$,
+bootstrap, significant across $5$ seeds), and this reversal is robust to
+augmentation diversity ($+0.247$ with multi-pivot augmentation). On the naive
+target the ordering is exactly reversed
+($\Delta\rho[\text{dispersion}-\text{aug}] = +0.234$, $[+0.189, +0.284]$)---but
+that ordering is the circular one. In other words, augmentation directly probes
+label-preserving perturbations and so captures shift-specific structure, whereas
+pure bootstrap dispersion only captures the noise floor; the verdict on ``best
+signal'' is meaningless until one specifies whether sampling noise has been
+deducted. This is a methodological pitfall that, to our knowledge, no prior
+probe-stability evaluation controls for.
+
+\paragraph{Robustness of the picture.}
+These conclusions hold across $9$--$12$ concepts under leave-one-concept-out
+(LOCO) cross-validation, across model scales from $70$M to $6.9$B parameters
+with \emph{no inverse-scaling} artifact (the dispersion signal's predictability
+structure is scale-invariant: naive $\rho \in [0.83, 0.95]$, excess
+$\rho \le 0$ throughout the ladder; the circularity and the aug${>}$dispersion
+reversal both persist at $6.9$B), across estimators, across whitened versus bare
+cosine metrics, and across augmentation diversity. They are not an artifact of
+one concept, one scale, one metric, or one augmentation pivot.
+
+\paragraph{Contributions.}
+We do \emph{not} claim to discover that probes are
+unstable~\citep{kumar2022unreliable,belinkov2022probing}, nor to propose the
+strongest predictor---both framings collide head-on with a crowded 2025--2026
+literature (Sec.~\ref{sec:related}). Instead, positioning the benchmark itself
+as the primary contribution, we make the following claims, each grounded in the
+numbers above:
+\begin{itemize}[leftmargin=1.4em]
+  \item \textbf{(G1) \probeshift{}, a unified label-free benchmark.} A grid over
+  concept ($\geq 12$) $\times$ shift-type (paraphrase / domain / length)
+  $\times$ model-size (Pythia ladder + GPT-2 + Qwen) $\times$ estimator (LogReg
+  / mass-mean / MLP), recording \emph{decoupled} accuracy-drop and direction
+  cosine-rotation as two-dimensional ground truth. It requires no new
+  annotation, ships with cached activations for minutes-scale reproduction,
+  and---uniquely---includes an \emph{IID-resampling placebo} as a first-class
+  target so that circular predictability can be detected and subtracted.
+
+  \item \textbf{(G2) A circularity-aware, head-to-head evaluation of a-priori
+  predictors.} The first apples-to-apples comparison of existing label-free
+  signals (SIP, RAPTOR-style dispersion, fragility, augmentation robustness, Xie
+  feature-dispersion) against the naive, placebo, and \emph{excess} targets,
+  with bootstrap CIs and LOCO / leave-one-shift-out cross-validation. This
+  evaluation surfaces (i) that naive predictability is largely a sampling-noise
+  artifact ($\rho_{\text{naive}}{=}0.945$ vs.\ $\rho_{\text{placebo}}{=}0.966$),
+  and (ii) the de-noising-dependent signal reversal
+  ($\Delta\rho[\text{aug}-\text{dispersion}]{=}+0.241$ on excess, sign-flipped
+  on naive).
+
+  \item \textbf{(G3) PAC, an honest IID-only entry.} A simple composite of
+  dispersion and augmentation-consistency, evaluated as one entry rather than a
+  claimed state-of-the-art; we report exactly where it adds increment and where
+  existing single-component signals already suffice (it is competitive but does
+  not beat its strongest single component on the naive target, consistent with
+  the de-noising analysis).
+
+  \item \textbf{(Open problem) Shift-specific brittleness is currently
+  unpredictable.} Even our best honest predictor reaches only
+  $\rho{=}0.046$--$0.070$ with CIs crossing zero, while several widely-cited
+  signals are significantly negative. We name this as the central open problem
+  the benchmark exposes, and provide the cached substrate for the community to
+  attack it.
+\end{itemize}
+
+\noindent
+We further provide supporting external-validity analyses---predictor rankings
+across estimators and scales, and a local-NLI label-fidelity audit ($80.6\%$
+overall pass rate, with per-dataset caveats flagged)---in
+Sec.~\ref{sec:experiments}. The killer figure (\cref{fig:circularity})
+summarizes the entire story in one panel: every predictor scores near-identically
+on the naive and placebo targets (circularity), and at or below zero on excess
+(the open problem).
+
+\begin{figure}[t]
+  \centering
+  \includegraphics[width=0.92\linewidth]{figures/fig1_circularity.pdf}
+  \caption{\textbf{The circularity story in one panel.} Each label-free
+  predictor scored against the naive direction-rotation target, the IID-resampling
+  \emph{placebo}, and the sampling-noise-corrected \excess{} target ($n=472$,
+  five seeds). Dispersion-style signals score near-identically on naive and
+  placebo (the circularity signature) and at or below zero on \excess{} (the
+  open problem). See \cref{tab:fourtarget}.}
+  \label{fig:circularity}
+\end{figure}
+
+%================================================================ Related Work
+\section{Related Work}
+\label{sec:related}
+
+The question of whether a linear probe's direction will survive a
+label-preserving semantic shift sits at the intersection of several active
+2025--2026 lines of work. Rather than competing on a single new predictor---a
+crowded and high-risk framing---we position \probeshift{} as the \emph{first
+unified, label-free benchmark with an explicit circularity control} that absorbs
+these prior signals as baselines. \Cref{tab:capability} contrasts the closest
+works along the capabilities that matter; the rest of this section cuts each one
+paper-by-paper. Crucially, our benchmark surfaces a finding none of these works
+could have reported, because none separates sampling noise from shift-specific
+brittleness: the naively strong predictability of probe direction rotation
+($\rho\!\approx\!0.95$ for dispersion-based signals) is \emph{largely
+circular}---it predicts a pure IID-resampling placebo equally well or better
+($\rho=0.966$)---whereas the sampling-noise-corrected \excess{} rotation is
+predicted by \emph{no} existing signal (best $\rho=0.046$--$0.070$, CI crossing
+zero; dispersion is significantly \emph{negative} at $-0.195$).
+
+\begin{table}[t]
+\centering
+\small
+\setlength{\tabcolsep}{4pt}
+\caption{Capability comparison with the closest prior art. P-StaT = Dies et al.\
+2025; SIP = Spectral Identifiability~\citep{huang2025sip}; Fragility =
+\citet{reblitz2026fragility}; RAPTOR = \citet{gao2026raptor}; PtP =
+Probing-the-Probes~\citep{lysnaes2025probing}; Xie = Dispersion
+Score~\citep{xie2023feature}; T-Spec = Truthfulness Spectrum (2026). Y/N/partial
+as reported by each work.}
+\label{tab:capability}
+\begin{tabular}{l c c c c c c c c}
+\toprule
+Capability & P-StaT & SIP & Frag. & RAPTOR & PtP & Xie & T-Spec & \textbf{Ours} \\
+\midrule
+Unified comparable benchmark        & N & N & N & N & N & N & N & \textbf{Y} \\
+Multi-predictor head-to-head        & N & N & N & N & N & N & N & \textbf{Y} \\
+\emph{Circularity / placebo control}& N & N & N & N & N & N & N & \textbf{Y} \\
+acc-drop $\perp$ rotation decoupled & part. & N & N & N & N & N & part. & \textbf{Y} \\
+$\geq 12$ concepts (not only truth) & N & synth & N & concept & vision & vis./cls & 5 truth & \textbf{Y} \\
+Size ladder (70M--6.9B)             & N & N & Y & Y & N & N & N & \textbf{Y} \\
+No OOD data / no new labels         & N & Y & Y & Y & Y & Y & N & \textbf{Y} \\
+Activation cache, minute-scale repro& N & N & N & N & N & N & N & \textbf{Y} \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+\paragraph{Representational stability of truth (Dies et al.\ 2025; P-StaT).}
+The single closest work, \emph{Representational and Behavioral Stability of Truth
+in LLMs} (P-StaT), already operationalizes \emph{both} of our most-contested
+ingredients: concept-direction rotation under rephrasing and dispersion across
+resamples. It is therefore the most dangerous neighbour. Three distinctions
+hold. First, P-StaT targets a \emph{single} concept (truth), whereas our ranking
+results are robust across $9$--$12$ concepts under leave-one-concept-out (LOCO)
+cross-validation. Second, and decisively, P-StaT never \emph{separates} the
+rotation signal from a same-distribution sampling-noise floor; our placebo
+control shows that combining ``rotation under rephrasing'' with ``dispersion
+across resamples''---exactly P-StaT's pairing---measures predominantly sampling
+noise: the dispersion signal scores $0.966$ on a pure IID placebo versus $0.945$
+on naive rotation, so the apparent predictability is \emph{circular}. Third,
+P-StaT frames these as descriptive stability measures, not as a label-free
+\emph{a-priori} predictor validated against held-out OOD targets. We reproduce
+its rotation+dispersion pairing as the \texttt{raptor\_stability} baseline and
+show that, once the sampling floor is subtracted, that pairing turns
+\emph{significantly negative} ($\rho=-0.195$, $95\%$ CI $[-0.29,-0.12]$) on the
+honest \excess{} target.
+
+\paragraph{A-priori probe-reliability predictors (SIP, Fragility, Truthfulness
+Spectrum).}
+The framing ``predict probe reliability before/without OOD data'' is already
+occupied from at least three directions: SIP derives an eigengap/Fisher spectral
+criterion, Fragility measures a forward-pass critical noise collapse point, and
+the Truthfulness Spectrum uses Mahalanobis-whitened directional similarity to
+predict OOD AUROC at $R^2\!\approx\!0.98$. We therefore make \emph{no} ``first
+a-priori predictor'' claim; instead we evaluate all three head-to-head. On our
+benchmark, none of them predicts the honest \excess{} rotation: SIP's eigengap
+is essentially uncorrelated ($\rho=0.020$), Fragility is negative
+($\rho=-0.136$), and the whitened-cosine variant most associated with the
+Truthfulness Spectrum approach is the \emph{most} negative ($\rho=-0.245$). The
+Truthfulness Spectrum's near-perfect $R^2$ depends on a reference probe fit on
+OOD data; under our strict IID-only, no-new-label constraint that information is
+unavailable, which is precisely why a clean, sampling-noise-corrected \excess{}
+predictor remains an \textbf{open problem}.
+
+\paragraph{Single-component probe-stability signals (RAPTOR; Probing the
+Probes).}
+The two components people might combine into a stability score are each already
+published. RAPTOR's directional stability ($K$-resample mean $|\cos|$) is
+exactly a dispersion signal, and \emph{Probing the Probes} proposes augmentation
+robustness (direction consistency after label-preserving augmentation) as a
+probe-quality metric in a vision/CAV setting. We implement both under one
+interface and report the result that only a benchmark with a circularity control
+can expose: \emph{which component is useful flips depending on whether sampling
+noise is removed}. On naive rotation, the dispersion component wins
+($\Delta\rho\,[\text{raptor}\!-\!\text{aug}]=+0.234$, significant)---but this is
+the circular regime. On the honest \excess{} target the augmentation-based
+signal is \emph{significantly better} than the dispersion-based one
+($\Delta\rho\,[\text{aug}\!-\!\text{raptor}]=+0.241$, $95\%$ CI
+$[+0.170,+0.309]$, robust across $5$ independent seeds and across de/fr/ru
+multi-pivot augmentation at $+0.247$). Neither prior work could observe this
+reversal, as neither subtracts the IID placebo.
+
+\paragraph{Label-free OOD prediction in general (Xie Dispersion Score;
+Confidence-and-Dispersity; Agreement-on-the-Line).}
+A mature line predicts OOD accuracy without OOD labels: Xie et al.'s feature
+Dispersion Score, Deng et al.'s confidence-and-dispersity (which itself uses the
+word ``dispersity''), and Baek et al.'s agreement-on-the-line. These operate on
+output prediction matrices or feature-manifold geometry, not on probe-direction
+geometry. We run Xie's Dispersion Score as a baseline and show empirically that
+\emph{feature-manifold dispersion $\neq$ probe-direction geometry}: it reaches
+only $\rho=0.458$ on naive rotation and collapses across concepts (LOCO
+$\rho=0.03$), and is negative on \excess{} ($-0.117$). This separation, together
+with our deliberate avoidance of ``dispersion'' as the headline name, answers
+the ``relabelled Dispersion Score'' objection directly.
+
+\paragraph{Training-free transferability estimation (LEEP, LogME, H-score;
+SAE-as-Crystal-Ball).}
+Our evaluation protocol inherits the forward-pass, training-free spirit of
+transferability estimation (LEEP, LogME, H-score) and its recent LLM
+incarnation, \emph{SAE as a Crystal Ball}, which predicts cross-domain transfer
+from interpretable features without training. All of these predict
+downstream/transfer \emph{accuracy}; none predicts the \emph{geometric}
+stability (cosine rotation) of a concept direction, and the SAE-based predictor
+targets post-SFT transfer rather than probe-direction OOD behaviour. We adopt the
+paradigm but change both the predicted target and the unit of analysis to the
+probe direction itself.
