auto-sync 2026-07-02T15:18:17Z workspace (part 4)

workspace/results/paper_analysis.json

@@ -1,6 +1,6 @@
 {
   "best_clean_key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005",
-  "generated_utc": "2026-07-02T14:59:38+00:00",
+  "generated_utc": "2026-07-02T15:05:57+00:00",
   "mechanism_gap": {
     "best_clean_vs_direct_same_ckpt": 0.1060869565217391,
     "best_clean_vs_h16": 0.09159420289855075,
@@ -12656,6 +12656,34 @@
       "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_summary.json",
       "std_success": 0.018595959341849776
     },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r0025_t0025": {
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "label": "K6-matched transported residual field re-grounding, small train-oracle selector calibration",
+      "missing": true,
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_oraclecal_train800_k4_r0025_t0025_summary.json"
+    },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r005_t005": {
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "label": "K6-matched transported residual field re-grounding, train-oracle calibrated selector",
+      "missing": true,
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_oraclecal_train800_k4_r005_t005_summary.json"
+    },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_rank005": {
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "label": "K6-matched transported residual field re-grounding, train-oracle rank calibration",
+      "missing": true,
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_oraclecal_train800_k4_rank005_summary.json"
+    },
+    "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_type005": {
+      "headline_metric": "mean_success",
+      "headline_metric_label": "deployment",
+      "label": "K6-matched transported residual field re-grounding, train-oracle type calibration",
+      "missing": true,
+      "source": "results/h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_oraclecal_train800_k4_type005_summary.json"
+    },
     "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_typesuccess001": {
       "ci95_success": 0.04240637681159422,
       "headline_ci95_success": 0.04240637681159422,

workspace/results/paper_analysis.md

@@ -1,6 +1,6 @@
 # Paper Analysis
 
-Generated: `2026-07-02T14:59:38+00:00`
+Generated: `2026-07-02T15:05:57+00:00`
 
 ## Main Seed Statistics
 
@@ -72,6 +72,10 @@ Generated: `2026-07-02T14:59:38+00:00`
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_taskrel | K6-matched transported residual field re-grounding, source-score 0.01, task-relative retrieval chart | 3 | deployment | 37.22% +/- 1.66 | 37.22% | +/- 4.12 | 58.49% | 0.512 | +7.48 pp |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_taskrelz | K6-matched transported residual field re-grounding, source-score 0.01, task-relative z-score retrieval chart | 3 | deployment | 36.99% +/- 2.02 | 36.99% | +/- 5.01 | 58.49% | 0.511 | +7.25 pp |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005 | K6-matched transported residual field re-grounding, task-conditioned source-score prior | 3 | deployment | 38.90% +/- 1.86 | 38.90% | +/- 4.62 | 59.96% | 0.513 | +9.16 pp |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r005_t005 | K6-matched transported residual field re-grounding, train-oracle calibrated selector | 0 | deployment | missing | missing | missing | missing | missing | missing |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r0025_t0025 | K6-matched transported residual field re-grounding, small train-oracle selector calibration | 0 | deployment | missing | missing | missing | missing | missing | missing |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_rank005 | K6-matched transported residual field re-grounding, train-oracle rank calibration | 0 | deployment | missing | missing | missing | missing | missing | missing |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_type005 | K6-matched transported residual field re-grounding, train-oracle type calibration | 0 | deployment | missing | missing | missing | missing | missing | missing |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_oraclek8 | K6-matched transported residual field re-grounding, source-score 0.01 candidate-oracle K8 | 3 | candidate-oracle | 44.35% +/- 1.98 | 38.84% | +/- 4.91 | 59.95% | 0.512 | +14.61 pp |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore0005 | K6-matched transported residual field re-grounding, exact drop-mask train source-score prior 0.005 | 3 | deployment | 38.72% +/- 1.71 | 38.72% | +/- 4.24 | 59.90% | 0.511 | +8.99 pp |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore0015 | K6-matched transported residual field re-grounding, exact drop-mask train source-score prior 0.015 | 3 | deployment | 38.78% +/- 1.82 | 38.78% | +/- 4.51 | 59.94% | 0.511 | +9.04 pp |

