ntc-ai/evidence-dial-grounding-tests

Evidence Dial: tests of evidence use beyond template cues

These fresh tests were frozen after our original synthetic benchmark gave an AlignTune-trained LoRA256/256. An independent token/input/scoring audit confirmed that score, but found that all16 question frames occurred in training and that DocumentB perfectly predicted CONFLICT across the original448 policy examples. The original scores remain on record. They do not establish general grounding.

Config	Cases	Purpose
cue_audit	80,16 fresh families	Two agreeing documents, a wrong-scope second document, contradiction within one document, two irrelevant documents and agreeing nonstandard simulator facts.16 of each pattern.
balanced_grounding	768,32 fresh families	Four answer categories ×three document counts ×two evidence orders per family. Each category occurs equally at each count.384 order pairs.
fresh_confirmation	768,32 additional families	Same balanced design, new R032..R063 families/values; frozen before the corrected Jev-curated retry trained.384 new order pairs.

The larger suite uses new fields, values, plan names and question/rendering forms. Every document has three factual statements; missing cases contain both wrong-scope and wrong-field distractors. Gold is computed from the distinct values actually recorded for the exact query scope/field: no value→UNKNOWN, one distinct value→that value, incompatible values→CONFLICT. A roundtrip parser checks the rendered statements. CuratorKIT schema and exact-dedup gates passed. Maximum chat input646 tokens; no evidence truncation or post-output filtering.

These are controlled synthetic diagnostics. Templates are shared within each new test, and each768-case cohort contains32 families. They are not a natural-corpus evaluation or768 independent facts. The separate frozen900-case FaithEval benchmark provides external transfer evidence.

All cases, IDs, gold answers, source and data hashes were frozen before any model generation on these tests. Manual review corrected simulator-question grammar before generation; the unused v1 draft is retained locally, with zero model calls. Active original balanced data is v2; its bytes and the cue-audit files remain unchanged. Those first follow-ups preserved the original training and selected Particle G0.001/strength1. Their later reuse for the corrected adapter is labelled a repeated research comparison, not fresh confirmation.

The additional fresh_confirmation cohort was frozen before the separate Jev-curated retry committed any update. Its six arms are OFF, original hinted particle1, original64-case LoRA, corrected particle1/2 and a same48-case LoRA. Corrected strength selection uses the original128 development cases before its first test output. No test winner replaces that choice. The data here contain no training targets for either adapter.

The exact original prompt, greedy decoding, nativeEOS3 and64-new-token cap apply to all arms. Strict finished-answer success, all per-case gains/losses and family-clustered intervals are reported separately, including unfavorable results. No model weights, private contacts or third-party passages are included here.

Own fictional data: CC0-1.0. Models and derived adapters retain Tiny Aya's CC-BY-NC-4.0 non-commercial licence.

Code, source builders and original results

Retained benchmark measurements · Corrected training curation