ntc-ai/evidence-dial-particle

Evidence Dial — Jev-curated RAW128 particle adapter

Transfer gain with a policy regression. On the repeated 900-case FaithEval subset, this adapter scores 501/900 (55.7%) versus LoRA trained on the same 48 cases, 270/900 (30.0%), and original 64-case LoRA, 375/900 (41.7%). The paired gain is 25.67 points [22.69, 28.67] over the matched control, and 14.00 [11.12, 16.85] over original LoRA. On fresh balanced policies it loses: 209/768 (27.2%) versus matched LoRA 477/768 (62.1%) and OFF 354/768 (46.1%), with particle minus LoRA −34.90 points [−38.41, −31.38].

On FaithEval, missing-evidence successes are 278/300 versus matched LoRA 104/300 and original LoRA 166/300. Strict counterfactual MC scores are 217/300 versus 159/300 and 180/300; inconsistent evidence remains weak at 6/300 versus 7/300 and 29/300. All strata and regressions are reported.

Formatting accounts for much of the MC gap. A post hoc diagnostic also accepts a correct letter followed by its exact correct option text, keeping QA scores unchanged. It gives MC counts 220/214/221 and overall counts 504/325/416. The overall advantage remains 19.89 points [17.18, 22.68] over matched LoRA and 9.78 [7.14, 12.46] over original LoRA. This supports improved abstention on this subset; it does not establish better semantic MC choice accuracy. The primary scorer and development selection remain unchanged. This is repeated research comparison, not pristine confirmation or an official leaderboard score.

Problem: answer from supplied organizational documents, return UNKNOWN for missing evidence and CONFLICT for incompatible equally authoritative evidence. The model mostly abstains: UNKNOWN on678/768 fresh cases at strength1. It answers only17/384 answerable/altered cases correctly, versus LoRA361/384. All76 gains over LoRA are missing-evidence cases;344 other cases are losses. A constant UNKNOWN policy scores25% on this balanced corpus. Missing-evidence accuracy alone does not establish better grounding.

The separate original answer-hinted adapter and every earlier result remain available; none is attributed to this method.

Frozen Tiny Aya produced explanations and final answers from evidence without reference-answer hints. At most16 fixed seeded attempts per each64 source case were allowed. Exact final-answer checking and actual NanoGPT Jev1.14 correctness≥.90/quality≥3 of4 retained the first approved attempt:48 cases (16 answerable,16 altered,9 missing,7 conflict), with16 excluded cases listed. All455 native attempts,65 Jev calls including8 fixtures, seed/RNG/token/hidden traces and118 paired positions are in the curation archive.

Neutral uses the unchanged evaluation prompt; positive adds the approved explanation/answer as verified context. Both use the same causal answer prefix. The sampled trajectory establishes provenance; the target is the separately cached positive-context native2048-value state. The shared Particle Sliders response API and official ParticleGAN0.9.5 game train the activation-routed 128×4 bank and rank8 q/k/v/o adapters across36 layers. There is no independently sampled language-model z. See the formulation.

Single fixed G.001/D.018/bank.03, batch2, seed20261110 and128 updates; RAW current weights with ema_decay=None. Scoped experimental fresh initialization preserves pretrained weights and zero up-projections. No saved state is initialized again. The adapter has4,766,912 parameters versus LoRA3,833,856; no footprint or speed advantage is claimed. Training samples118 answer positions, while the matched LoRA samples48 cases. Objectives, exposure units and preparation costs differ.

Development scored39/128 at1 and32/128 at2 versus OFF86; strength2 emitted UNKNOWN throughout development. Strength1 was frozen before confirmation by equal-case finished-answer ON−OFF success alone. No test reselection, SSE tiebreak or automatic rate/horizon extension occurred. The bounded catastrophic-output checks passed; that does not establish accuracy or absence of excess abstention.

Base: CohereLabs/tiny-aya-global, revision af89d219b53ed9b13b8a4645f9c8028973510324. Attribution: Cohere Labs. Derived weights are CC-BY-NC-4.0, for non-commercial use. Obtain gated base access through your own account. This public custom safetensors packet contains no base weights, teacher, critic or optimizer and cannot resume training.

Four retained native generations at strength1 reproduced exact input/output token IDs from this publication copy. Learned tensor values are unchanged; the proof is bundled as distribution-equivalence.json. This validates distribution equivalence, not model quality.

The complete corrected comparisons follow. Except for fresh confirmation, these are repeated research cohorts. Counts use the frozen strict scorer; intervals resample families or source questions.

Cohort	OFF	Particle1	Particle2	Same48 LoRA	Particle1 − LoRA, pp ·95% CI
Fresh confirmation768	354	209	192	477	-34.90 [-38.41, -31.38]
Original policies256	169	69	64	254	-72.27 [-75.39, -68.75]
FaithEval900	323	501	468	270	+25.67 [+22.69, +28.67]
Cue audit80	53	16	16	48	-40.00 [-40.00, -40.00]
Balanced stress768	353	218	193	479	-33.98 [-37.24, -30.73]

The cue-audit interval is degenerate because every family has the same paired difference under its five replicated patterns; it is not a claim of zero uncertainty beyond those templates. All stratum/format comparisons are descriptive and unadjusted.

Install the pinned project and core, then use the native protocol:

evidence-dial infer --adapter ntc-ai/evidence-dial-particle --row demo/example.json --strength 1 --out example-output.json --gpu 1

Greedy, EOS3,64 output tokens and2048 input tokens match evaluation. Overlength evidence is refused without truncation. Strength0 bypasses the adapter;2 uses the same RAW values at twice their contribution and can regress. This is a single-seed English research experiment, not a production reliability guarantee.

All benchmark results · Saved-output explorer · FULL/RAW checkpoints

CuratorKIT → Particle Sliders / AlignTune → AuditKIT is the actual pipeline. The Cohere2 binding, curation, corpus and evaluation are new hackathon work; Particle Sliders core is pre-existing. Upstream sources retain their licences.