outputs/threads/engine_task.md

You are the autonomous experimental execution thread for Identifiable Latent Rubrics. The user explicitly requested parallel Codex threads under tmux so work survives terminal disconnection. Continue until all feasible real-data experiments below have actually run and been inspected. Do not merely plan. Do not ask ordinary engineering confirmations. Do not fabricate results. Do not spawn further agents or Codex sessions.

Workspace /home/ningyuxi/workspace/latent-rubric; Python .venv/bin/python. Existing default Codex account is ChatGPT; do not copy/change credentials or CODEX_HOME. Read README, reports/progress.md, reports/deviations.md, reports/metric_definitions.md, reports/api_budget.md, configs, and relevant code. Root is simultaneously coordinating and writing final reports; other tmux threads only audit/statistics.

OWNERSHIP: You exclusively own real experiment execution and may fix scripts/train_heads.py, scripts/evaluate_heads.py, scripts/run_ablations.py, models/, losses/, intervention/, representations/, data/, configs/ as needed. Do not modify analysis/ or final_report.md/paper_results.md; root owns them. Write progress to reports/engine_status.md, decisions to reports/engine_decisions.md, final handoff reports/engine_handoff.md, and outputs/ENGINE_DONE.json only when work actually finished. Do NOT git commit/reset/checkout or modify other workspaces. Root will commit reviewed changes. Write urgent findings for root to outputs/threads/engine_message.md. Do not wait for root to resolve routine fixes within your ownership.

HARD DEVICE SCOPE: physical GPU1 only, UUID GPU-2a9c799b-ab9e-c677-db0e-379f480cb00f. Do not operate other GPUs or services. Current full Qwen3-32B extraction is running under latent-rubric-extract32.service and is almost finished. Do not interrupt or duplicate it. Existing scripts/gpu_job.sh releases only our latent-rubric-gpu1-reserve.service, checks no unexpected GPU1 processes, executes a job and restores a 110 GiB idle memory reservation. XDG_RUNTIME_DIR=/run/user/1006 DBUS_SESSION_BUS_ADDRESS=unix:path=/run/user/1006/bus. Keep the reservation when GPU1 not needed. Lightweight head training may coexist with this reservation if sufficient memory; large model loading must use gpu_job.sh. Qwen3-32B and Qwen3-14B exact pinned dense weights already downloaded, unquantized BF16. Both original-response tokenizer caches exist. Frozen 32B features cache outputs/features/32b has 21,360 rows, layers [16,32,48,56,64], last and mean FP16. Require COMPLETE.json before training and verify service exit; cache resumes, never rerun without reason. H200 allocation limit125GiB.

COMPLETED: 67 controlled synthetic runs, figures, reports. Real HelpSteer2 schema verified and data prepared; train/dev/test prompt-disjoint. 6111/654/352 non-tie preference pairs, 18378/1944/1038 responses. Human attributes separate attributes.parquet. Baseline/ILR fitting must NEVER read human fine-grained attributes except explicit dense upper bound; never use TEST labels to train or tune. Oracle axis naming uses DEV only. All models save source snapshots, config, checkpoints, predictions and metrics. Main representation chosen in advance: 32B block48 response-mean. Four named semantic axes correctness/coherence/complexity/verbosity + one residual. Helpfulness never atomic. Linear main; MLP robustness. Statistical tests and final report owned by root.

EXECUTE SEQUENTIALLY BY DEPENDENCY:

1. After 32B complete, run one baseline seed and inspect train/dev losses (dev only for decisions), debug if necessary, then five seeds each Scalar-BT, Latent-BT, Ortho, dense Attribute+BT. Existing scripts/train_baselines.sh and train_heads.py. Evaluate via scripts/evaluate_heads.py, inspect actual baseline geometry. Do not infer useful rank-four real subspace solely from one global reward; record negative results.
2. Paid intervention pipeline is implemented but NO research edits have been generated yet. User specifically changed verifier to DeepSeek-V4.1-Flash; DO NOT USE DeepSeek-V4-Pro. Generator GLM-5.3, verifier target-blinded DeepSeek-V4.1-Flash. Both endpoints smoke tested. API code accesses existing user-owned mode600 credential reference; never print/copy secrets. Project budget config caps conservative planning-cost200CNY,6000 attempts,6M completion tokens. Provider actual rates unverified; record actual usage, don't claim invoice costs. Pre-ledger smokes listed in reports. Respect cap. Reuse cached edits/verification for every seed/ablation/backbone.
3. Begin with --bases 5 (40 possible edits), verify with intervention.verify, inspect actual outputs and criterion/filter quality. Then expand --bases50 and verify. Only if quality adequate expand toward --bases250 --heldout-bases50 (2400 maximum edits cumulatively), verify using 6-12 API workers sensibly. CLI supports --workers. Strict filter signed target>=1, off-target all0, medium/high confidence. Positive correctness/coherence may be infeasible. No human attribute selection of training edits. Pool prompts are disjoint by criterion and split; all originate TRAIN responses. Separate held-out intervention base ranks350..399 never train anchors. Document all deviations/quality failures rather than force target result.
4. Materialize final interventions.parquet before caching their vectors (expansion changes feature_index order). Tokenize and extract all feasible modified responses with same 32B layers/poolings. Use --input data/processed/interventions.parquet --output data/processed/tokenized_32b_interventions.parquet to representations.tokenize; its reports/token_lengths32b filenames currently overwrite originals, so fix output/report suffix to preserve original reports first. Extract --tokens that path --output outputs/features/32b_interventions under gpu_job.sh. Original features already cached; only modified rows new.
5. Train ILR for each feasible budget10,25,50,100,250 per criterion (nested seed-specific subsets); five seeds. MainN=max supported budget<=50, selected by availability, not test metrics. If fewer10 strict in some criteria, improve controllability based TRAIN intervention quality or perform lower-budget clearly documented experiments; don't simply stop. Required ablations: raw/direction/strict at matched N; four axes vs four+residual; linear vs specified MLP; all five layers and both poolings; shuffled preference and flipped intervention controls. scripts/run_ablations.py implements most and saves main_design.json but inspect before use. Important models>=5seeds. If complexity clearly fails, implement/evaluate reduced correctness/coherence/verbosity model and report complexity failure explicitly. Root can evaluate interpretation jointly; do not hide it.
6. Essential14B replication after32B core results: frozen same dataset/features, at least Latent-BT and ILR five seeds. Model weights/tokenizer caches already available. Extract original14B and intervention14B features on GPU1 only, then train and evaluate; share identical interventions, no new API costs. Layer/pooling follow main work, can use all cached representations cheaply. Optional LoRA/SAE are not required and must not block core study.
7. Evaluate actual preference, attribute correlations/purity, cross-seed axes/subspaces, held-out intervention target/leakage/residual distributions, TEST human natural quasi-pairs, and64 unseen rubric combinations (same-prompt rank accuracy, tau/rho) using scripts/evaluate_heads.py. Check all outputs actually complete. Write detailed engine handoff with run IDs, performed/missing experiment inventory, negative results, code fixes, resource state, API usage, and mainN. Root owns final tables/bootstrap/figures/research report, so leave all raw arrays for them. At end restore GPU1 reservation and verify nvidia-smi/service status. Mark ENGINE_DONE only when all feasible above is completed, not because of time/token pressure.

You may monitor dependencies with short sleeps; this tmux session is durable. Update engine_status every substantial stage. Do not prematurely terminate after launching a background process; monitor, debug, inspect outputs, continue. Existing synthetic results mean this is an active research project, not scaffolding.