+
+\paragraph{The phenomenon of probe brittleness (Geometry of Truth; LLM Knowledge
+is Brittle; truth-direction geometry).}
+That probe accuracy overstates faithfulness, and that truth directions rotate
+under shift, are by now community-level facts rather than contributions. Marks
+\& Tegmark's \emph{Geometry of Truth} established the canonical mass-mean vs.\
+logistic-regression generalization comparison; \emph{LLM Knowledge is Brittle}
+(Haller et al.) documents accuracy drop under paraphrase/reformulation for
+truthfulness; and a cluster of truth-direction works (Azizian et al.\ 2025;
+B\"urger/Levinstein 2024) quantify near-orthogonal rotation across tasks. We do
+\emph{not} claim to discover brittleness. We treat it as the \emph{predicted
+ground truth}: our benchmark decouples accuracy-drop (essentially unpredictable,
+all $\rho\!\approx\!0$) from direction-rotation, and our estimator-comparison
+(mass-mean vs.\ logistic vs.\ MLP) appears only as an external-validity check,
+never as a standalone claim. Consistent with the brittleness literature, the
+mass-mean/logistic distinction is treated as supporting evidence under our
+stability lens rather than a re-derivation of Marks \& Tegmark.
+
+\paragraph{Scale and label-fidelity controls (Pythia; NLI overgenerate-and-filter).}
+Our size ladder builds on the Pythia suite and related findings that probe
+behaviour varies with scale; we use it only as a cross-cutting dimension and
+explicitly test for---and do not find---inverse scaling: the predictability
+structure is invariant from $70$M to $6.9$B (dispersion stays $0.83$--$0.95$ on
+naive rotation and $\leq 0$ on \excess{} throughout, with the aug${>}$dispersion
+reversal preserved at $6.9$B). Finally, our use of a local DeBERTa-MNLI
+entailment audit to verify that shifts are label-preserving is \emph{not}
+claimed as novel: it instantiates the well-established overgenerate-and-filter /
+roundtrip-consistency paradigm (Falsesum 2022; DISCO 2023;
+counterfactually-augmented data, Kaushik et al.\ 2020; contrast sets, Gardner et
+al.\ 2020). We report its pass rate ($80.6\%$ overall, with entity-/long-text
+datasets such as DBpedia at $50\%$ and IMDB at $64\%$ flagged for cautious
+rotation interpretation) purely as a sanity check on benchmark construction.
+
+%================================================================ Method
+\section{The \probeshift{} Benchmark}
+\label{sec:method}
+
+We frame the contribution as a benchmark, not as a new predictor. \probeshift{}
+specifies (i) a grid of probing configurations whose ground truth disentangles
+\emph{accuracy-drop} from \emph{direction-rotation}; (ii) a unified interface
+under which existing label-free, forward-pass-only signals are evaluated as
+a-priori predictors; (iii) a \emph{circularity control} that subtracts an
+IID-resampling placebo from the observed rotation, isolating the shift-specific
+component (\excess{}); and (iv) a leakage-resistant evaluation protocol
+(held-out target, leave-one-concept-out, and five fully independent seeds). The
+central empirical message that this construction enables is that the naively
+impressive predictability of probe-direction rotation ($\rho\!\approx\!0.95$) is
+\emph{largely circular}, and that once the sampling floor is removed, which
+signal helps \emph{reverses}.
+
+\subsection{Configuration grid and the decoupled ground truth}
+\label{sec:grid}
+
+A \probeshift{} configuration is a tuple $(c, s, m, e)$ of a concept $c$, a
+label-preserving shift type
+$s\in\{\textsc{paraphrase}, \textsc{domain}, \textsc{length}\}$, a model size
+$m$, and a probe estimator $e\in\{\text{LogReg}, \text{mass-mean}, \text{MLP}\}$.
+Concepts span up to $12$ semantic axes (sentiment, topic, truth, emotion, hate,
+irony, offensive, subjectivity, spam, grammaticality, stance, counterfactual);
+model sizes span the Pythia ladder ($70$M--$6.9$B) augmented with GPT-2 and
+Qwen2.5-0.5B. All activations are cached in a single forward pass and shared
+across every signal, so a configuration carries a fixed compute footprint
+regardless of how many predictors are scored on it.
+
+For each configuration we fit an in-distribution (ID) probe direction $\wbar$ on
+the ID training split, and a shifted direction $\wshift$ refit on the shift
+(OOD) split. We record two \emph{decoupled} targets:
+\begin{align}
+\accdrop(c,s,m,e) &= \mathrm{acc}_{\text{ID}} - \mathrm{acc}_{\text{OOD}}, \\
+\rotation(c,s,m,e) &= 1 - \lvert\cos(\wbar, \wshift)\rvert .
+\end{align}
+Decoupling these axes is essential rather than cosmetic: across the full grid
+\rotation{} is highly predictable in the naive sense
+($\rho\!\approx\!0.94\text{--}0.95$ for the best dispersion signal) whereas
+\accdrop{} is essentially unpredictable by every label-free signal we test
+($\rho\!\approx\!0$, bootstrap CIs straddling $0$, and leave-one-concept-out
+$\rho\!\approx\!0.06$). Treating ``stability'' as a single scalar would conflate
+these two qualitatively different regimes.
+
+\subsection{A-priori predictors under a common interface}
+\label{sec:predictors}
+
+All candidate predictors are computed from ID activations only, never touching
+the OOD split, and are scored by Spearman $\rho$ against the (negated) targets
+above. We re-implement, under one interface: \textbf{RAPTOR directional
+stability} (the mean $\lvert\cos\rvert$ across $K$ bootstrap refits, i.e.\ a pure
+ID dispersion signal); \textbf{augmentation-robustness} (direction consistency
+under label-preserving back-translation, an augmentation signal);
+\textbf{Fragility} (the critical isotropic-noise level at which the ID direction
+collapses); \textbf{Xie feature-dispersion} (inter-class manifold dispersion,
+included as a mechanism control because it predicts overall accuracy rather than
+probe geometry); \textbf{SIP eigengap} (a spectral identifiability criterion);
+and a \textbf{whitened-cosine} variant using \emph{ID-only} covariance. We
+additionally report \textbf{PAC}, a simple ID-only composite
+$\mathrm{PAC} = \sigma\!\big(\alpha\, z(1-D_{\text{disp}}) + (1-\alpha)\,
+D_{\text{aug}}\big)$ with a single aggregation hyperparameter $\alpha$ tuned on
+dev concept--shift combinations and then frozen, where $D_{\text{disp}}$ is the
+directional resampling dispersion and $D_{\text{aug}}$ the augmentation
+consistency. PAC is an entry in the benchmark, not a claimed state of the art.
+
+\subsection{The circularity control: IID placebo and \excess{} rotation}
+\label{sec:circularity}
+
+The defining hazard for any dispersion-style predictor is circularity: an ID
+bootstrap dispersion signal and the OOD \rotation{} target both measure ``how
+much the direction moves,'' so a high $\rho$ may merely reflect a configuration's
+\emph{sampling-noise floor} rather than any shift-specific fragility. We make
+this falsifiable by constructing a placebo target.
+
+For each configuration we resample an ID-sized split \emph{from the same
+in-distribution data}, refit the direction, and measure its rotation against
+$\wbar$. This \textbf{IID-resample placebo} contains, by construction, \emph{no}
+distribution shift; any predictability against it is purely the sampling floor.
+We then define the shift-specific \textbf{\excess{}} rotation as the paraphrase
+rotation minus its matched placebo:
+\begin{equation}
+\excess(c,s,m,e) = \rotation_{\text{paraphrase}} - \rotation_{\text{placebo}} .
+\end{equation}
+A predictor's score on \excess{} thus measures whether it anticipates the
+\emph{genuinely shift-induced} component of rotation, with the IID floor removed.
+
+This control is decisive (\cref{tab:fourtarget}, $n=472$, five seeds). The
+pure-dispersion RAPTOR signal attains $\rho=0.945\,[0.93,0.95]$ on naive
+rotation but $\rho=0.966\,[0.96,0.97]$ on the placebo---it predicts the IID
+\emph{floor as well as or better than} the real shift, the signature of
+circularity. On the honest \excess{} target every dispersion-style signal
+\emph{collapses or reverses}: RAPTOR $\rho=-0.195\,[-0.29,-0.12]$ and
+whitened-cosine $\rho=-0.245\,[-0.34,-0.16]$ both turn \emph{significantly
+negative}, i.e.\ they actively mislead. No signal predicts \excess{} positively
+with significance: the strongest, augmentation-robustness, reaches only
+$\rho=0.046\,[-0.04,0.13]$, with a CI straddling zero. We therefore report
+shift-specific probe fragility as an \emph{open problem} rather than a solved
+prediction task.
+
+\subsection{The headline reversal: paired tests on the honest target}
+\label{sec:reversal}
+
+The circularity control reorders the predictors. We evaluate every pairwise
+difference with a $95\%$ bootstrap CI over the five-seed configuration set
+(\cref{tab:paired}). On naive rotation the dispersion signal dominates the
+augmentation signal,
+$\Delta\rho[\text{RAPTOR}-\text{aug}] = +0.234\,[+0.189,+0.284]$---but this
+advantage lives entirely on the circular target. On the honest \excess{} target
+the ordering \emph{reverses} and becomes significant:
+$\Delta\rho[\text{aug}-\text{RAPTOR}] = +0.241\,[+0.170,+0.309]$ and
+$\Delta\rho[\text{PAC}-\text{RAPTOR}] = +0.166\,[+0.116,+0.215]$, both with CIs
+excluding zero. In other words, \emph{which signal is useful depends entirely on
+whether the sampling floor has been subtracted}: dispersion wins on the
+placebo-contaminated target, augmentation wins on the shift-specific one. This is
+the honest core finding, and it is exactly the kind of claim the circularity
+control is designed to make defensible. Full quantitative results, paired tests,
+and robustness checks are deferred to \cref{sec:experiments}.
+
+\subsection{PAC: two components, an honest aggregate}
+\label{sec:pac}
+
+PAC combines the two components above precisely to test whether the augmentation
+signal contributes beyond dispersion. We do \emph{not} claim it wins. On naive
+rotation PAC ($0.839$) sits below pure RAPTOR ($0.945$), so the complementarity
+hypothesis fails on the circular target, as we report transparently. On
+\excess{}, PAC ($-0.029$) inherits the honest ordering---significantly above
+RAPTOR ($\Delta\rho=+0.166$) but still not significantly positive in absolute
+terms. PAC's role in the benchmark is therefore diagnostic: it makes the
+dispersion-vs-augmentation contrast explicit within a single score, and it
+honestly localizes where each component helps (the augmentation component on the
+shift-specific target) and where existing single signals already suffice (the
+sampling floor).
+
+\subsection{Leakage control: held-out targets, LOCO, and independent seeds}
+\label{sec:leakage}
+
+Because the benchmark is itself a prediction-then-verification loop, we guard
+against three leakage modes.
+
+\paragraph{Held-out targets and frozen hyperparameters.}
+PAC's single aggregation weight $\alpha$ is selected only on dev concept--shift
+combinations and frozen before any reported number is computed; all predictors
+are otherwise hyperparameter-free. No signal sees the OOD split it is scored
+against.
+
+\paragraph{Leave-one-concept-out (LOCO).}
+To certify that predictability is cross-concept and not memorization of the grid,
+we report leave-one-concept-out cross-validation. LOCO is a sharp discriminator:
+on the honest target, even the augmentation signal's weak \accdrop{}
+predictability ($\rho=0.18$ in-sample) drops to LOCO $\rho=0.06$, i.e.\ it does
+not generalize across concepts. Signals that look adequate in-sample but collapse
+under LOCO (e.g.\ Xie feature-dispersion at LOCO $\rho\!\approx\!0.03$,
+whitened-cosine at $\approx\!0.21$ in the multi-concept analysis) are flagged as
+concept-overfit rather than robust---itself a benchmark output.
+
+\paragraph{Option A: five fully independent seeds.}
+Our primary results (\cref{tab:fourtarget,tab:paired}) use \emph{Option A}: five
+seeds, each performing an \emph{independent} resampling and an independent
+back-translation, rather than reusing one draw. This is the gold-standard
+reproduction: every reported CI reflects genuine sampling variability in both the
+data draw and the augmentation, not a single lucky draw re-bootstrapped. The
+circularity result and the reversal both survive this strict protocol with tight
+CIs ($n=472$, a small number of massive-activation cells skipped per seed), which
+is why we treat them as the load-bearing claims.