workspace/results/paper_table_status.json

@@ -1771,6 +1771,82 @@
       "best_config": null,
       "gain_vs_h16_policy": 0.09159420289855075
     },
+    {
+      "key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r005_t005",
+      "label": "K6-matched transported residual field re-grounding, train-oracle calibrated selector",
+      "path": "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_oraclecal_train800_k4_r005_t005_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "selector calibration learned from train-split candidate-oracle rollouts and evaluated on held-out validation",
+      "fallback_success": null,
+      "pending_job": "",
+      "path_exists": false,
+      "status": "pending",
+      "success": null,
+      "std_success": null,
+      "completed_seeds": null,
+      "num_completed": null,
+      "best_config": null,
+      "gain_vs_h16_policy": null
+    },
+    {
+      "key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r0025_t0025",
+      "label": "K6-matched transported residual field re-grounding, small train-oracle selector calibration",
+      "path": "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_oraclecal_train800_k4_r0025_t0025_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "smaller train-oracle rank/type selector calibration learned without validation outcomes",
+      "fallback_success": null,
+      "pending_job": "",
+      "path_exists": false,
+      "status": "pending",
+      "success": null,
+      "std_success": null,
+      "completed_seeds": null,
+      "num_completed": null,
+      "best_config": null,
+      "gain_vs_h16_policy": null
+    },
+    {
+      "key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_rank005",
+      "label": "K6-matched transported residual field re-grounding, train-oracle rank calibration",
+      "path": "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_oraclecal_train800_k4_rank005_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "rank-only selector calibration learned from train-split candidate-oracle rollouts",
+      "fallback_success": null,
+      "pending_job": "",
+      "path_exists": false,
+      "status": "pending",
+      "success": null,
+      "std_success": null,
+      "completed_seeds": null,
+      "num_completed": null,
+      "best_config": null,
+      "gain_vs_h16_policy": null
+    },
+    {
+      "key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_type005",
+      "label": "K6-matched transported residual field re-grounding, train-oracle type calibration",
+      "path": "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_oraclecal_train800_k4_type005_summary.json",
+      "clean_deployment": "yes",
+      "same_state_proposals": "no",
+      "expert_proposal": "no",
+      "story_role": "type-only selector calibration learned from train-split candidate-oracle rollouts",
+      "fallback_success": null,
+      "pending_job": "",
+      "path_exists": false,
+      "status": "pending",
+      "success": null,
+      "std_success": null,
+      "completed_seeds": null,
+      "num_completed": null,
+      "best_config": null,
+      "gain_vs_h16_policy": null
+    },
     {
       "key": "transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_oraclek8",
       "label": "K6-matched transported residual field re-grounding, source-score 0.01 candidate-oracle K8",

workspace/results/paper_table_status.md

@@ -96,6 +96,10 @@ Baseline h=16 policy: 29.74%
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_taskrel | K6-matched transported residual field re-grounding, source-score 0.01, task-relative retrieval chart | complete | 37.22% | +7.48 pp | yes | no | no | object-relative local chart ablation for transported residual support selection |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_metric_taskrelz | K6-matched transported residual field re-grounding, source-score 0.01, task-relative z-score retrieval chart | complete | 36.99% | +7.25 pp | yes | no | no | normalized object-relative local chart ablation for transported residual support selection |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005 | K6-matched transported residual field re-grounding, task-conditioned source-score prior | complete | 38.90% | +9.16 pp | yes | no | no | task-conditioned train-source utility prior chosen from per-task source-score sensitivity |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r005_t005 | K6-matched transported residual field re-grounding, train-oracle calibrated selector | pending | pending | pending | yes | no | no | selector calibration learned from train-split candidate-oracle rollouts and evaluated on held-out validation |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r0025_t0025 | K6-matched transported residual field re-grounding, small train-oracle selector calibration | pending | pending | pending | yes | no | no | smaller train-oracle rank/type selector calibration learned without validation outcomes |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_rank005 | K6-matched transported residual field re-grounding, train-oracle rank calibration | pending | pending | pending | yes | no | no | rank-only selector calibration learned from train-split candidate-oracle rollouts |
+| transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_type005 | K6-matched transported residual field re-grounding, train-oracle type calibration | pending | pending | pending | yes | no | no | type-only selector calibration learned from train-split candidate-oracle rollouts |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_oraclek8 | K6-matched transported residual field re-grounding, source-score 0.01 candidate-oracle K8 | complete | 44.35% | +14.61 pp | diagnostic | no | no | diagnostic top-8 proposal-oracle ceiling for the current best source-score-calibrated selector |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore0005 | K6-matched transported residual field re-grounding, exact drop-mask train source-score prior 0.005 | complete | 38.72% | +8.99 pp | yes | no | no | very small train-source utility prior sensitivity for selector calibration |
 | transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore0015 | K6-matched transported residual field re-grounding, exact drop-mask train source-score prior 0.015 | complete | 38.78% | +9.04 pp | yes | no | no | midpoint train-source utility prior sensitivity for selector calibration |

