Artifact Card
Artifact Type
This artifact is a review-time hosted evaluation-record resource for an evaluation-methodology paper. It is not a new benchmark suite and should not be evaluated as a coverage-oriented dataset release. It contains claim-indexed synthetic/task-generated records, scorer outputs, and locked verification assets for the paper's tables and figures.
Croissant-compatible metadata is included for the materialized evaluation records. The records are not proposed as a standalone benchmark distribution; they are evidence artifacts for auditing the VCOS scoring contract. The canonical contract-instantiation resource is grouped under assets/vcos_instantiation_library/, covering 3 mechanism families, 4 paper-facing family cards, 16 carriers, and 192 deterministic instances as scale-credibility evidence rather than benchmark prevalence evidence. The reviewer-core source under src/arena/canary_crucible/ exposes the H1/H2 protocol mechanisms that convert visible mixed-version packets into semantic outcomes and claim-readiness gates.
Intended Use
Reviewers should use this artifact to verify that each paper claim points to a concrete source asset, expected denominator or metric, and offline verification command. The verifier recomputes the main H1/H2/H3 claim checks from released normalized records where the raw provider-private logs are excluded: h1_claim_recompute_records.jsonl, h2_claim_recompute_records.jsonl, and h3_rcr_paired_scorer_records.jsonl.
For method reuse, src/arena/canary_crucible/eval_protocol/ contains the shared claim-readiness and outcome-stratification helpers; families/builtin/stale_policy_binding_state.py contains the H1 visible-state oracle and semantic classifier; families/builtin/stateful_session.py contains the H2 realized-packet dynamic oracle, action-object scorer, and control-regime preflight. The package omits live provider wrappers and paper renderers because they are not part of the reusable evaluation protocol core.
Evidence Scope
H1 and H2 are paper-facing mixed-version evaluation audits. H3 is a 640-pair paired terminal-invariance scorer-validity stress test, not a broad external prevalence estimate. ToolSandbox provides WITNESS_ONLY strict laundering evidence and a receipt-salience mechanism audit as paper-facing appendix sidecar evidence, while the failed diversity scout blocks any cross-model strong-externality claim. Tau-style action/state records are retained as release provenance only: clean-current controls pass, while neutral and preregistered completion-pressure conflicts mostly induce evidence requests or abstention, so no positive holdout is claim-bearing. Tau-style records are not a full tau-bench leaderboard, an official database-evaluator integration, or a PDF claim. The VCOS instantiation library is contract-reuse evidence, not a claim-bearing benchmark suite. The standalone artifact verifies released records, figures, manifests, and locked hashes; it does not rebuild paper assets or the PDF.
Privacy And Provider Metadata
The review package excludes live API keys, account identifiers, raw provider headers, provider request IDs, raw response bodies, and token-usage records. It keeps normalized model/action/scorer records and aggregate summaries needed for auditability.
Provider-route provenance is deliberately bounded to reviewer-auditable, non-private fields:
| Model lane | Provider route | Official? | Logged fields | Claim role |
|---|---|---|---|---|
| qwen | official | yes | model slug/date/normalized action | primary lane |
| minimax | official | yes | model slug/date/normalized action | primary or diagnostic, depending on gate |
| glm | official | yes | model slug/date/normalized action | primary lane |
| GPT lane | routed provider | no; intermediary route | model slug/date/normalized action only | bounded external/model lane |
| Claude lane | routed provider if present | no; intermediary route | model slug/date/normalized action only | diagnostic only |
Release Plan
For camera-ready release, code and non-provider-private synthetic/task-generated records will be released under a permissive open-source license, subject to upstream benchmark licenses. Raw provider identifiers, account metadata, and private API logs are excluded.