Evidence before narrative

Publication-facing artifacts are accepted only when they preserve the causal, statistical, and provenance contracts below.

Four hard rules

1. query_message_index < action_message_index for every chronological sample.
2. Repeated targets are multi-positive examples, not in-batch negatives.
3. all_splits and train_split candidate pools are reported together.
4. Mean gains never replace heldout losses, candidate coverage, or session-clustered uncertainty.

Artifact classes

Class	Tracked	Purpose
Compact JSON summaries	Yes	Canonical metrics, metadata, and audit outcomes
Provenance manifests	Yes	Dataset, model revision, config, seed, and source linkage
Paper tables and figures	Yes	Human-readable presentation generated from accepted evidence
Row-level predictions	No	Large local evidence used for paired tests
Embedding caches and datasets	No	Rebuildable or upstream-controlled material
PDFs and arXiv tarballs	No	Generated publication outputs

The complete policy is maintained in ARTIFACT_POLICY.md.

Claim ledger

Claim	Status	Required caveat
Chronological state helps Hermes search-query retrieval	Supported	Target type and candidate-pool dependent
Full combined gain is a memory effect	Rejected	Explicit target prior explains a larger share
Support gating is no-loss under tool shift	Rejected	Cold heldout tools can lose
URL results provide clean state evidence	Rejected	Literal repetition dominates the task
External open-vocabulary retrieval is solved	Not supported	TAU train-only coverage averages 11.92%

Documentation gate

scripts/docs_contract.py checks that homepage, sidebars, README, evidence docs, result values, and critical artifact paths stay aligned. scripts/docs_site_e2e.py builds the site and checks both locales, sitemap coverage, H1s, links, assets, and Mermaid rendering. CI runs both gates.

When a result changes, update the machine-readable artifact first, then paper and documentation in the same pull request.