workspace/scripts/build_oracle_selector_calibration.py

@@ -0,0 +1,294 @@
+#!/usr/bin/env python
+from __future__ import annotations
+
+import argparse
+import glob
+import json
+from collections import defaultdict
+from pathlib import Path
+from typing import Any
+
+
+def _mean(values: list[float]) -> float | None:
+    if not values:
+        return None
+    return float(sum(values) / len(values))
+
+
+def _clip(value: float, limit: float | None) -> float:
+    if limit is None or limit <= 0:
+        return value
+    return max(-float(limit), min(float(limit), value))
+
+
+def _branch_values(row: dict[str, Any], objective: str) -> list[float]:
+    if objective == "score":
+        values = row.get("candidate_oracle_branch_scores")
+    elif objective == "progress":
+        values = row.get("candidate_oracle_branch_progress")
+    elif objective == "success":
+        values = [
+            1.0 if bool(value) else 0.0
+            for value in row.get("candidate_oracle_branch_successes", [])
+        ]
+    else:
+        raise ValueError("objective must be 'score', 'progress', or 'success'")
+    if not isinstance(values, list):
+        return []
+    parsed: list[float] = []
+    for value in values:
+        try:
+            parsed.append(float(value))
+        except (TypeError, ValueError):
+            parsed.append(0.0)
+    return parsed
+
+
+def _rank_biases(
+    values_by_rank: list[list[float]],
+    *,
+    scale: float,
+    min_count: int,
+    max_abs_bias: float | None,
+) -> tuple[list[float], list[float | None], list[int]]:
+    means = [_mean(values) for values in values_by_rank]
+    counts = [len(values) for values in values_by_rank]
+    baseline = means[0] if means and means[0] is not None and counts[0] >= min_count else None
+    biases: list[float] = []
+    for rank, mean in enumerate(means):
+        if rank == 0 or baseline is None or mean is None or counts[rank] < min_count:
+            biases.append(0.0)
+        else:
+            biases.append(_clip(float(scale) * (float(mean) - float(baseline)), max_abs_bias))
+    return biases, means, counts
+
+
+def _type_bonuses(
+    values_by_type: dict[str, list[float]],
+    *,
+    scale: float,
+    min_count: int,
+    max_abs_bonus: float | None,
+) -> tuple[dict[str, float], dict[str, float | None], dict[str, int], float | None]:
+    pooled = [value for values in values_by_type.values() for value in values]
+    baseline = _mean(pooled)
+    means = {
+        candidate_type: _mean(values)
+        for candidate_type, values in sorted(values_by_type.items())
+    }
+    counts = {
+        candidate_type: len(values)
+        for candidate_type, values in sorted(values_by_type.items())
+    }
+    if baseline is None or len(pooled) < min_count:
+        return {}, means, counts, baseline
+    bonuses: dict[str, float] = {}
+    for candidate_type, mean in means.items():
+        if mean is None or counts[candidate_type] < min_count:
+            continue
+        bonuses[candidate_type] = _clip(
+            float(scale) * (float(mean) - float(baseline)),
+            max_abs_bonus,
+        )
+    return bonuses, means, counts, baseline
+
+
+def _iter_rollout_paths(patterns: list[str]) -> list[Path]:
+    paths: list[Path] = []
+    for pattern in patterns:
+        matches = [Path(path) for path in glob.glob(pattern)]
+        paths.extend(matches or [Path(pattern)])
+    unique = sorted({path.resolve() for path in paths})
+    missing = [path for path in unique if not path.exists()]
+    if missing:
+        raise FileNotFoundError(f"Missing rollout file(s): {missing}")
+    return unique
+
+
+def build_oracle_selector_calibration(
+    rollout_paths: list[Path],
+    *,
+    objective: str,
+    max_rank: int,
