ntc-ai/evidence-dial-benchmark

Evidence Dial — complete retained benchmark

Transfer gain, policy regression. The corrected Jev-curated particle adapter beats LoRA on the repeated FaithEval subset and loses on fresh balanced policy confirmation. All 17,868 native case-arm measurements over 2,772 cases are retained. The original answer-hinted release is preserved byte-for-byte under original-answer-hinted-v1; the complete corrected comparison is under curated-retry-v1.

Problem statement: Track 1, problem 3 — answer from organizational documents, return UNKNOWN when evidence is absent, and CONFLICT when equally authoritative records disagree. Every arm uses the same strong prompt, native greedy decoding, EOS3 and64 output tokens. Inputs over2,048 tokens are excluded before sampling; evidence is never truncated. Strict success requires an exact supported alias, UNKNOWN or CONFLICT. Multiple choice accepts exactly the correct letter or option text. Format differences count.

Corrected particle versus LoRA trained on the same48 cases

Cohort	OFF	Particle1	Particle2	LoRA	Particle1 − LoRA, pp ·95% interval
Fresh confirmation768	354	209	192	477	−34.90 [−38.41,−31.38]
Original policies256 · repeated	169	69	64	254	−72.27 [−75.39,−68.75]
FaithEval900 · repeated	323	501	468	270	+25.67 [22.69,28.67]
Cue audit80 · repeated	53	16	16	48	−40.00 [−40.00,−40.00]
Balanced stress768 · repeated	353	218	193	479	−33.98 [−37.24,−30.73]

Strength1 was frozen on the128 original development cases before any new test output. Development was negative:39/128 at1 and32/128 at2 versusOFF86/128 and matched LoRA128/128. It is the better of the two declared nominal particle strengths, not a positive development result or a test-based choice. No new rate, horizon, seed search or SSE tiebreak was added after results.

On FaithEval, selected particle scores217/300 counterfactual,278/300 missing and6/300 inconsistent; matched LoRA scores159/104/7. Original64-case LoRA scores 180/166/29,375/900 overall. Particle beats that original control by14.00pp [11.12,16.85] overall. These are repeated research comparisons on a fixed 900-case subset, not pristine confirmation or an official leaderboard score. Verified matched answerable QA controls were unavailable.

The strict MC advantage is largely formatting. A post hoc diagnostic accepts an unambiguous correct label followed by its exact correct option text, keeping QA scores unchanged. It restores3 particle,55 matched-LoRA and41 original-LoRA answers. MC counts become220/214/221 of300, with paired difference intervals spanning zero. Overall counts become504/325/416: particle still gains19.89pp [17.18,22.68] over matched LoRA and9.78 [7.14,12.46] over original LoRA. The primary protocol/selection is unchanged; this diagnostic does not establish better semantic choice accuracy. Its exact rule, four positive/negative fixtures, restored IDs and all source prediction hashes are in diagnostics/.

On fresh policies the selected particle emits literal UNKNOWN678/768 times and answers only17/384 answerable/altered cases correctly, versus LoRA361/384. All76 gains over LoRA are missing cases;344 others are losses. A constant UNKNOWN policy scores25% on the balanced corpus. Both corrected adapters score 0/192 conflicts, versusOFF6/192. The transfer benefit cannot be generalized to reliable open-ended organizational grounding.

Files and reproduction

curated-retry-v1/results.json includes every strict rate, paired gain/loss, 10,000-replicate family/source-question bootstrap interval, per-type diagnostic, all FaithEval strata, QA substring/MC option-text diagnostics, native timing, document-count effects and order invariance. predictions/ retains native outputs, output token IDs, prompt hashes and measured timing/token counts. paired-cases.jsonl supports comparison of all original and corrected arms.

Arm keys particle_1/2 and lora mean corrected particle and its48-case control. hinted_particle_1/2 and lora_original mean original hinted particle and64-case LoRA. Original particle2 was not planned for fresh confirmation or balanced stress and stays absent there. Missing measurements are never filled with zero. Original G0.012 stopped32 for qualified repetition; G0.004 completed128 but stays blocked for uncertain mathematical-notation observer scope. Their states remain in the training archive; neither is promoted by this release.

All17,868 strict outcomes, intervals, strata, subgroups and timing summaries were independently recomputed from retained responses. Every actual native input sequence was checked privately against matched OFF; teacher flags are false, inputs fit2,048 tokens and outputs end at native EOS3 or the64-token cap. Public audits contain hashes/counts, excluding source input token arrays and AuditKIT fields that could reconstruct FaithEval passages. FaithEval paired rows link to pinned upstream IDs/revisions without republishing questions/passages. Exact materialization and score commands require no new generation or judge call.

diagnostics/ contains four-case native equivalence proofs for both distribution copies, one exact saved-seed token/hidden-state replay, original/corrected cache output-head diagnostics, native-response SSE, actual exposure/loss journals and cost coverage. SSE is diagnostic only and worsens on this retry; it never selects a model. Model/critic/bank weight EMA is disabled and current RAW values are used. No completed update, teacher, proof or saved output was replayed to repair a resource refusal. The LoRA's post-training floor refusal and dev-only recovery are retained without changing its128 completed updates or8 saved answers.

Intervals quantify case/family variation for one training seed. The cue-audit interval is degenerate because each of its16 families has the same difference across five replicated patterns; it does not establish zero uncertainty beyond those templates. Stratum/format comparisons are descriptive and unadjusted. The original100% LoRA score shares all16 question frames and a DocumentB/conflict shortcut with training; its raw exact matches were audited, but the harder corpora expose that ceiling's limited scope.

Resource reports account for about7.00 observed/estimated GPU lease-hours across the whole pipeline, with explicit legacy gaps. One original50.98s diagnostic omitted CUDA mapping; its physical device is unverified, with no effect-score or selection impact. Shared-GPU timing includes contention and excludes cold load in per-query generation figures. Any dollar figure assumes an illustrative $1/GPU-hour, not paid pricing. Jev's reported estimated_cost0 is not verified zero billing. No closed answer-model API comparison is fabricated.

Corrected particle · Matched LoRA · Saved-output explorer · Curated samples/states · Frozen stronger tests · FULL/RAW archives

Licences

Own fictional source evidence is CC0; project code is MIT. Tiny Aya and derived weights/native-state tensors are CC-BY-NC-4.0 for noncommercial use. FaithEval source data retain their upstream terms; full source passages are not bundled. Kit implementations retain their original licences. No gated base weights are included. The Cohere2 binding, corpus, curation and evaluation were built for the hackathon; Particle Sliders core existed beforehand.