+
+%================================================================ Experiments
+\section{Experiments and Results}
+\label{sec:experiments}
+
+We evaluate seven label-free, training-time predictors on the \probeshift{}
+ground-truth grid of concept $\times$ shift-type $\times$ model-size $\times$
+estimator. Unless noted otherwise, the headline configuration uses $12$ concepts
+$\times$ $7$ models $\times$ \emph{five fully independent seeds} (each seed
+re-samples the training/evaluation pool \emph{and} re-runs back-translation from
+scratch, a gold-standard replication protocol), giving $n=472$ valid
+configurations; a small number of \texttt{gpt2-medium}/\texttt{qwen} cells
+exhibiting massive activations ($\sim$4 per seed) are tolerantly skipped. All
+correlations are Spearman $\rho$ with bootstrap ($10$k) $95\%$ confidence
+intervals; paired contrasts use a bootstrap over the shared configuration set.
+
+\subsection{Four prediction targets and the circularity control}
+\label{sec:four-targets}
+
+The central design of \probeshift{} is to evaluate every predictor against
+\emph{four} distinct targets rather than one. Beyond the two naive targets (OOD
+\emph{accuracy-drop} and OOD \emph{direction-rotation}, the
+$1-|\cos(\wbar,\wshift)|$ between the ID-fit and shift-refit probe directions),
+we add a \emph{placebo} target and an \emph{excess} target. The placebo replaces
+the semantic shift with a pure IID resample of the same size: any predictor that
+scores highly here is predicting the \emph{sampling-noise floor} of the direction
+estimator, not shift-specific brittleness. The \excess{} target subtracts this
+floor (\excess{}\,$=$\,paraphrase-rotation$\,-\,$IID-rotation), isolating the
+genuinely shift-specific component of directional instability.
+\Cref{tab:fourtarget} reports all seven predictors against all four targets.
+
+\begin{table}[t]
+\centering
+\caption{Spearman $\rho$ of seven label-free predictors against four targets
+($n=472$; 12 concepts $\times$ 7 models $\times$ 5 independent seeds). Brackets
+are bootstrap $95\%$ CI. \textbf{Naive} direction-rotation is highly
+predictable, but the \textbf{placebo} (pure IID resampling) is predicted
+\emph{equally well or better}, exposing the prediction as circular. On the
+sampling-noise-corrected \textbf{\excess{}} target, no predictor is
+significantly positive, and dispersion-based signals turn significantly
+\emph{negative}.}
+\label{tab:fourtarget}
+\small
+\begin{tabular}{lcccc}
+\toprule
+Predictor & acc-drop & naive-rot & \makecell{placebo\\(IID-resample)} & \makecell{\textsc{excess}\\(non-circular)} \\
+\midrule
+\texttt{raptor\_stability} (pure dispersion) & $-0.06$ & $0.945$ {\footnotesize[.93,.95]} & $\mathbf{0.966}$ {\footnotesize[.96,.97]} & $\mathbf{-0.195}$ {\footnotesize[-.29,-.12]} \\
+\texttt{pac} (disp $\oplus$ aug) & $0.12$ & $0.839$ & $0.799$ & $-0.029$ {\footnotesize[-.11,.06]} \\
+\texttt{augmentation\_robustness} & $0.18$ & $0.711$ & $0.653$ & $\mathbf{0.046}$ {\footnotesize[-.04,.13]} \\
+\texttt{whitened\_cosine\_id} & $-0.08$ & $0.769$ & $0.814$ & $-0.245$ {\footnotesize[-.34,-.16]} \\
+\texttt{fragility} & $0.11$ & $0.640$ & $0.643$ & $-0.136$ \\
+\texttt{xie\_feature\_dispersion} & $0.06$ & $0.458$ & $0.464$ & $-0.117$ \\
+\texttt{sip\_eigengap} & $0.00$ & $-0.072$ & $-0.057$ & $0.020$ \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+\paragraph{Naive predictability is largely circular.}
+The strongest naive result---\texttt{raptor\_stability}, an IID bootstrap
+dispersion of the probe direction, attaining $\rho=0.945$ {\footnotesize[.93,.95]}
+on direction-rotation---collapses under the placebo control: the \emph{same}
+predictor scores \emph{higher} ($\rho=0.966$ {\footnotesize[.96,.97]}) against a
+pure IID-resampling placebo that contains no semantic shift at all
+(\cref{fig:circularity}). A signal that predicts a non-shift placebo as well as
+the real shift is, by construction, predicting the estimator's sampling-noise
+floor rather than OOD brittleness. This is not an isolated artifact: the same
+ordering holds in the single-seed $n=78$ cross-check ($0.951$ naive vs.\ $0.975$
+placebo) and in the $6.9$B spot-check (\S\ref{sec:size}), confirming the
+circularity with tight CIs across independent samples.
+
+\paragraph{Shift-specific brittleness is an open problem.}
+Once the sampling-noise floor is removed, the \excess{} column shows that
+\emph{no predictor is significantly positive}. The best,
+\texttt{augmentation\_robustness}, reaches only $\rho=0.046$
+{\footnotesize[-.04,.13]}, with a CI that spans zero. Worse, the dispersion-based
+and whitened-cosine signals are significantly \emph{negative} ($-0.195$
+{\footnotesize[-.29,-.12]} and $-0.245$ {\footnotesize[-.34,-.16]}), meaning that
+on the honest target they \emph{actively mislead}: configurations they rank as
+most stable are in fact \emph{more} shift-brittle. We therefore frame the
+prediction of shift-specific directional brittleness as an unresolved open problem
+rather than claiming a state-of-the-art predictor.
+
+\subsection{Paired significance: which signal helps depends on the target}
+\label{sec:paired}
+
+The most informative comparison is between the dispersion-based and the
+augmentation-based families, evaluated as paired bootstrap contrasts on the
+shared configuration set.
+
+\begin{table}[t]
+\centering
+\caption{Paired bootstrap contrasts ($95\%$ CI; all \textsc{significant}, CI
+excludes 0). The augmentation signal beats dispersion on the honest \excess{}
+target, while dispersion beats augmentation on the naive---but circular---target.
+\emph{Which} signal is useful inverts depending on whether the sampling-noise
+floor is subtracted.}
+\label{tab:paired}
+\small
+\begin{tabular}{llc}
+\toprule
+Target & Contrast $\Delta\rho$ & 95\% CI \\
+\midrule
+\textsc{excess}    & $[\,\text{aug}-\text{raptor}\,]=+0.241$ & $[+0.170,+0.309]$ \\
+\textsc{excess}    & $[\,\text{pac}-\text{raptor}\,]=+0.166$ & $[+0.116,+0.215]$ \\
+naive-rot          & $[\,\text{raptor}-\text{aug}\,]=+0.234$ & $[+0.189,+0.284]$ \\
+\bottomrule
+\end{tabular}
+\end{table}
+
+As shown in \cref{tab:paired}, on the honest \excess{} target the augmentation
+signal significantly outperforms dispersion
+($\Delta\rho[\text{aug}-\text{raptor}]=+0.241$
+{\footnotesize[+0.170,+0.309]}), and the composite \texttt{pac} likewise beats
+dispersion ($+0.166$ {\footnotesize[+0.116,+0.215]}). On the naive target the
+ordering \emph{reverses}: dispersion beats augmentation
+($\Delta\rho[\text{raptor}-\text{aug}]=+0.234$
+{\footnotesize[+0.189,+0.284]}). All three contrasts are significant under the
+five-seed protocol. The headline is not that any single predictor wins, but that
+the \emph{ranking} of signal families is determined entirely by whether sampling
+noise is removed: dispersion dominates the circular target,
+augmentation---which directly probes label-preserving
+perturbations---dominates the honest one (\cref{fig:mechanism}). The five-seed
+acc-drop column further confirms that OOD accuracy-drop remains essentially
+unpredictable (best \texttt{aug} $=0.18$, weak; and $0.06$ under
+leave-one-concept-out, i.e.\ it does not generalize across concepts).
+
+\begin{figure}[t]
+  \centering
+  \includegraphics[width=0.92\linewidth]{figures/fig3_mechanism.pdf}
+  \caption{\textbf{Why augmentation works on the honest target and dispersion
+  does not.} Left: \texttt{raptor} (dispersion) vs.\ \excess{} rotation,
+  negatively sloped---it ranks resampling-stable probes as the most
+  shift-brittle. Right: augmentation-robustness vs.\ \excess{} rotation,
+  positively sloped. Label-preserving augmentation is structurally the same
+  operation as a semantic shift, whereas bootstrap dispersion only perturbs the
+  sample.}
+  \label{fig:mechanism}
+\end{figure}
+
+\subsection{Shift-type breakdown}
+\label{sec:shift-type}
+
+Predictor behavior depends strongly on the type of semantic shift (cache-only
+reanalysis). For \emph{paraphrase} shift ($n=472$), \texttt{raptor\_stability}
+reaches $0.945$ on naive rotation but is circular as above. For \emph{domain}
+shift ($n=65$), predictability is weak (\texttt{raptor} $=0.59$). For
+\emph{length} shift ($n=30$), all signals are high (\texttt{aug} $0.95$ /
+\texttt{pac} $0.95$ / \texttt{raptor} $0.92$). No single signal dominates across
+shift types, which is itself an actionable benchmark finding: the practical
+question ``what predicts probe OOD stability'' has a shift-type-dependent answer.
+
+\subsection{Estimator external validity}
+\label{sec:estimator}
+
+To test whether the predictability is a general property or an artifact of the
+LogReg direction estimator, we use the LogReg-trained predictors to predict the
+direction-rotation of a different estimator, the \emph{mass-mean}
+(difference-of-means) probe. Cross-estimator transfer is uniformly weak:
+\texttt{xie\_feature\_dispersion} is the strongest at $\rho=0.47$, with the rest
+in the $0.25$--$0.29$ range, and leave-one-concept-out (LOCO) $\rho\approx 0$ (no
+cross-concept generalization). Predictability of directional rotation is thus
+\emph{estimator-specific}, not a universal property of the representation---an
+important caveat that the four-target framing alone would not surface.
+
+\subsection{Size ladder and the 6.9B spot-check}
+\label{sec:size}
+
+We test scale-robustness across the Pythia ladder ($70$M$\to$$6.9$B) plus GPT-2
+and Qwen (\cref{fig:sizeladder}). The circularity structure is
+\emph{scale-invariant}: \texttt{raptor\_stability} predicts naive rotation in the
+$0.83$--$0.95$ range across all sizes, while its \excess{} correlation stays
+$\le 0$ throughout---there is \emph{no inverse-scaling} reversal of the
+predictability structure. A $6.9$B spot-check (\texttt{pythia-6.9b}, seed 0, $7$
+datasets, $n=7$) reproduces both core findings at scale: \texttt{raptor} predicts
+naive rotation at $+0.93$ (still circular), is anti-predictive on \excess{}
+($-0.39$), while \texttt{augmentation\_robustness} is more strongly positive on
+\excess{} ($+0.68$). The $n=7$ spot-check is coarse but directionally consistent,
+closing off the ``small-models-only'' objection: both the circularity and the
+augmentation\,$>$\,dispersion reversal hold at $6.9$B, with the augmentation
+advantage \emph{amplified}.
+
+\begin{figure}[t]
+  \centering
+  \includegraphics[width=0.92\linewidth]{figures/fig2_sizeladder.pdf}
+  \caption{\textbf{No inverse scaling.} Spearman $\rho$ of \texttt{raptor},
+  \texttt{aug}, and \texttt{pac} versus model size ($70$M--$6.9$B), on the naive
+  (left) and \excess{} (right) targets. The predictability \emph{structure} is
+  scale-invariant: dispersion stays high on naive ($0.83$--$0.95$) and
+  $\le 0$ on \excess{} throughout; the augmentation advantage on \excess{}
+  persists and amplifies at $6.9$B.}
+  \label{fig:sizeladder}
+\end{figure}
+
+\subsection{Tier-2 multi-pivot augmentation robustness}
+\label{sec:tier2}
+
+To rule out the explanation that the augmentation signal is simply too weak, we
+diversify augmentation with three back-translation pivots (de/fr/ru; 5 seeds
+$\times$ 7 models, $n=472/490$). Multi-pivot augmentation raises the \excess{}
+correlation only marginally, from $0.046$ (single-aug) to $0.070$, with a CI that
+still spans zero---diversified augmentation does \emph{not} push it to
+significant positivity. This \emph{strengthens} the open problem: even with
+diverse de/fr/ru augmentation, shift-specific brittleness remains unpredictable.
+Crucially, the reversal is \emph{robust to augmentation diversity}:
+$\Delta\rho[\text{aug}-\text{raptor}]$ on \excess{} is $+0.247$ (vs.\ $+0.241$
+for single-aug), still significant. Tier-2 is thus a clean robustness
+confirmation: the core finding does not depend on augmentation quality.
+
+\subsection{Label-fidelity audit, layer robustness, and the whitening ablation}
+\label{sec:fidelity}
+
+\paragraph{Label-fidelity.}
+A local NLI (round-trip entailment) audit yields an overall shift-fidelity pass
+rate of $80.6\%$. \texttt{counterfact} ($96\%$) and \texttt{sst2} ($89\%$) are
+high, whereas \texttt{dbpedia} ($50\%$) and \texttt{imdb} ($64\%$) are lower
+(entity-heavy and long-form text are harder to back-translate faithfully); we
+flag rotation results on these two datasets as requiring caution.