+    rank_scale: float,
+    type_scale: float,
+    min_count: int,
+    max_abs_rank_bias: float | None,
+    max_abs_type_bonus: float | None,
+) -> dict[str, Any]:
+    if max_rank <= 0:
+        raise ValueError("max_rank must be positive")
+    if min_count <= 0:
+        raise ValueError("min_count must be positive")
+
+    rank_by_task: dict[str, list[list[float]]] = defaultdict(
+        lambda: [[] for _ in range(max_rank)]
+    )
+    type_by_task: dict[str, dict[str, list[float]]] = defaultdict(
+        lambda: defaultdict(list)
+    )
+    global_ranks: list[list[float]] = [[] for _ in range(max_rank)]
+    global_types: dict[str, list[float]] = defaultdict(list)
+    rows_seen = 0
+    rows_used = 0
+    skipped_branches = 0
+
+    for path in rollout_paths:
+        payload = json.loads(path.read_text())
+        for row in payload.get("rows", []):
+            if not isinstance(row, dict):
+                continue
+            rows_seen += 1
+            task_id = str(row.get("task_id") or "")
+            values = _branch_values(row, objective)
+            candidate_types = row.get("candidate_oracle_types")
+            valid_mask = row.get("candidate_oracle_valid_mask")
+            if not task_id or not values or not isinstance(candidate_types, list):
+                continue
+            if not isinstance(valid_mask, list):
+                valid_mask = [True] * len(values)
+            branch_count = min(max_rank, len(values), len(candidate_types), len(valid_mask))
+            if branch_count <= 0:
+                continue
+            rows_used += 1
+            for rank in range(branch_count):
+                if not bool(valid_mask[rank]):
+                    skipped_branches += 1
+                    continue
+                value = float(values[rank])
+                candidate_type = str(candidate_types[rank])
+                rank_by_task[task_id][rank].append(value)
+                global_ranks[rank].append(value)
+                type_by_task[task_id][candidate_type].append(value)
+                global_types[candidate_type].append(value)
+
+    field_rank_biases_by_task: dict[str, list[float]] = {}
+    rank_utility_means_by_task: dict[str, list[float | None]] = {}
+    rank_counts_by_task: dict[str, list[int]] = {}
+    for task_id in sorted(rank_by_task):
+        biases, means, counts = _rank_biases(
+            rank_by_task[task_id],
+            scale=rank_scale,
+            min_count=min_count,
+            max_abs_bias=max_abs_rank_bias,
+        )
+        field_rank_biases_by_task[task_id] = biases
+        rank_utility_means_by_task[task_id] = means
+        rank_counts_by_task[task_id] = counts
+    global_rank_biases, global_rank_means, global_rank_counts = _rank_biases(
+        global_ranks,
+        scale=rank_scale,
+        min_count=min_count,
+        max_abs_bias=max_abs_rank_bias,
+    )
+    field_rank_biases_by_task["*"] = global_rank_biases
+    rank_utility_means_by_task["*"] = global_rank_means
+    rank_counts_by_task["*"] = global_rank_counts
+
+    candidate_type_bonuses_by_task: dict[str, dict[str, float]] = {}
+    type_utility_means_by_task: dict[str, dict[str, float | None]] = {}
+    type_counts_by_task: dict[str, dict[str, int]] = {}
+    type_baselines_by_task: dict[str, float | None] = {}
+    for task_id in sorted(type_by_task):
+        bonuses, means, counts, baseline = _type_bonuses(
+            type_by_task[task_id],
+            scale=type_scale,
+            min_count=min_count,
+            max_abs_bonus=max_abs_type_bonus,
+        )
+        candidate_type_bonuses_by_task[task_id] = bonuses
+        type_utility_means_by_task[task_id] = means
+        type_counts_by_task[task_id] = counts
