MEASUREMENT DICTIONARY (LOCAL STUDY)
assignment/trial_id: exact original schedule ID. global_assignment_index: original frozen index. assignment_index: local phase index retained for source tracing. source_phase: immutable originating run folder.
correct: all external semantic requirement groups true on final installed plan in a COMPLETE trial. Interrupted trial primary outcome remains null/unknown; last durable plan is retained separately, not presented as normal final outcome.
partial_score: proportion of template-specific groups passing; descriptive, not calibrated across templates.
model_calls: successfully returned evaluated-model API records in completed trials. Partial turns retained for interrupted trials; interrupted request total is unknown and not added as measured calls/tokens.
tool_calls: trusted finite-interface operation counter, not a vendor tool API or host command count. Each plan install/proposal request counts one; requested actual visible verification counts one; eligible non-continue finish assessment counts one. Read-only proposals do not install plans.
actual_tokens: backend prompt_eval_count + eval_count across returned calls. Each INPUT count independently agrees with pinned tokenizer. OUTPUT counts are backend-reported and cap-checked, not independently reconstructed. Counts include cached prompt tokens as reported; they are not measured FLOPs, energy, uncached compute or monetary cost.
wall_seconds: monotonic elapsed trial time including API wait, loading if first call, trusted operations and grading. api_wall_seconds: sum of API wait durations. Timings are not an isolated-host performance benchmark.
actual_visible_check_on_final_plan: passing public structural check matches final artifact hash and version. This check omits semantic requirements and cannot certify correctness.
controller_accepted: eligible solved finish admitted by policy. Does not mean external correctness.
model_solved_claims: eligible solved claims; can exceed one/trial after rejected attempts. Nonfinal staged claims retained separately. wrong_solved_claims: these eligible plans fail external semantics. Zero claim denominator means undefined rate.
wrong_controller_acceptance: accepted solved finish whose final plan fails external semantics. Descriptive rates identify numerator, observed and assigned denominators.
stop_reason: successful/unsolved finish, call/budget/deadline stop; infrastructure stop lives in batch/incident records, not silently recoded as model failure.
trace_head/count: stored hash-chain anchor and length. Detects mismatch against retained anchors, not an independent timestamp or cryptographic proof of benign execution.
freeze_sha256: actual original local freeze bytes. Unchanged source hashes are checked at audit. Post-freeze audit/continuation amendments have separate files and disclosures.
semantic units: 120 prespecified parameter draws, 115 distinct evaluation payloads, six fixed templates. Three repeats and two related model variants do not create new semantic problems.
identification_bounds: lower/upper value of original paired contract estimand when unknown outcomes range over [0,1]; endpoints are sensitivity limits, not observed outcomes.
percentile_95_envelope: lower bootstrap percentile of the lower-effect endpoint to upper percentile of the upper-effect endpoint, paired resampling within fixed cells. Generator-conditional sensitivity only; no novel-template/model-population coverage or missing-at-random claim.
deduplicated_contract_descriptive: post-hoc unique-payload-weighted contrast within cells, cells equally weighted, incomplete groups omitted. Does not replace original primary effect.

Post-freeze measurement clarification: the frozen trial controller enforces an effective 16,192 aggregate input+output token cap, below the outer 16,384 protocol ceiling; prompts advertise 16,192. This reporting correction does not alter admission or outcomes. Separate seeds did not ensure semantic holdout: 6 evaluation recovery draws/5 payloads match preflight splits, retained with a post-hoc exclusion sensitivity.

Runtime-pin interpretation: top-level model_tag/weights fields preserve legacy 1.5B compatibility qualification. MAIN models are the two 0.5B tags in the frozen assignment schedule and models dictionary. Audits compare every request and returned model identity against that schedule, not the legacy default field.

Host instrumentation: original57-input freeze remains preserved. A separately pinned diagnostic observer and later bounded100ms Windows mutex around private heartbeat I/O were introduced after retained heartbeat failures. No scientific prompt, task, model setting, seed, grader or primary estimand changed. Monitoring timing can differ. Rejected native share-delete candidate was never admitted to model runs.