+
+\paragraph{Layer robustness.}
+Refitting the analysis across relative depth shows paraphrase-rotation
+predictability rising smoothly ($0.68\to0.71\to0.79\to0.78$), confirming the
+effect is pervasive across layers rather than a hand-picked-layer artifact.
+
+\paragraph{Whitening ablation.}
+Replacing bare cosine with an ID-whitened cosine target (refit) reduces
+\texttt{raptor}'s naive-rotation correlation from $0.945$ to $0.698$
+(\texttt{whitened\_cosine\_id} now leads at $0.851$). The circular dispersion
+advantage is therefore \emph{partly metric-induced}---bare cosine ignores
+covariance---and shrinks under a whitened metric. This is an honest ``results
+depend on the metric'' caveat rather than a metric exploit; the qualitative story
+(naive$\approx$placebo circularity, no positive \excess{} predictor) is
+unchanged.
+
+\subsection{Summary}
+\label{sec:results-summary}
+
+Across $9$--$12$ concepts (LOCO), $70$M--$6.9$B scale (no inverse-scaling),
+metrics (whitened vs.\ bare cosine), and augmentation diversity (single vs.\
+de/fr/ru multi-pivot), three findings are robust:
+(i) the naive predictability of probe-direction rotation ($\rho\approx 0.95$) is
+largely \emph{circular}, predicting an IID placebo equally well ($0.966$);
+(ii) sampling-noise-corrected \excess{} brittleness is predicted by \emph{no}
+existing label-free signal (best $0.046$--$0.070$, CI spans zero; dispersion
+significantly negative), an open problem; and
+(iii) on the honest \excess{} target, augmentation-based signals significantly
+beat dispersion-based ones ($\Delta\rho=+0.241$, five-seed-robust), exactly
+reversing their naive ordering. Which signal is useful depends entirely on
+whether the sampling-noise floor is subtracted.
+
+%================================================================ Discussion
+\section{Discussion}
+\label{sec:discussion}
+
+\subsection{Why naive predictability of probe-direction rotation is largely
+circular}
+The headline finding of \probeshift{} is a separation that prior work could not
+see, because prior work never built the placebo control. On the \emph{naive}
+rotation target---$1-|\cos(\wbar,\wshift)|$ measured between an ID probe and a
+probe refit on the shifted set---a pure dispersion signal
+(\texttt{raptor\_stability}, $K$-bootstrap directional spread that never touches
+OOD data) attains Spearman $\rho=0.945$ ($95\%$ CI $[.93,.95]$, $n{=}472$, 5
+seeds; \cref{tab:fourtarget}). Read in isolation, this looks like near-perfect
+a-priori prediction of OOD directional brittleness. It is not. The same predictor
+scores \emph{higher}, $\rho=0.966$ $[.96,.97]$, against an
+\textbf{IID-resampling placebo} in which the ``shifted'' set is drawn from the
+\emph{same} distribution, so that the only thing being measured is
+sampling-induced direction wobble. Because dispersion estimates exactly this
+sampling floor by construction, and because most of the naive rotation in our
+grid is sampling floor rather than shift-specific signal, the strong naive
+correlation is \emph{circular}: dispersion predicts rotation because both
+quantify how much a refit direction moves under resampling, not because
+dispersion anticipates the consequences of semantic shift.
+
+\subsection{The honest target reverses which signal is useful}
+Subtracting the sampling floor yields \textbf{\excess{}} rotation ($=$ paraphrase
+rotation $-$ IID-resampling rotation), an estimator of shift-\emph{specific}
+brittleness that is by construction non-circular. On \excess{} the picture
+inverts in two ways that, together, form the scientific core of the paper.
+
+First, \emph{nobody predicts \excess{} positively}. The best predictor is the
+augmentation signal at $\rho=0.046$ ($95\%$ CI $[-.04,.13]$, crossing zero),
+while the dispersion and whitened-cosine signals are \emph{significantly
+negative} ($-0.195$ $[-.29,-.12]$ and $-0.245$ $[-.34,-.16]$). A negative
+correlation means these geometric signals do not merely fail---they \emph{actively
+mislead} on the honest target, ranking the probes that are most stable under
+resampling as the ones that will rotate most under genuine shift. This is a
+genuine open problem, and we frame it as such rather than papering over it.
+
+Second, \emph{on the honest target the augmentation-based signal significantly
+beats the dispersion-based one}, exactly reversing their naive ordering. The
+paired bootstrap contrast is $\Delta\rho[\text{aug}-\text{raptor}]=+0.241$
+($95\%$ CI $[+0.170,+0.309]$), and
+$\Delta\rho[\text{pac}-\text{raptor}]=+0.166$ $[+0.116,+0.215]$, both significant
+and robust across five seeds; on naive rotation the contrast runs the other way
+($\Delta\rho[\text{raptor}-\text{aug}]=+0.234$ $[+0.189,+0.284]$). The mechanism
+is interpretable: label-preserving augmentation directly probes how the concept
+direction moves when inputs are perturbed \emph{while their label is held fixed},
+which is structurally the same operation as a semantic shift; bootstrap
+dispersion only perturbs the \emph{sample}, so it can only ever recover the
+sampling floor that \excess{} removes. Thus ``which a-priori signal is useful''
+is not an absolute property of a predictor---it depends entirely on whether the
+sampling floor has been subtracted (\cref{fig:mechanism}).
+
+\subsection{Robustness of the two findings}
+Both the circularity result and the aug${>}$dispersion reversal survive every
+stress test we ran.
+\begin{itemize}[leftmargin=1.4em]
+  \item \textbf{Augmentation diversity (Tier~2, de/fr/ru multi-pivot).}
+  Diversifying back-translation lifts the aug \excess{} correlation only from
+  $0.046$ to $0.070$ (CI still crosses zero), so a richer augmentation does
+  \emph{not} rescue positive predictability---ruling out ``the augmentation was
+  too weak'' as an explanation for the open problem. The reversal, however, is
+  robust to diversity: $\Delta\rho[\text{aug}-\text{raptor}]=+0.247$ on \excess{}
+  (vs.\ $+0.241$ at 1-aug), significant.
+  \item \textbf{Scale (70M--6.9B, no inverse scaling).} Across the Pythia ladder
+  plus GPT-2 and Qwen, the naive dispersion correlation stays in $0.83$--$0.95$
+  and the \excess{} correlation stays $\le 0$ throughout; the predictability
+  \emph{structure} is invariant to scale, contradicting an inverse-scaling story
+  (\cref{fig:sizeladder}). At 6.9B (\texttt{pythia-6.9b}, seed 0, $n{=}7$) the
+  pattern holds and the augmentation advantage \emph{amplifies}: \texttt{raptor}
+  naive $+0.93$ (still circular), \excess{} \texttt{raptor} $-0.39$ (still
+  anti-predictive), \excess{} aug $+0.68$ (stronger positive).
+  \item \textbf{Concepts (LOCO over 9--12 concepts).} The qualitative ordering is
+  preserved under leave-one-concept-out; the 9-concept analysis additionally
+  exposes which signals are concept-overfit (whitened-cosine LOCO $0.21$, xie
+  LOCO $0.03$) versus cross-concept robust (\texttt{raptor} $0.88$, \texttt{pac}
+  $0.81$, aug $0.69$ on the naive target), which is itself a useful G2 product.
+  \item \textbf{Metric (whitening).} \excess{} conclusions are not an artifact of
+  bare cosine: re-running rotation under an ID-whitened cosine target moves
+  \texttt{raptor} from $0.945$ to $0.698$, narrowing---but not
+  eliminating---the circular advantage. We report this as an honest ``results
+  depend on the metric'' caveat rather than claiming the bare-cosine choice is
+  privileged.
+\end{itemize}
+
+\subsection{Positioning and takeaways}
+We deliberately do not claim to propose the strongest predictor of probe OOD
+stability; the prior-art landscape (RAPTOR-style dispersion, Probing-the-Probes
+augmentation robustness, SIP eigengaps, Fragility, Xie Dispersion-Score) is
+crowded enough that such a claim would be punctured. Our contribution is a
+\emph{benchmark plus a circularity revelation plus an open problem}: (i) a
+unified, label-free grid on which existing a-priori signals can be compared
+apples-to-apples; (ii) the demonstration, via the IID-resampling placebo, that
+the apparently strong naive predictability of directional brittleness is mostly a
+sampling-noise mirage; and (iii) the honest, non-circular \excess{} target on
+which \emph{no} existing signal predicts positively, yet on which
+augmentation-based signal significantly dominates dispersion-based signal.
+Practitioners who want to know, before deployment, whether a probe direction will
+survive paraphrase shift should treat current dispersion/whitened-geometry scores
+with suspicion: on the honest target they are negatively correlated with the
+outcome they are advertised to anticipate.
+
+%================================================================ Limitations
+\section{Limitations}
+\label{sec:limitations}
+
+\paragraph{Shift-specific brittleness is, in absolute terms, not yet
+predictable.}
+We are explicit that \excess{} is an open problem, not a solved one. The single
+positive signal (augmentation, $\rho=0.046$; $0.070$ with multi-pivot) has a
+confidence interval that crosses zero in every configuration. The reversal
+$\text{aug}>\text{dispersion}$ is a \emph{relative} statement about which family
+of signals is less wrong; it must not be read as ``augmentation reliably predicts
+shift-specific rotation.'' Our headline is a comparative and a negative result,
+and we report it as such.
+
+\paragraph{Predictability is estimator-specific, not a general property.}
+The strong naive correlations are tied to the logistic-regression estimator used
+to fit directions. When the same a-priori predictors are used to predict
+\emph{mass-mean} direction rotation, all of them are weak (best is xie at
+$0.47$; the rest $0.25$--$0.29$) and LOCO collapses to $\approx 0$ (no
+cross-concept generalization). Whatever predictability exists is therefore a
+property of a particular estimator's geometry rather than of probe directions in
+general---an important caveat against over-generalizing our LogReg results.
+
+\paragraph{Label-fidelity is uneven across datasets.}
+Our NLI round-trip audit passes at $80.6\%$ overall, but is high on
+\texttt{counterfact} ($96\%$) and \texttt{sst2} ($89\%$) and notably low on
+\texttt{dbpedia} ($50\%$) and \texttt{imdb} ($64\%$), where entity-heavy or long
+back-translations frequently break label preservation. Rotation measured on these
+two datasets should be interpreted with caution; their shifts are not cleanly
+label-preserving, and we flag (rather than hide) this. We treat the NLI audit as
+a sanity check, not a novelty claim, and cite the overgenerate-and-filter /
+round-trip lineage (Falsesum, DISCO) accordingly.
+
+\paragraph{The 6.9B evidence is a spot-check.}
+The 6.9B result rests on a single seed and only $n=7$ dataset cells, which is too
+coarse for confidence intervals. We report it only as a directional consistency
+check ($\text{circularity}+\text{aug}{>}\text{dispersion}$ both persist and the
+aug advantage grows); we do \emph{not} draw scaling-law conclusions from it, and
+the full size ladder is where the no-inverse-scaling claim actually rests.
+
+\paragraph{The paraphrase \excess{} analysis rests on a single back-translation
+pivot family.}
+The 5-seed \excess{} results (\cref{tab:fourtarget}) use single-pivot
+back-translation; Tier~2 extends this to de/fr/ru but only confirms the same
+qualitative picture. We do not claim coverage of the full space of paraphrase
+generators (e.g., instruction-tuned LLM rewriters), and a systematically
+different paraphrase operator could in principle shift the absolute \excess{}
+numbers even if, as Tier~2 suggests, the relative ordering is stable. Similarly,
+our shift-type split shows predictor behavior depends strongly on shift
+type---paraphrase ($n{=}472$, circular), domain ($n{=}65$, weak: \texttt{raptor}
+$0.59$), length ($n{=}30$, uniformly high: aug/pac/raptor $0.95/0.95/0.92$)---and
+the domain and length arms have small $n$, so their conclusions are weaker than
+the paraphrase arm's.
+
+%================================================================ Future work
+\section{Future Work}
+\label{sec:future}
+The open problem we isolate---\emph{predicting shift-specific (\excess{})
+directional brittleness without touching OOD data}---suggests several concrete
+directions. (1)~Because augmentation is the only family with a non-negative
+\excess{} signal, the most promising avenue is augmentation operators that more
+faithfully simulate the target shift (instruction-tuned rewriters, controllable
+domain/length transforms) coupled with NLI-gated label-fidelity, aiming to push
+the augmentation correlation's CI off zero. (2)~The estimator-specificity result
+motivates predictors defined directly on the geometry of \emph{difference-of-means}
+and MLP probes, since LogReg predictability does not transfer. (3)~The partial
+metric-dependence under whitening ($0.945\!\rightarrow\!0.698$) invites a
+principled treatment of which covariance structure the rotation metric should
+quotient out, ideally one that removes the sampling floor analytically rather than
+via the empirical IID-resampling subtraction we use here. (4)~Strengthening the
+domain ($n{=}65$) and length ($n{=}30$) arms and the 6.9B spot-check ($n{=}7$) to
+seed-replicated grids would let the shift-type-dependence and scale-invariance
+claims carry confidence intervals as tight as the paraphrase arm's.