+        type_baselines_by_task[task_id] = baseline
+    global_type_bonuses, global_type_means, global_type_counts, global_type_baseline = (
+        _type_bonuses(
+            global_types,
+            scale=type_scale,
+            min_count=min_count,
+            max_abs_bonus=max_abs_type_bonus,
+        )
+    )
+    candidate_type_bonuses_by_task["*"] = global_type_bonuses
+    type_utility_means_by_task["*"] = global_type_means
+    type_counts_by_task["*"] = global_type_counts
+    type_baselines_by_task["*"] = global_type_baseline
+
+    return {
+        "source_rollouts": [str(path) for path in rollout_paths],
+        "calibration_source": "candidate_oracle_rollout",
+        "objective": objective,
+        "max_rank": int(max_rank),
+        "rank_scale": float(rank_scale),
+        "type_scale": float(type_scale),
+        "min_count": int(min_count),
+        "max_abs_rank_bias": max_abs_rank_bias,
+        "max_abs_type_bonus": max_abs_type_bonus,
+        "rows_seen": rows_seen,
+        "rows_used": rows_used,
+        "skipped_branches": skipped_branches,
+        "field_rank_biases_by_task": field_rank_biases_by_task,
+        "rank_utility_means_by_task": rank_utility_means_by_task,
+        "rank_counts_by_task": rank_counts_by_task,
+        "candidate_type_bonuses_by_task": candidate_type_bonuses_by_task,
+        "type_utility_means_by_task": type_utility_means_by_task,
+        "type_counts_by_task": type_counts_by_task,
+        "type_baselines_by_task": type_baselines_by_task,
+    }
+
+
+def main(argv: list[str] | None = None) -> int:
+    parser = argparse.ArgumentParser(
+        description=(
+            "Build train-split selector calibration from candidate-oracle rollout traces."
+        )
+    )
+    parser.add_argument("--rollout", action="append", required=True)
+    parser.add_argument("--out", type=Path, required=True)
+    parser.add_argument("--objective", choices=("score", "progress", "success"), default="score")
+    parser.add_argument("--max-rank", type=int, default=4)
+    parser.add_argument("--rank-scale", type=float, default=0.05)
+    parser.add_argument("--type-scale", type=float, default=0.05)
+    parser.add_argument("--min-count", type=int, default=20)
+    parser.add_argument("--max-abs-rank-bias", type=float, default=0.02)
+    parser.add_argument("--max-abs-type-bonus", type=float, default=0.02)
+    args = parser.parse_args(argv)
+
+    max_abs_rank_bias = args.max_abs_rank_bias if args.max_abs_rank_bias > 0 else None
+    max_abs_type_bonus = args.max_abs_type_bonus if args.max_abs_type_bonus > 0 else None
+    rollout_paths = _iter_rollout_paths(args.rollout)
+    result = build_oracle_selector_calibration(
+        rollout_paths,
+        objective=args.objective,
+        max_rank=args.max_rank,
+        rank_scale=args.rank_scale,
+        type_scale=args.type_scale,
+        min_count=args.min_count,
+        max_abs_rank_bias=max_abs_rank_bias,
+        max_abs_type_bonus=max_abs_type_bonus,
+    )
+    args.out.parent.mkdir(parents=True, exist_ok=True)
+    args.out.write_text(json.dumps(result, indent=2) + "\n")
+    print(
+        json.dumps(
+            {
+                key: value
+                for key, value in result.items()
+                if key
+                not in {
+                    "field_rank_biases_by_task",
+                    "rank_utility_means_by_task",
+                    "rank_counts_by_task",
+                    "candidate_type_bonuses_by_task",
+                    "type_utility_means_by_task",
+                    "type_counts_by_task",
+                    "type_baselines_by_task",
+                }
+            },
+            indent=2,
+        )
+    )
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

