refresh/rlhf-decoding-primitives

Task gallery / Machine learning / Version 1

Reference solution (oracle): Fail
No-op resistance check: Pass
QA verdict: Not recorded
Aggregate eligibility: Excluded / other version

Quality exclusion

Known task defect recorded
Reference solution fails

These per-version observations remain visible, but this version does not contribute to aggregate pass rates.

The task

Reimplement the post-training decoding-and-collation library in /app/objectives.py so that every function reproduces the sealed reference module reforacle exactly. reforacle is importable and callable (import reforacle) but its source cannot be read. It is the only ground truth; there is no written specification of what any function computes or of the conventions it follows. Recover its behavior by calling it. /app/…

Excerpt only; not a complete executable task specification. Task ID task_b6b7ff00f321461e · version 1. Selection status: not fully recorded.

Results by model and effort

This exact task version only. Every rate is passes / valid runs with a Wilson 95% interval; n≤2 remains counts only. Macro aggregation does not apply within a single task. Baselines are summarized in the quality gates, not ranked. A passed no-op resistance check means that doing nothing did not solve the task.

Current evidence

Model / effort	Verifier result [95% Wilson]	Diagnostic labels / valid	HARNESS_ERROR · excluded	Outcome classes
Claude Opus 5 · default (unverified)	3/3 · 100.0% [43.9%, 100.0%] · preliminary	3 / 3	0	GOOD_FAILURE: 0 BAD_FAILURE: 0 GOOD_SUCCESS: 3 BAD_SUCCESS: 0

Legacy evidence

Model / effort	Verifier result [95% Wilson]	Diagnostic labels / valid	HARNESS_ERROR · excluded	Outcome classes

Every recorded run

9 recorded trials, including controls. Diagnostic labels do not replace verifier reward. Labels and quality checks are source declarations, not independently attested here. Internal evidence remains private; a missing public URL is shown explicitly. Records with invalid or absent rewards are not scored.

Run / model	Date / source	Reward / eligibility	Diagnostic class / source	Resources	Evidence
`trial_9a523fffff844872` oracle@@	Oct 12, 2026 console	0 · excludedharness_error	HARNESS_ERROR frozen-classifier	Tokens: — Turns: — Cost: —	Not publicly available
`trial_e97b207f57b0422f` nop@@	Oct 12, 2026 console	0 · excludedharness_error	HARNESS_ERROR frozen-classifier	Tokens: — Turns: — Cost: —	Not publicly available
`trial_419e4b9c24474298` oracle@@	Oct 12, 2026 console	1 · valid	GOOD_SUCCESS frozen-classifier	Tokens: — Turns: — Cost: —	Not publicly available
`trial_ae2489fb59404cd8` nop@@	Oct 12, 2026 console	0 · valid	GOOD_FAILURE frozen-classifier	Tokens: — Turns: — Cost: —	Not publicly available
`trial_e071620114a44f6c` Claude Opus 5 · default (unverified)	Oct 12, 2026 console	1 · valid	GOOD_SUCCESS frozen-classifier	Tokens: 59,078 Turns: 57 Cost: $4.0970	Not publicly available
`trial_75c41d893655444b` Claude Opus 5 · default (unverified)	Oct 12, 2026 console	1 · valid	GOOD_SUCCESS frozen-classifier	Tokens: 45,035 Turns: 50 Cost: $3.4586	Not publicly available
`trial_a8a8cee79ba144ac` Claude Opus 5 · default (unverified)	Oct 12, 2026 console	No reward · excludedno_reward	No diagnostic label No diagnostic source	Tokens: — Turns: — Cost: —	Not publicly available
`trial_6d61dac56a364fe6` Claude Opus 5 · default (unverified)	Oct 12, 2026 console	1 · valid	GOOD_SUCCESS frozen-classifier	Tokens: 48,929 Turns: 49 Cost: $3.3467	Not publicly available
`trial_8612295f3e524556` Claude Opus 5 · default (unverified)	Oct 12, 2026 console	No reward · excludedno_reward	No diagnostic label No diagnostic source	Tokens: — Turns: — Cost: —	Not publicly available