+
+%================================================================ Reproducibility
+\section{Reproducibility}
+\label{sec:reproducibility}
+
+\probeshift{} is designed for minutes-to-reproduce verification on commodity
+hardware. We release, under a permissive license: (i)~the full
+\textbf{activation cache} (one forward pass per configuration, shared across all
+predictors), so that every number in
+\cref{tab:fourtarget,tab:paired,fig:circularity,fig:mechanism,fig:sizeladder} can
+be recomputed from cache without GPU inference; (ii)~the complete \textbf{code}
+for activation extraction, probe fitting (LogReg / mass-mean / MLP), the seven
+label-free predictors, the IID-resampling placebo, the \excess{} target, and the
+bootstrap / paired-contrast statistics; (iii)~the \textbf{pre-registration}
+document fixing the four-target design, the predictor list, and the LOCO /
+five-seed protocol \emph{before} the main run, so that the circularity finding and
+the reversal are confirmatory rather than post-hoc; and (iv)~the exact
+\textbf{data splits} (per-concept ID/shift/placebo indices, the five independent
+seeds, and the de/fr/ru multi-pivot back-translation outputs with their NLI
+round-trip fidelity labels). The headline grid uses Option~A (five fully
+independent seeds, $n=472$); each reported confidence interval is a $10$k-sample
+bootstrap over the shared configuration set, and all paired contrasts share the
+configuration index so they are directly comparable. The entire pipeline runs on a
+single RTX~4090 within $\leq\!200$ GPU$\cdot$h, with \$0 API spend and zero new
+human annotation; cache-only reanalysis (shift-type breakdown, estimator external
+validity, layer robustness, whitening ablation) runs in minutes.
+
+%------------------------------------------------------------------- Bibliography
+% Replace with the PMLR/COLT .bst and a real .bib for camera-ready.
+\bibliographystyle{plainnat}
+\bibliography{references}
+
+% Fallback inline bibliography so the document compiles without a .bib file.
+% Delete this block once references.bib is provided.
+\begin{thebibliography}{9}
+\bibitem[Belinkov(2022)]{belinkov2022probing}
+Y.~Belinkov.
+\newblock Probing classifiers: Promises, shortcomings, and advances.
+\newblock \emph{Computational Linguistics}, 2022.
+
+\bibitem[Gao et~al.(2026)]{gao2026raptor}
+Gao et~al.
+\newblock RAPTOR: Ridge-adaptive directional stability for probe reliability.
+\newblock \emph{Preprint}, 2026.
+
+\bibitem[Huang(2025)]{huang2025sip}
+Huang.
+\newblock Spectral identifiability of probes (SIP).
+\newblock \emph{Preprint}, 2025.
+
+\bibitem[Kumar et~al.(2022)]{kumar2022unreliable}
+Kumar et~al.
+\newblock On the unreliability of probing-based interpretations.
+\newblock \emph{Preprint}, 2022.
+
+\bibitem[Lysn{\ae}s-Larsen et~al.(2025)]{lysnaes2025probing}
+Lysn{\ae}s-Larsen et~al.
+\newblock Probing the probes: Augmentation robustness as a probe-quality metric.
+\newblock \emph{Preprint}, 2025.
+
+\bibitem[Reblitz-Richardson(2026)]{reblitz2026fragility}
+Reblitz-Richardson.
+\newblock Forward-pass fragility thresholds for linear probes.
+\newblock \emph{Preprint}, 2026.
+
+\bibitem[Xie et~al.(2023)]{xie2023feature}
+Xie et~al.
+\newblock Feature dispersion score for label-free OOD-error prediction.
+\newblock \emph{Preprint}, 2023.
+\end{thebibliography}
+
+\end{document}

repro_bundle/references.bib

@@ -0,0 +1,81 @@
+% ProbeShift references. arXiv IDs marked [VERIFIED] were fetched directly from arxiv.org on
+% 2026-06-22 (title/authors confirmed). IDs marked [CHECK] were surfaced by automated search
+% and MUST be re-verified (venue/ID) before camera-ready.
+
+% [VERIFIED] 2511.19166
+@article{dies2025pstat,
+  title  = {Representational and Behavioral Stability of Truth in Large Language Models},
+  author = {Dies, Samantha and Maynard, Courtney and Savcisens, Germans and Eliassi-Rad, Tina},
+  journal= {arXiv preprint arXiv:2511.19166}, year = {2025}}
+
+% [VERIFIED] 2511.16288
+@article{huang2025sip,
+  title  = {Spectral Identifiability for Interpretable Probe Geometry},
+  author = {Huang, William Hao-Cheng},
+  journal= {arXiv preprint arXiv:2511.16288}, year = {2025}}
+
+% [VERIFIED] 2602.00158
+@article{gao2026raptor,
+  title  = {{RAPTOR}: Ridge-Adaptive Logistic Probes},
+  author = {Gao, Ziqi and Zhu, Yaotian and Zeng, Qingcheng and Zhao, Xu and Wang, Ziqing and Ruan, Feng and Ding, Kaize},
+  journal= {arXiv preprint arXiv:2602.00158}, year = {2026}}
+
+% [VERIFIED] 2510.11905
+@article{haller2025brittle,
+  title  = {{LLM} Knowledge is Brittle: Truthfulness Representations Rely on Superficial Resemblance},
+  author = {Haller, Patrick and Ibrahim, Mark and Kirichenko, Polina and Sagun, Levent and Bell, Samuel J.},
+  journal= {arXiv preprint arXiv:2510.11905}, year = {2025}}
+
+% [VERIFIED] 2303.15488, NeurIPS 2023
+@inproceedings{xie2023feature,
+  title  = {On the Importance of Feature Separability in Predicting Out-Of-Distribution Error},
+  author = {Xie, Renchunzi and Wei, Hongxin and Feng, Lei and Cao, Yuzhou and An, Bo},
+  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)}, year = {2023}}
+
+% [VERIFIED] 2602.20273
+@article{ying2026truthspectrum,
+  title  = {The Truthfulness Spectrum Hypothesis},
+  author = {Ying, Zhuofan Josh and Ravfogel, Shauli and Kriegeskorte, Nikolaus and Hase, Peter},
+  journal= {arXiv preprint arXiv:2602.20273}, year = {2026}}
+
+% [VERIFIED] 2310.06824, COLM 2024
+@article{marks2024geometry,
+  title  = {The Geometry of Truth: Emergent Linear Structure in {LLM} Representations of True/False Datasets},
+  author = {Marks, Samuel and Tegmark, Max},
+  journal= {Conference on Language Modeling (COLM); arXiv:2310.06824}, year = {2024}}
+
+% [CHECK] 2511.04312
+@article{lysnaes2025probing,
+  title  = {Probing the Probes: Methods and Metrics for Concept Alignment},
+  author = {Lysn{\ae}s-Larsen, M. and Eggen, P. and Str{\"u}mke, I.},
+  journal= {arXiv preprint arXiv:2511.04312}, year = {2025}}
+
+% [CHECK] 2606.11375
+@article{reblitz2026fragility,
+  title  = {When Probing Accuracy Saturates, Fragility Resolves},
+  author = {Reblitz-Richardson, O. and others},
+  journal= {arXiv preprint arXiv:2606.11375}, year = {2026}}
+
+% [CHECK] 2605.27958
+@article{deception2026pressure,
+  title  = {Pressure-Testing Deception Probes in {LLM}s: Scaling, Robustness, and the Geometry of Deceptive Representations},
+  author = {Anonymous},
+  journal= {arXiv preprint arXiv:2605.27958}, year = {2026}}
+
+@inproceedings{hewitt2019control,
+  title  = {Designing and Interpreting Probes with Control Tasks},
+  author = {Hewitt, John and Liang, Percy},
+  booktitle = {EMNLP}, year = {2019}}
+
+@article{belinkov2022probing,
+  title  = {Probing Classifiers: Promises, Shortcomings, and Advances},
+  author = {Belinkov, Yonatan}, journal = {Computational Linguistics}, volume = {48}, number = {1}, year = {2022}}
+
+@inproceedings{park2024linear,
+  title  = {The Linear Representation Hypothesis and the Geometry of Large Language Models},
+  author = {Park, Kiho and Choe, Yo Joong and Veitch, Victor},
+  booktitle = {ICML}, year = {2024}}
+
+@inproceedings{biderman2023pythia,
+  title  = {Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling},
+  author = {Biderman, Stella and others}, booktitle = {ICML}, year = {2023}}

repro_bundle/references.bib.bak

@@ -0,0 +1,71 @@
+% ProbeShift references. arXiv IDs marked [VERIFIED] were fetched directly from arxiv.org on
+% 2026-06-22 (title/authors confirmed). IDs marked [CHECK] were surfaced by automated search
+% and MUST be re-verified (venue/ID) before camera-ready.
+
+@article{dies2025pstat,             % [VERIFIED] 2511.19166
+  title  = {Representational and Behavioral Stability of Truth in Large Language Models},
+  author = {Dies, Samantha and Maynard, Courtney and Savcisens, Germans and Eliassi-Rad, Tina},
+  journal= {arXiv preprint arXiv:2511.19166}, year = {2025}}
+
+@article{huang2025sip,              % [VERIFIED] 2511.16288
+  title  = {Spectral Identifiability for Interpretable Probe Geometry},
+  author = {Huang, William Hao-Cheng},
+  journal= {arXiv preprint arXiv:2511.16288}, year = {2025}}
+
+@article{gao2026raptor,             % [VERIFIED] 2602.00158
+  title  = {{RAPTOR}: Ridge-Adaptive Logistic Probes},
+  author = {Gao, Ziqi and Zhu, Yaotian and Zeng, Qingcheng and Zhao, Xu and Wang, Ziqing and Ruan, Feng and Ding, Kaize},
+  journal= {arXiv preprint arXiv:2602.00158}, year = {2026}}
+
+@article{haller2025brittle,         % [VERIFIED] 2510.11905
+  title  = {{LLM} Knowledge is Brittle: Truthfulness Representations Rely on Superficial Resemblance},
+  author = {Haller, Patrick and Ibrahim, Mark and Kirichenko, Polina and Sagun, Levent and Bell, Samuel J.},
+  journal= {arXiv preprint arXiv:2510.11905}, year = {2025}}
+
+@inproceedings{xie2023feature,      % [VERIFIED] 2303.15488, NeurIPS 2023
+  title  = {On the Importance of Feature Separability in Predicting Out-Of-Distribution Error},
+  author = {Xie, Renchunzi and Wei, Hongxin and Feng, Lei and Cao, Yuzhou and An, Bo},
+  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)}, year = {2023}}
+
+@article{ying2026truthspectrum,     % [VERIFIED] 2602.20273
+  title  = {The Truthfulness Spectrum Hypothesis},
+  author = {Ying, Zhuofan Josh and Ravfogel, Shauli and Kriegeskorte, Nikolaus and Hase, Peter},
+  journal= {arXiv preprint arXiv:2602.20273}, year = {2026}}
+
+@article{marks2024geometry,         % [VERIFIED] 2310.06824, COLM 2024
+  title  = {The Geometry of Truth: Emergent Linear Structure in {LLM} Representations of True/False Datasets},
+  author = {Marks, Samuel and Tegmark, Max},
+  journal= {Conference on Language Modeling (COLM); arXiv:2310.06824}, year = {2024}}
+
+@article{lysnaes2025probing,        % [CHECK] 2511.04312
+  title  = {Probing the Probes: Methods and Metrics for Concept Alignment},
+  author = {Lysn{\ae}s-Larsen, M. and Eggen, P. and Str{\"u}mke, I.},
+  journal= {arXiv preprint arXiv:2511.04312}, year = {2025}}
+
+@article{reblitz2026fragility,      % [CHECK] 2606.11375
+  title  = {When Probing Accuracy Saturates, Fragility Resolves},
+  author = {Reblitz-Richardson, O. and others},