workspace/scripts/build_paper_analysis.py

@@ -561,6 +561,42 @@ METHODS = [
             "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_summary.json"
         ),
     ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r005_t005",
+        label="K6-matched transported residual field re-grounding, train-oracle calibrated selector",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_"
+            "oraclecal_train800_k4_r005_t005_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r0025_t0025",
+        label="K6-matched transported residual field re-grounding, small train-oracle selector calibration",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_"
+            "oraclecal_train800_k4_r0025_t0025_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_rank005",
+        label="K6-matched transported residual field re-grounding, train-oracle rank calibration",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_"
+            "oraclecal_train800_k4_rank005_summary.json"
+        ),
+    ),
+    MethodSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_type005",
+        label="K6-matched transported residual field re-grounding, train-oracle type calibration",
+        summary_path=(
+            "h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_"
+            "besttransport_margin0p00_k6_srcscore_task_pick001_stack005_"
+            "oraclecal_train800_k4_type005_summary.json"
+        ),
+    ),
     MethodSpec(
         key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_oraclek8",
         label="K6-matched transported residual field re-grounding, source-score 0.01 candidate-oracle K8",

workspace/scripts/build_paper_table_status.py

@@ -927,6 +927,42 @@ SPECS = [
         expert_proposal="no",
         story_role="task-conditioned train-source utility prior chosen from per-task source-score sensitivity",
     ),
+    ResultSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r005_t005",
+        label="K6-matched transported residual field re-grounding, train-oracle calibrated selector",
+        path="h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_oraclecal_train800_k4_r005_t005_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="selector calibration learned from train-split candidate-oracle rollouts and evaluated on held-out validation",
+    ),
+    ResultSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_r0025_t0025",
+        label="K6-matched transported residual field re-grounding, small train-oracle selector calibration",
+        path="h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_oraclecal_train800_k4_r0025_t0025_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="smaller train-oracle rank/type selector calibration learned without validation outcomes",
+    ),
+    ResultSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_rank005",
+        label="K6-matched transported residual field re-grounding, train-oracle rank calibration",
+        path="h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_oraclecal_train800_k4_rank005_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="rank-only selector calibration learned from train-split candidate-oracle rollouts",
+    ),
+    ResultSpec(
+        key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore_task_pick001_stack005_oraclecal_train800_k4_type005",
+        label="K6-matched transported residual field re-grounding, train-oracle type calibration",
+        path="h16_transport_field_reground_fieldonly_k6clean_dropnoopwg_b12_v1_besttransport_margin0p00_k6_srcscore_task_pick001_stack005_oraclecal_train800_k4_type005_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="type-only selector calibration learned from train-split candidate-oracle rollouts",
+    ),
     ResultSpec(
         key="transport_field_reground_fieldonly_k6matched_b12_clean_k6_dropnoopwg_retargeted_srcscore001_oraclek8",
         label="K6-matched transported residual field re-grounding, source-score 0.01 candidate-oracle K8",

workspace/scripts/slurm/build_oracle_selector_calibration.sbatch

@@ -0,0 +1,43 @@
+#!/bin/bash
+#SBATCH --job-name=build_oracle_cal
+#SBATCH --account=def-yalda
+#SBATCH --time=00:10:00
+#SBATCH --cpus-per-task=1
+#SBATCH --mem=2G
+#SBATCH --array=0-2
+#SBATCH --output=outputs/hpc/logs/%x_%A_%a.out
+#SBATCH --error=outputs/hpc/logs/%x_%A_%a.err
+
+set -euo pipefail
+
+PROJECT_DIR="${PROJECT_DIR:-$SLURM_SUBMIT_DIR}"
+RUN_ROOT="${RUN_ROOT:-/scratch/$USER/dovla/experiments/dovla_h16_policy_ckpt_runs}"
+OBJECTIVE="${OBJECTIVE:?Set OBJECTIVE to the seed-indexed run objective}"
+SEED="${SLURM_ARRAY_TASK_ID:-0}"
+ROLLOUT_NAME="${ROLLOUT_NAME:?Set ROLLOUT_NAME to the train oracle rollout JSON}"
+OUT_NAME="${OUT_NAME:-oracle_selector_calibration.json}"
+CAL_OBJECTIVE="${CAL_OBJECTIVE:-score}"
+MAX_RANK="${MAX_RANK:-4}"
+RANK_SCALE="${RANK_SCALE:-0.05}"
+TYPE_SCALE="${TYPE_SCALE:-0.05}"
+MIN_COUNT="${MIN_COUNT:-20}"
+MAX_ABS_RANK_BIAS="${MAX_ABS_RANK_BIAS:-0.02}"
+MAX_ABS_TYPE_BONUS="${MAX_ABS_TYPE_BONUS:-0.02}"
+PYTHON="${PYTHON:-python3}"
+
+ROLLOUT="$RUN_ROOT/$OBJECTIVE/seed_$SEED/$ROLLOUT_NAME"
+OUT="$RUN_ROOT/$OBJECTIVE/seed_$SEED/$OUT_NAME"
+
+cd "$PROJECT_DIR"
+mkdir -p outputs/hpc/logs "$(dirname "$OUT")"
+
+"$PYTHON" scripts/build_oracle_selector_calibration.py \
+  --rollout "$ROLLOUT" \
+  --out "$OUT" \
+  --objective "$CAL_OBJECTIVE" \
+  --max-rank "$MAX_RANK" \
+  --rank-scale "$RANK_SCALE" \
+  --type-scale "$TYPE_SCALE" \
+  --min-count "$MIN_COUNT" \
+  --max-abs-rank-bias "$MAX_ABS_RANK_BIAS" \
+  --max-abs-type-bonus "$MAX_ABS_TYPE_BONUS"