# Evidence Contract

Panvascular intervention research is starting to write traceability into an agent's output contract: ScienceBuddy-PV fixes a shared convolutional model on the public ARCADE coronary angiography benchmark and evaluates only an evidence gate and action audit, where the gate accepts 76.7% of stenosis-frame candidates against 100% for a score-only comparator. On the locked test set, stenosis Dice is 0.231 and AUROC is 0.957 at a pixel prevalence of 0.00975, while reference lesion candidates are accepted 44.3% and rejected 55.7% of the time. The result is an image-level technical validation; recursive harness optimization and model learning remain protocols rather than executed experiments.

Panvascular intervention research is starting to write traceability into an agent's output contract: ScienceBuddy-PV fixes a shared convolutional model on the public ARCADE coronary angiography benchmark and evaluates only an evidence gate and action audit, where the gate accepts 76.7% of stenosis-frame candidates against 100% for a score-only comparator. On the locked test set, stenosis Dice is 0.231 and AUROC is 0.957 at a pixel prevalence of 0.00975, while reference lesion candidates are accepted 44.3% and rejected 55.7% of the time. The result is an image-level technical validation; recursive harness optimization and model learning remain protocols rather than executed experiments.

The work proposes two enforceable contracts for LLM agents over databases: a plan contract that statically checks plan structure, result-field references, and grounded identifiers before execution, and an evidence contract that records belief state, completeness, and provenance so claims can be verified; in TGMS, result-field checking turns invalid references into repairable rejections, completeness propagation detects truncated-count failures in all 15 controlled cases, and claim gating leaves none of 199 emitted answers with an unsupported gated claim at the cost of 21 fewer answers.

The work proposes two enforceable contracts for LLM agents over databases: a plan contract that statically checks plan structure, result-field references, and grounded identifiers before execution, and an evidence contract that records belief state, completeness, and provenance so claims can be verified; in TGMS, result-field checking turns invalid references into repairable rejections, completeness propagation detects truncated-count failures in all 15 controlled cases, and claim gating leaves none of 199 emitted answers with an unsupported gated claim at the cost of 21 fewer answers.

EIO-Agents introduces an open specification with two layers: the Evaluation Intelligence Ontology (EIO) defines typed evidence, versioned behavioral predicates with evidence contracts, claims, witness rules, and proof status, while the Portable Evaluation Record (PER) preserves one evaluation's evidence-to-decision chain as a canonical, content-addressed record; in a reference credit-underwriting evaluation, 243 claims project, validate, and explain to a BLOCK decision, whereas an identical capped score in another record yields REVIEW because its proxy observation did not reproduce.

EIO-Agents introduces an open specification with two layers: the Evaluation Intelligence Ontology (EIO) defines typed evidence, versioned behavioral predicates with evidence contracts, claims, witness rules, and proof status, while the Portable Evaluation Record (PER) preserves one evaluation's evidence-to-decision chain as a canonical, content-addressed record; in a reference credit-underwriting evaluation, 243 claims project, validate, and explain to a BLOCK decision, whereas an identical capped score in another record yields REVIEW because its proxy observation did not reproduce.