+  journal= {arXiv preprint arXiv:2606.11375}, year = {2026}}
+
+@article{deception2026pressure,     % [CHECK] 2605.27958
+  title  = {Pressure-Testing Deception Probes in {LLM}s: Scaling, Robustness, and the Geometry of Deceptive Representations},
+  author = {Anonymous},
+  journal= {arXiv preprint arXiv:2605.27958}, year = {2026}}
+
+@inproceedings{hewitt2019control,
+  title  = {Designing and Interpreting Probes with Control Tasks},
+  author = {Hewitt, John and Liang, Percy},
+  booktitle = {EMNLP}, year = {2019}}
+
+@article{belinkov2022probing,
+  title  = {Probing Classifiers: Promises, Shortcomings, and Advances},
+  author = {Belinkov, Yonatan}, journal = {Computational Linguistics}, volume = {48}, number = {1}, year = {2022}}
+
+@inproceedings{park2024linear,
+  title  = {The Linear Representation Hypothesis and the Geometry of Large Language Models},
+  author = {Park, Kiho and Choe, Yo Joong and Veitch, Victor},
+  booktitle = {ICML}, year = {2024}}
+
+@inproceedings{biderman2023pythia,
+  title  = {Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling},
+  author = {Biderman, Stella and others}, booktitle = {ICML}, year = {2023}}

repro_bundle/requirements.txt

@@ -0,0 +1,17 @@
+# Core
+numpy>=1.26
+scipy>=1.11
+scikit-learn>=1.4
+pandas>=2.1
+
+# Deep learning / models (install the CUDA build of torch matching your 4090 driver)
+torch>=2.2
+transformers>=4.44
+datasets>=2.20
+accelerate>=0.33
+sentencepiece>=0.2          # for NLLB / Marian / Qwen tokenizers
+sacremoses>=0.1             # for Marian (opus-mt) back-translation
+
+# Utilities
+tqdm>=4.66
+pyyaml>=6.0

repro_bundle/results_A/results/audit.jsonl

@@ -0,0 +1,70 @@
+{"dataset": "sst2", "seed": 0, "flip_rate": 0.09750000000000003, "n": 800}
+{"dataset": "imdb", "seed": 0, "flip_rate": 0.355, "n": 800}
+{"dataset": "ag_news", "seed": 0, "flip_rate": 0.1875, "n": 800}
+{"dataset": "dbpedia", "seed": 0, "flip_rate": 0.49875, "n": 800}
+{"dataset": "counterfact", "seed": 0, "flip_rate": 0.043749999999999956, "n": 800}
+{"dataset": "emotion", "seed": 0, "flip_rate": 0.06125000000000003, "n": 800}
+{"dataset": "tweet_hate", "seed": 0, "flip_rate": 0.26375000000000004, "n": 800}
+{"dataset": "tweet_irony", "seed": 0, "flip_rate": 0.22499999999999998, "n": 800}
+{"dataset": "tweet_offensive", "seed": 0, "flip_rate": 0.22375, "n": 800}
+{"dataset": "subj", "seed": 0, "flip_rate": 0.08875, "n": 800}
+{"dataset": "spam", "seed": 0, "flip_rate": 0.27125, "n": 800}
+{"dataset": "cola", "seed": 0, "flip_rate": 0.13749999999999996, "n": 800}
+{"dataset": "stance", "seed": 0, "flip_rate": 0.15384615384615385, "n": 208}
+{"dataset": "amazon_cf", "seed": 0, "flip_rate": 0.12250000000000005, "n": 800}
+{"dataset": "sst2", "seed": 1, "flip_rate": 0.10124999999999995, "n": 800}
+{"dataset": "imdb", "seed": 1, "flip_rate": 0.38749999999999996, "n": 800}
+{"dataset": "ag_news", "seed": 1, "flip_rate": 0.20750000000000002, "n": 800}
+{"dataset": "dbpedia", "seed": 1, "flip_rate": 0.5137499999999999, "n": 800}
+{"dataset": "counterfact", "seed": 1, "flip_rate": 0.03249999999999997, "n": 800}
+{"dataset": "emotion", "seed": 1, "flip_rate": 0.07625000000000004, "n": 800}
+{"dataset": "tweet_hate", "seed": 1, "flip_rate": 0.26, "n": 800}
+{"dataset": "tweet_irony", "seed": 1, "flip_rate": 0.23124999999999996, "n": 800}
+{"dataset": "tweet_offensive", "seed": 1, "flip_rate": 0.23375, "n": 800}
+{"dataset": "subj", "seed": 1, "flip_rate": 0.08250000000000002, "n": 800}
+{"dataset": "spam", "seed": 1, "flip_rate": 0.24250000000000005, "n": 800}
+{"dataset": "cola", "seed": 1, "flip_rate": 0.13624999999999998, "n": 800}
+{"dataset": "stance", "seed": 1, "flip_rate": 0.1826923076923077, "n": 208}
+{"dataset": "amazon_cf", "seed": 1, "flip_rate": 0.11375000000000002, "n": 800}
+{"dataset": "sst2", "seed": 2, "flip_rate": 0.09875, "n": 800}
+{"dataset": "imdb", "seed": 2, "flip_rate": 0.35624999999999996, "n": 800}
+{"dataset": "ag_news", "seed": 2, "flip_rate": 0.20625000000000004, "n": 800}
+{"dataset": "dbpedia", "seed": 2, "flip_rate": 0.5037499999999999, "n": 800}
+{"dataset": "counterfact", "seed": 2, "flip_rate": 0.04125000000000001, "n": 800}
+{"dataset": "emotion", "seed": 2, "flip_rate": 0.05874999999999997, "n": 800}
+{"dataset": "tweet_hate", "seed": 2, "flip_rate": 0.22750000000000004, "n": 800}
+{"dataset": "tweet_irony", "seed": 2, "flip_rate": 0.22624999999999995, "n": 800}
+{"dataset": "tweet_offensive", "seed": 2, "flip_rate": 0.23875000000000002, "n": 800}
+{"dataset": "subj", "seed": 2, "flip_rate": 0.09875, "n": 800}
+{"dataset": "spam", "seed": 2, "flip_rate": 0.24124999999999996, "n": 800}
+{"dataset": "cola", "seed": 2, "flip_rate": 0.10875000000000001, "n": 800}
+{"dataset": "stance", "seed": 2, "flip_rate": 0.14903846153846156, "n": 208}
+{"dataset": "amazon_cf", "seed": 2, "flip_rate": 0.08750000000000002, "n": 800}
+{"dataset": "sst2", "seed": 3, "flip_rate": 0.11250000000000004, "n": 800}
+{"dataset": "imdb", "seed": 3, "flip_rate": 0.3425, "n": 800}
+{"dataset": "ag_news", "seed": 3, "flip_rate": 0.19874999999999998, "n": 800}
+{"dataset": "dbpedia", "seed": 3, "flip_rate": 0.47875, "n": 800}
+{"dataset": "counterfact", "seed": 3, "flip_rate": 0.043749999999999956, "n": 800}
+{"dataset": "emotion", "seed": 3, "flip_rate": 0.07250000000000001, "n": 800}
+{"dataset": "tweet_hate", "seed": 3, "flip_rate": 0.25125, "n": 800}
+{"dataset": "tweet_irony", "seed": 3, "flip_rate": 0.24250000000000005, "n": 800}
+{"dataset": "tweet_offensive", "seed": 3, "flip_rate": 0.2025, "n": 800}
+{"dataset": "subj", "seed": 3, "flip_rate": 0.10375000000000001, "n": 800}
+{"dataset": "spam", "seed": 3, "flip_rate": 0.23624999999999996, "n": 800}
+{"dataset": "cola", "seed": 3, "flip_rate": 0.11250000000000004, "n": 800}
+{"dataset": "stance", "seed": 3, "flip_rate": 0.16826923076923073, "n": 208}
+{"dataset": "amazon_cf", "seed": 3, "flip_rate": 0.09875, "n": 800}
+{"dataset": "sst2", "seed": 4, "flip_rate": 0.12124999999999997, "n": 800}
+{"dataset": "imdb", "seed": 4, "flip_rate": 0.36250000000000004, "n": 800}
+{"dataset": "ag_news", "seed": 4, "flip_rate": 0.17500000000000004, "n": 800}
+{"dataset": "dbpedia", "seed": 4, "flip_rate": 0.48624999999999996, "n": 800}
+{"dataset": "counterfact", "seed": 4, "flip_rate": 0.03874999999999995, "n": 800}
+{"dataset": "emotion", "seed": 4, "flip_rate": 0.07250000000000001, "n": 800}
+{"dataset": "tweet_hate", "seed": 4, "flip_rate": 0.24750000000000005, "n": 800}
+{"dataset": "tweet_irony", "seed": 4, "flip_rate": 0.24624999999999997, "n": 800}
+{"dataset": "tweet_offensive", "seed": 4, "flip_rate": 0.23624999999999996, "n": 800}
+{"dataset": "subj", "seed": 4, "flip_rate": 0.08125000000000004, "n": 800}
+{"dataset": "spam", "seed": 4, "flip_rate": 0.24124999999999996, "n": 800}
+{"dataset": "cola", "seed": 4, "flip_rate": 0.11499999999999999, "n": 800}
+{"dataset": "stance", "seed": 4, "flip_rate": 0.1875, "n": 208}
+{"dataset": "amazon_cf", "seed": 4, "flip_rate": 0.13, "n": 800}

repro_bundle/results_A/results/stats_ranking.jsonl

@@ -0,0 +1,7 @@
+{"target": "acc_drop", "n_configs": 472, "table": [{"predictor": "augmentation_robustness", "spearman": 0.17628948336103584, "p": 0.00011810508174001726, "kendall": 0.11986019460492477, "ci95": [0.0864486183837455, 0.2656858828070762], "loco_mean": 0.06306440054904434, "n": 472}, {"predictor": "pac", "spearman": 0.11790738396920045, "p": 0.010354502513996207, "kendall": 0.07946596854652109, "ci95": [0.03199899524267362, 0.19739551469136], "loco_mean": 0.03027311003896434, "n": 472}, {"predictor": "fragility", "spearman": 0.10714381169126685, "p": 0.019896547808716226, "kendall": 0.0819652122104107, "ci95": [0.02184423176267955, 0.19486119226176976], "loco_mean": -0.08367198887077822, "n": 472}, {"predictor": "xie_feature_dispersion", "spearman": 0.05468516346193012, "p": 0.23570006318676667, "kendall": 0.0401513209128409, "ci95": [-0.03209005896781303, 0.1425023272515047], "loco_mean": -0.012901634988281898, "n": 472}, {"predictor": "sip_eigengap", "spearman": 0.000509027245282521, "p": 0.9911998467763682, "kendall": 0.0014395350322827368, "ci95": [-0.08621837109724974, 0.09490063671198525], "loco_mean": 0.10342197319983981, "n": 472}, {"predictor": "raptor_stability", "spearman": -0.062397712100437104, "p": 0.17594389124505538, "kendall": -0.04115892744258281, "ci95": [-0.14459614732709314, 0.02266644621275267], "loco_mean": -0.06425062206029349, "n": 472}, {"predictor": "whitened_cosine_id", "spearman": -0.07663916907355152, "p": 0.0963002843104417, "kendall": -0.05037133000022321, "ci95": [-0.16246440320558697, 0.010007628269806309], "loco_mean": -0.09396010191341637, "n": 472}]}
+{"target": "rotation", "n_configs": 472, "table": [{"predictor": "raptor_stability", "spearman": 0.9449489763429444, "p": 3.6596555462199364e-230, "kendall": 0.7988412681276762, "ci95": [0.9320857918652855, 0.9546405882585741], "loco_mean": 0.714095456119887, "n": 472}, {"predictor": "pac", "spearman": 0.8386192401102935, "p": 4.389790032967171e-126, "kendall": 0.6566807009967972, "ci95": [0.8041701035011758, 0.8686758317302297], "loco_mean": 0.7336833473426215, "n": 472}, {"predictor": "whitened_cosine_id", "spearman": 0.7690200093623065, "p": 2.168995457934309e-93, "kendall": 0.5761092518622476, "ci95": [0.7221527504365517, 0.8086111694029999], "loco_mean": 0.17490397542579705, "n": 472}, {"predictor": "augmentation_robustness", "spearman": 0.7108589060252217, "p": 7.570490911324105e-74, "kendall": 0.5244521213429774, "ci95": [0.6647373626280134, 0.7590285725368467], "loco_mean": 0.6466917724774698, "n": 472}, {"predictor": "fragility", "spearman": 0.6402369351573364, "p": 8.412937618018045e-56, "kendall": 0.5176148935000774, "ci95": [0.5762098860164254, 0.6965181301409472], "loco_mean": 0.34843798881555416, "n": 472}, {"predictor": "xie_feature_dispersion", "spearman": 0.4577071159234756, "p": 8.112397111719808e-26, "kendall": 0.30139623592068804, "ci95": [0.3719011054200997, 0.5362764648442968], "loco_mean": -0.15587895297039037, "n": 472}, {"predictor": "sip_eigengap", "spearman": -0.07233275965780582, "p": 0.11656426348042145, "kendall": -0.045983526648003925, "ci95": [-0.1676638264409657, 0.0276888325777201], "loco_mean": 0.27295969447731966, "n": 472}]}
