BradJu's picture
Publish authorized non-adopted experiments; selected model unchanged
c72da75 verified
|
Raw History Blame Contribute Delete
2.69 kB

Negative-experiment appendix: tool trajectory learning and contract regression

Author: Chonghao Ju. Local release candidate; not published or approved for publication. Neither experiment replaces the released AlignTune ownership gate or its 23/30 shopping workflow. Paid iteration is stopped.

Experiment Base Ownership adapter Trajectory adapter
Original trajectory experiment 0/72 direct tools 34/72 with a deterministic controller 39/72 direct tools
Expanded contract, all direct tools 0/72 1/72 0/72

These rows use different prompts and, for ownership, different architectures. They are not one ranking. The 72 tasks are exposed synthetic cases, not a new independent test.

Offline reproduction

Python 3.9+; no packages, network, GPU, credentials or inference:

python3 -I -B replay.py

The script verifies hashes, 196 teacher trajectories, every executed tool call in both experiments, exact metrics and family-bootstrap intervals. It copies records to a temporary directory for re-scoring; packaged records are never overwritten. Expected output ends in APPENDIX_PASS. Offline replay recomputes scores from saved answers; it does not regenerate model predictions.

Contents

  • data/: 304 source-matched synthetic tasks (196 training, 36 development, 72 held-out/exposed), exact teacher calls, literal-value overrides and simulated replies.
  • recorded/distillation/: 216 evaluation trajectories, original local metrics, and training-mask receipt.
  • recorded/regression/: 234 trajectories (18 development and 216 regression), exact summary, paired changes and separately labeled post-hoc format diagnosis.
  • code/: historical environment, prompts, training loss checks and scoring; portable inference helpers.
  • adapter/: non-adopted trajectory LoRA, 30.7 MB; no base weights. Only the saved base-path metadata is normalized. Weight bytes unchanged.
  • config/, provenance.json: versions, settings, criteria and transformations.
  • INFERENCE.md: optional fresh-inference/training instructions. They require separate compute and are never invoked by replay.

No private profile/reserve/library texts, authentication files, billing/account receipts, cloud launchers, private Hub addresses or cloud logs are included. Source task text matches the already-public authored component dataset exactly. New model outputs, compiled traces, overrides and the trajectory adapter are still proposed exports awaiting user confirmation.

See REPORT.md for findings and limitations, LICENSE-SCOPE.md for the distinct code/data/model permissions, and FILES.csv for the exact proposed publication list.