+{"target": "iid_split_rotation", "n_configs": 472, "table": [{"predictor": "raptor_stability", "spearman": 0.9663956649462877, "p": 1.9769716421360586e-279, "kendall": 0.844200942819101, "ci95": [0.9582833668806262, 0.972249511018557], "loco_mean": 0.822013981169579, "n": 472}, {"predictor": "whitened_cosine_id", "spearman": 0.8142104839116457, "p": 4.6349247787002345e-113, "kendall": 0.6190938860700277, "ci95": [0.7784315483263554, 0.8457827980307192], "loco_mean": 0.2507519638805218, "n": 472}, {"predictor": "pac", "spearman": 0.7989523437605205, "p": 7.370681520368353e-106, "kendall": 0.6131202993990427, "ci95": [0.7548202970707872, 0.8362192541519101], "loco_mean": 0.7092535567689743, "n": 472}, {"predictor": "augmentation_robustness", "spearman": 0.6528079273309736, "p": 1.1636128383807618e-58, "kendall": 0.47211126704811257, "ci95": [0.594467447577123, 0.7075880242965953], "loco_mean": 0.5938518661656708, "n": 472}, {"predictor": "fragility", "spearman": 0.6434621748707247, "p": 1.6002788335451959e-56, "kendall": 0.5162034040583596, "ci95": [0.5777199855374996, 0.6972436181916808], "loco_mean": 0.30639625747257265, "n": 472}, {"predictor": "xie_feature_dispersion", "spearman": 0.46410427354367867, "p": 1.377354979546913e-26, "kendall": 0.31332541653172113, "ci95": [0.38047007067527516, 0.5387551616620547], "loco_mean": -0.07487413986761712, "n": 472}, {"predictor": "sip_eigengap", "spearman": -0.05687906812360867, "p": 0.21740724040384907, "kendall": -0.03140843113445152, "ci95": [-0.15064310621609026, 0.04683266118594591], "loco_mean": 0.24522656676053, "n": 472}]}
+{"target": "excess_rotation", "n_configs": 472, "table": [{"predictor": "augmentation_robustness", "spearman": 0.045730142358639335, "p": 0.3214915785134481, "kendall": 0.029813955162114502, "ci95": [-0.04291475943890026, 0.13280723679010745], "loco_mean": 0.11571837660547339, "n": 472}, {"predictor": "sip_eigengap", "spearman": 0.019548618802250367, "p": 0.6718436438563953, "kendall": 0.01587965653174074, "ci95": [-0.06592842183087877, 0.11021518205534078], "loco_mean": 0.11096384112403068, "n": 472}, {"predictor": "pac", "spearman": -0.028921926535337233, "p": 0.5307837018089083, "kendall": -0.022059088128396125, "ci95": [-0.11461587366461326, 0.05697485566708844], "loco_mean": 0.0566288049921826, "n": 472}, {"predictor": "xie_feature_dispersion", "spearman": -0.11740393878758819, "p": 0.010688017745394272, "kendall": -0.08521357371621864, "ci95": [-0.20901249520866028, -0.030594774861803075], "loco_mean": -0.17614889459054547, "n": 472}, {"predictor": "fragility", "spearman": -0.13634890441214695, "p": 0.0029947554168688454, "kendall": -0.0995317208632813, "ci95": [-0.2216850177242833, -0.05109443013239468], "loco_mean": 0.01926842737612711, "n": 472}, {"predictor": "raptor_stability", "spearman": -0.19517738512287972, "p": 1.9527049508461923e-05, "kendall": -0.12812623700025189, "ci95": [-0.28507442645385095, -0.11649292054466626], "loco_mean": -0.11715629043368607, "n": 472}, {"predictor": "whitened_cosine_id", "spearman": -0.24497186857439832, "p": 7.030580997726293e-08, "kendall": -0.1661088920076289, "ci95": [-0.3367444423080759, -0.15504169609386367], "loco_mean": -0.11500811181491921, "n": 472}]}
+{"target": "acc_drop", "n_configs": 472, "table": [{"predictor": "augmentation_robustness", "spearman": 0.17628948336103584, "p": 0.00011810508174001726, "kendall": 0.11986019460492477, "ci95": [0.0864486183837455, 0.2656858828070762], "loco_mean": 0.06306440054904434, "n": 472}, {"predictor": "pac", "spearman": 0.11790738396920045, "p": 0.010354502513996207, "kendall": 0.07946596854652109, "ci95": [0.03199899524267362, 0.19739551469136], "loco_mean": 0.03027311003896434, "n": 472}, {"predictor": "fragility", "spearman": 0.10714381169126685, "p": 0.019896547808716226, "kendall": 0.0819652122104107, "ci95": [0.02184423176267955, 0.19486119226176976], "loco_mean": -0.08367198887077822, "n": 472}, {"predictor": "xie_feature_dispersion", "spearman": 0.05468516346193012, "p": 0.23570006318676667, "kendall": 0.0401513209128409, "ci95": [-0.03209005896781303, 0.1425023272515047], "loco_mean": -0.012901634988281898, "n": 472}, {"predictor": "sip_eigengap", "spearman": 0.000509027245282521, "p": 0.9911998467763682, "kendall": 0.0014395350322827368, "ci95": [-0.08621837109724974, 0.09490063671198525], "loco_mean": 0.10342197319983981, "n": 472}, {"predictor": "raptor_stability", "spearman": -0.062397712100437104, "p": 0.17594389124505538, "kendall": -0.04115892744258281, "ci95": [-0.14459614732709314, 0.02266644621275267], "loco_mean": -0.06425062206029349, "n": 472}, {"predictor": "whitened_cosine_id", "spearman": -0.07663916907355152, "p": 0.0963002843104417, "kendall": -0.05037133000022321, "ci95": [-0.16246440320558697, 0.010007628269806309], "loco_mean": -0.09396010191341637, "n": 472}]}
+{"target": "rotation", "n_configs": 472, "table": [{"predictor": "raptor_stability", "spearman": 0.9449489763429444, "p": 3.6596555462199364e-230, "kendall": 0.7988412681276762, "ci95": [0.9320857918652855, 0.9546405882585741], "loco_mean": 0.714095456119887, "n": 472}, {"predictor": "pac", "spearman": 0.8386192401102935, "p": 4.389790032967171e-126, "kendall": 0.6566807009967972, "ci95": [0.8041701035011758, 0.8686758317302297], "loco_mean": 0.7336833473426215, "n": 472}, {"predictor": "whitened_cosine_id", "spearman": 0.7690200093623065, "p": 2.168995457934309e-93, "kendall": 0.5761092518622476, "ci95": [0.7221527504365517, 0.8086111694029999], "loco_mean": 0.17490397542579705, "n": 472}, {"predictor": "augmentation_robustness", "spearman": 0.7108589060252217, "p": 7.570490911324105e-74, "kendall": 0.5244521213429774, "ci95": [0.6647373626280134, 0.7590285725368467], "loco_mean": 0.6466917724774698, "n": 472}, {"predictor": "fragility", "spearman": 0.6402369351573364, "p": 8.412937618018045e-56, "kendall": 0.5176148935000774, "ci95": [0.5762098860164254, 0.6965181301409472], "loco_mean": 0.34843798881555416, "n": 472}, {"predictor": "xie_feature_dispersion", "spearman": 0.4577071159234756, "p": 8.112397111719808e-26, "kendall": 0.30139623592068804, "ci95": [0.3719011054200997, 0.5362764648442968], "loco_mean": -0.15587895297039037, "n": 472}, {"predictor": "sip_eigengap", "spearman": -0.07233275965780582, "p": 0.11656426348042145, "kendall": -0.045983526648003925, "ci95": [-0.1676638264409657, 0.0276888325777201], "loco_mean": 0.27295969447731966, "n": 472}]}
+{"target": "iid_split_rotation", "n_configs": 472, "table": [{"predictor": "raptor_stability", "spearman": 0.9663956649462877, "p": 1.9769716421360586e-279, "kendall": 0.844200942819101, "ci95": [0.9582833668806262, 0.972249511018557], "loco_mean": 0.822013981169579, "n": 472}, {"predictor": "whitened_cosine_id", "spearman": 0.8142104839116457, "p": 4.6349247787002345e-113, "kendall": 0.6190938860700277, "ci95": [0.7784315483263554, 0.8457827980307192], "loco_mean": 0.2507519638805218, "n": 472}, {"predictor": "pac", "spearman": 0.7989523437605205, "p": 7.370681520368353e-106, "kendall": 0.6131202993990427, "ci95": [0.7548202970707872, 0.8362192541519101], "loco_mean": 0.7092535567689743, "n": 472}, {"predictor": "augmentation_robustness", "spearman": 0.6528079273309736, "p": 1.1636128383807618e-58, "kendall": 0.47211126704811257, "ci95": [0.594467447577123, 0.7075880242965953], "loco_mean": 0.5938518661656708, "n": 472}, {"predictor": "fragility", "spearman": 0.6434621748707247, "p": 1.6002788335451959e-56, "kendall": 0.5162034040583596, "ci95": [0.5777199855374996, 0.6972436181916808], "loco_mean": 0.30639625747257265, "n": 472}, {"predictor": "xie_feature_dispersion", "spearman": 0.46410427354367867, "p": 1.377354979546913e-26, "kendall": 0.31332541653172113, "ci95": [0.38047007067527516, 0.5387551616620547], "loco_mean": -0.07487413986761712, "n": 472}, {"predictor": "sip_eigengap", "spearman": -0.05687906812360867, "p": 0.21740724040384907, "kendall": -0.03140843113445152, "ci95": [-0.15064310621609026, 0.04683266118594591], "loco_mean": 0.24522656676053, "n": 472}]}

repro_bundle/results_final/results/audit.jsonl

@@ -0,0 +1,162 @@
+{"dataset": "sst2", "seed": 0, "flip_rate": 0.09750000000000003, "n": 800}
+{"dataset": "imdb", "seed": 0, "flip_rate": 0.355, "n": 800}
+{"dataset": "ag_news", "seed": 0, "flip_rate": 0.1875, "n": 800}
+{"dataset": "dbpedia", "seed": 0, "flip_rate": 0.49875, "n": 800}
+{"dataset": "counterfact", "seed": 0, "flip_rate": 0.043749999999999956, "n": 800}
+{"dataset": "emotion", "seed": 0, "flip_rate": 0.06125000000000003, "n": 800}
+{"dataset": "tweet_hate", "seed": 0, "flip_rate": 0.26375000000000004, "n": 800}
+{"dataset": "tweet_irony", "seed": 0, "flip_rate": 0.22499999999999998, "n": 800}
+{"dataset": "tweet_offensive", "seed": 0, "flip_rate": 0.22375, "n": 800}
+{"dataset": "subj", "seed": 0, "flip_rate": 0.08875, "n": 800}
+{"dataset": "spam", "seed": 0, "flip_rate": 0.27125, "n": 800}
+{"dataset": "cola", "seed": 0, "flip_rate": 0.13749999999999996, "n": 800}
+{"dataset": "stance", "seed": 0, "flip_rate": 0.15384615384615385, "n": 208}
+{"dataset": "amazon_cf", "seed": 0, "flip_rate": 0.12250000000000005, "n": 800}
+{"dataset": "sst2", "seed": 1, "flip_rate": 0.10124999999999995, "n": 800}
+{"dataset": "imdb", "seed": 1, "flip_rate": 0.38749999999999996, "n": 800}
+{"dataset": "ag_news", "seed": 1, "flip_rate": 0.20750000000000002, "n": 800}
+{"dataset": "dbpedia", "seed": 1, "flip_rate": 0.5137499999999999, "n": 800}
+{"dataset": "counterfact", "seed": 1, "flip_rate": 0.03249999999999997, "n": 800}
+{"dataset": "emotion", "seed": 1, "flip_rate": 0.07625000000000004, "n": 800}
+{"dataset": "tweet_hate", "seed": 1, "flip_rate": 0.26, "n": 800}
+{"dataset": "tweet_irony", "seed": 1, "flip_rate": 0.23124999999999996, "n": 800}
+{"dataset": "tweet_offensive", "seed": 1, "flip_rate": 0.23375, "n": 800}
+{"dataset": "subj", "seed": 1, "flip_rate": 0.08250000000000002, "n": 800}
+{"dataset": "spam", "seed": 1, "flip_rate": 0.24250000000000005, "n": 800}
+{"dataset": "cola", "seed": 1, "flip_rate": 0.13624999999999998, "n": 800}
+{"dataset": "stance", "seed": 1, "flip_rate": 0.1826923076923077, "n": 208}
+{"dataset": "amazon_cf", "seed": 1, "flip_rate": 0.11375000000000002, "n": 800}
+{"dataset": "sst2", "seed": 2, "flip_rate": 0.09875, "n": 800}
+{"dataset": "imdb", "seed": 2, "flip_rate": 0.35624999999999996, "n": 800}
+{"dataset": "ag_news", "seed": 2, "flip_rate": 0.20625000000000004, "n": 800}
+{"dataset": "dbpedia", "seed": 2, "flip_rate": 0.5037499999999999, "n": 800}
+{"dataset": "counterfact", "seed": 2, "flip_rate": 0.04125000000000001, "n": 800}
+{"dataset": "emotion", "seed": 2, "flip_rate": 0.05874999999999997, "n": 800}
+{"dataset": "tweet_hate", "seed": 2, "flip_rate": 0.22750000000000004, "n": 800}
+{"dataset": "tweet_irony", "seed": 2, "flip_rate": 0.22624999999999995, "n": 800}
+{"dataset": "tweet_offensive", "seed": 2, "flip_rate": 0.23875000000000002, "n": 800}
+{"dataset": "subj", "seed": 2, "flip_rate": 0.09875, "n": 800}
+{"dataset": "spam", "seed": 2, "flip_rate": 0.24124999999999996, "n": 800}
+{"dataset": "cola", "seed": 2, "flip_rate": 0.10875000000000001, "n": 800}
+{"dataset": "stance", "seed": 2, "flip_rate": 0.14903846153846156, "n": 208}
+{"dataset": "amazon_cf", "seed": 2, "flip_rate": 0.08750000000000002, "n": 800}
+{"dataset": "sst2", "seed": 3, "flip_rate": 0.11250000000000004, "n": 800}
+{"dataset": "imdb", "seed": 3, "flip_rate": 0.3425, "n": 800}
+{"dataset": "ag_news", "seed": 3, "flip_rate": 0.19874999999999998, "n": 800}
+{"dataset": "dbpedia", "seed": 3, "flip_rate": 0.47875, "n": 800}
+{"dataset": "counterfact", "seed": 3, "flip_rate": 0.043749999999999956, "n": 800}
+{"dataset": "emotion", "seed": 3, "flip_rate": 0.07250000000000001, "n": 800}
+{"dataset": "tweet_hate", "seed": 3, "flip_rate": 0.25125, "n": 800}
+{"dataset": "tweet_irony", "seed": 3, "flip_rate": 0.24250000000000005, "n": 800}
+{"dataset": "tweet_offensive", "seed": 3, "flip_rate": 0.2025, "n": 800}
+{"dataset": "subj", "seed": 3, "flip_rate": 0.10375000000000001, "n": 800}
+{"dataset": "spam", "seed": 3, "flip_rate": 0.23624999999999996, "n": 800}
+{"dataset": "cola", "seed": 3, "flip_rate": 0.11250000000000004, "n": 800}
+{"dataset": "stance", "seed": 3, "flip_rate": 0.16826923076923073, "n": 208}
+{"dataset": "amazon_cf", "seed": 3, "flip_rate": 0.09875, "n": 800}
+{"dataset": "sst2", "seed": 4, "flip_rate": 0.12124999999999997, "n": 800}
+{"dataset": "imdb", "seed": 4, "flip_rate": 0.36250000000000004, "n": 800}
+{"dataset": "ag_news", "seed": 4, "flip_rate": 0.17500000000000004, "n": 800}
+{"dataset": "dbpedia", "seed": 4, "flip_rate": 0.48624999999999996, "n": 800}
+{"dataset": "counterfact", "seed": 4, "flip_rate": 0.03874999999999995, "n": 800}
+{"dataset": "emotion", "seed": 4, "flip_rate": 0.07250000000000001, "n": 800}
+{"dataset": "tweet_hate", "seed": 4, "flip_rate": 0.24750000000000005, "n": 800}
+{"dataset": "tweet_irony", "seed": 4, "flip_rate": 0.24624999999999997, "n": 800}
+{"dataset": "tweet_offensive", "seed": 4, "flip_rate": 0.23624999999999996, "n": 800}
+{"dataset": "subj", "seed": 4, "flip_rate": 0.08125000000000004, "n": 800}
+{"dataset": "spam", "seed": 4, "flip_rate": 0.24124999999999996, "n": 800}
+{"dataset": "cola", "seed": 4, "flip_rate": 0.11499999999999999, "n": 800}
+{"dataset": "stance", "seed": 4, "flip_rate": 0.1875, "n": 208}
+{"dataset": "amazon_cf", "seed": 4, "flip_rate": 0.13, "n": 800}
+{"dataset": "sst2", "seed": 0, "flip_rate": 0.09750000000000003, "n": 800}
+{"dataset": "imdb", "seed": 0, "flip_rate": 0.355, "n": 800}
+{"dataset": "ag_news", "seed": 0, "flip_rate": 0.1875, "n": 800}
+{"dataset": "dbpedia", "seed": 0, "flip_rate": 0.49875, "n": 800}
+{"dataset": "counterfact", "seed": 0, "flip_rate": 0.043749999999999956, "n": 800}
+{"dataset": "emotion", "seed": 0, "flip_rate": 0.06125000000000003, "n": 800}
+{"dataset": "tweet_hate", "seed": 0, "flip_rate": 0.26375000000000004, "n": 800}
+{"dataset": "tweet_irony", "seed": 0, "flip_rate": 0.22499999999999998, "n": 800}
+{"dataset": "tweet_offensive", "seed": 0, "flip_rate": 0.22375, "n": 800}
+{"dataset": "subj", "seed": 0, "flip_rate": 0.08875, "n": 800}
+{"dataset": "spam", "seed": 0, "flip_rate": 0.27125, "n": 800}
+{"dataset": "cola", "seed": 0, "flip_rate": 0.13749999999999996, "n": 800}
+{"dataset": "stance", "seed": 0, "flip_rate": 0.15384615384615385, "n": 208}
+{"dataset": "amazon_cf", "seed": 0, "flip_rate": 0.12250000000000005, "n": 800}
+{"dataset": "sst2", "seed": 1, "flip_rate": 0.10124999999999995, "n": 800}
+{"dataset": "sst2", "seed": 0, "flip_rate": 0.09750000000000003, "n": 800}
+{"dataset": "imdb", "seed": 0, "flip_rate": 0.355, "n": 800}
+{"dataset": "ag_news", "seed": 0, "flip_rate": 0.1875, "n": 800}
+{"dataset": "dbpedia", "seed": 0, "flip_rate": 0.49875, "n": 800}
+{"dataset": "counterfact", "seed": 0, "flip_rate": 0.043749999999999956, "n": 800}
+{"dataset": "emotion", "seed": 0, "flip_rate": 0.06125000000000003, "n": 800}
+{"dataset": "tweet_hate", "seed": 0, "flip_rate": 0.26375000000000004, "n": 800}
+{"dataset": "tweet_irony", "seed": 0, "flip_rate": 0.22499999999999998, "n": 800}
+{"dataset": "tweet_offensive", "seed": 0, "flip_rate": 0.22375, "n": 800}
+{"dataset": "subj", "seed": 0, "flip_rate": 0.08875, "n": 800}
+{"dataset": "spam", "seed": 0, "flip_rate": 0.27125, "n": 800}
+{"dataset": "cola", "seed": 0, "flip_rate": 0.13749999999999996, "n": 800}
+{"dataset": "stance", "seed": 0, "flip_rate": 0.15384615384615385, "n": 208}
+{"dataset": "amazon_cf", "seed": 0, "flip_rate": 0.12250000000000005, "n": 800}
+{"dataset": "sst2", "seed": 1, "flip_rate": 0.10124999999999995, "n": 800}
+{"dataset": "imdb", "seed": 1, "flip_rate": 0.38749999999999996, "n": 800}
+{"dataset": "ag_news", "seed": 1, "flip_rate": 0.20750000000000002, "n": 800}
+{"dataset": "dbpedia", "seed": 1, "flip_rate": 0.5137499999999999, "n": 800}
+{"dataset": "counterfact", "seed": 1, "flip_rate": 0.03249999999999997, "n": 800}
+{"dataset": "emotion", "seed": 1, "flip_rate": 0.07625000000000004, "n": 800}
+{"dataset": "tweet_hate", "seed": 1, "flip_rate": 0.26, "n": 800}
+{"dataset": "tweet_irony", "seed": 1, "flip_rate": 0.23124999999999996, "n": 800}
+{"dataset": "tweet_offensive", "seed": 1, "flip_rate": 0.23375, "n": 800}
+{"dataset": "subj", "seed": 1, "flip_rate": 0.08250000000000002, "n": 800}
+{"dataset": "spam", "seed": 1, "flip_rate": 0.24250000000000005, "n": 800}
+{"dataset": "cola", "seed": 1, "flip_rate": 0.13624999999999998, "n": 800}
+{"dataset": "stance", "seed": 1, "flip_rate": 0.1826923076923077, "n": 208}
+{"dataset": "amazon_cf", "seed": 1, "flip_rate": 0.11375000000000002, "n": 800}
+{"dataset": "sst2", "seed": 2, "flip_rate": 0.09875, "n": 800}
+{"dataset": "imdb", "seed": 2, "flip_rate": 0.35624999999999996, "n": 800}
+{"dataset": "ag_news", "seed": 2, "flip_rate": 0.20625000000000004, "n": 800}
+{"dataset": "dbpedia", "seed": 2, "flip_rate": 0.5037499999999999, "n": 800}
+{"dataset": "counterfact", "seed": 2, "flip_rate": 0.04125000000000001, "n": 800}
+{"dataset": "emotion", "seed": 2, "flip_rate": 0.05874999999999997, "n": 800}
+{"dataset": "tweet_hate", "seed": 2, "flip_rate": 0.22750000000000004, "n": 800}
+{"dataset": "tweet_irony", "seed": 2, "flip_rate": 0.22624999999999995, "n": 800}
+{"dataset": "tweet_offensive", "seed": 2, "flip_rate": 0.23875000000000002, "n": 800}
+{"dataset": "subj", "seed": 2, "flip_rate": 0.09875, "n": 800}
+{"dataset": "spam", "seed": 2, "flip_rate": 0.24124999999999996, "n": 800}
+{"dataset": "cola", "seed": 2, "flip_rate": 0.10875000000000001, "n": 800}
+{"dataset": "stance", "seed": 2, "flip_rate": 0.14903846153846156, "n": 208}
+{"dataset": "amazon_cf", "seed": 2, "flip_rate": 0.08750000000000002, "n": 800}
+{"dataset": "sst2", "seed": 3, "flip_rate": 0.11250000000000004, "n": 800}
+{"dataset": "imdb", "seed": 3, "flip_rate": 0.3425, "n": 800}
+{"dataset": "ag_news", "seed": 3, "flip_rate": 0.19874999999999998, "n": 800}
+{"dataset": "dbpedia", "seed": 3, "flip_rate": 0.47875, "n": 800}
+{"dataset": "counterfact", "seed": 3, "flip_rate": 0.043749999999999956, "n": 800}
+{"dataset": "emotion", "seed": 3, "flip_rate": 0.07250000000000001, "n": 800}
+{"dataset": "tweet_hate", "seed": 3, "flip_rate": 0.25125, "n": 800}
+{"dataset": "tweet_irony", "seed": 3, "flip_rate": 0.24250000000000005, "n": 800}
+{"dataset": "tweet_offensive", "seed": 3, "flip_rate": 0.2025, "n": 800}
+{"dataset": "subj", "seed": 3, "flip_rate": 0.10375000000000001, "n": 800}
+{"dataset": "spam", "seed": 3, "flip_rate": 0.23624999999999996, "n": 800}
+{"dataset": "cola", "seed": 3, "flip_rate": 0.11250000000000004, "n": 800}
+{"dataset": "stance", "seed": 3, "flip_rate": 0.16826923076923073, "n": 208}
+{"dataset": "amazon_cf", "seed": 3, "flip_rate": 0.09875, "n": 800}
+{"dataset": "sst2", "seed": 4, "flip_rate": 0.12124999999999997, "n": 800}
+{"dataset": "imdb", "seed": 4, "flip_rate": 0.36250000000000004, "n": 800}
+{"dataset": "ag_news", "seed": 4, "flip_rate": 0.17500000000000004, "n": 800}
+{"dataset": "dbpedia", "seed": 4, "flip_rate": 0.48624999999999996, "n": 800}
+{"dataset": "counterfact", "seed": 4, "flip_rate": 0.03874999999999995, "n": 800}
+{"dataset": "emotion", "seed": 4, "flip_rate": 0.07250000000000001, "n": 800}
+{"dataset": "tweet_hate", "seed": 4, "flip_rate": 0.24750000000000005, "n": 800}
+{"dataset": "tweet_irony", "seed": 4, "flip_rate": 0.24624999999999997, "n": 800}
+{"dataset": "tweet_offensive", "seed": 4, "flip_rate": 0.23624999999999996, "n": 800}
+{"dataset": "subj", "seed": 4, "flip_rate": 0.08125000000000004, "n": 800}
+{"dataset": "spam", "seed": 4, "flip_rate": 0.24124999999999996, "n": 800}
+{"dataset": "cola", "seed": 4, "flip_rate": 0.11499999999999999, "n": 800}
+{"dataset": "stance", "seed": 4, "flip_rate": 0.1875, "n": 208}
+{"dataset": "amazon_cf", "seed": 4, "flip_rate": 0.13, "n": 800}
+{"dataset": "sst2", "seed": 0, "flip_rate": 0.09750000000000003, "n": 800}
+{"dataset": "ag_news", "seed": 0, "flip_rate": 0.1875, "n": 800}
+{"dataset": "counterfact", "seed": 0, "flip_rate": 0.043749999999999956, "n": 800}
+{"dataset": "emotion", "seed": 0, "flip_rate": 0.06125000000000003, "n": 800}
+{"dataset": "subj", "seed": 0, "flip_rate": 0.08875, "n": 800}
+{"dataset": "dbpedia", "seed": 0, "flip_rate": 0.49875, "n": 800}
+{"dataset": "tweet_hate", "seed": 0, "flip_rate": 0.26375000000000004, "n": 800}