Research Verification Gate

Research Verification Gate

Use this as the last gate before claiming a research output is ready.

Required checks

• the stated scope was actually answered
• important claims still have supporting evidence
• available structured artifacts (sources.jsonl, claims.jsonl, guards.jsonl, delivery.json, source ledgers, evidence maps) were inspected when present
• time-sensitive facts include concrete dates when needed
• requested format/style context was inspected for blog, article, report, or other format-matched writing
• the active writing-style profile from writing-style-settings.md was loaded and recorded for finalizable writing, including style_profile_ref, policy_hash, active_overlays, active_requirement_ids, and style_applied
• mathematical, TCS, graph-theoretic, formal-proof, or LaTeX writing loaded math-manuscript-style.md
• every input the conclusions rest on was read whole, or its truncation is disclosed: tool payloads reporting complete: false, capped subprocess output, partial retrievals, and summaries standing in for full sources
• remaining gaps are disclosed
• incomplete analysis is used when material scope is still unchecked, and when a load-bearing source was read only in part
• residual uncertainty is listed for unfinished or disputed load-bearing claims
• multi-LLM LGTM / same-family agreement alone is not bankable supporting evidence; require different-family re-derivation and/or a machine-checkable artifact for banking language
• review rounds that only reword without an evidence delta do not count as progress; do not continue until all reviewers approve
• open negative-space / blocked-route failures stay disclosed (halt and disclose)

Output contract

Produce a short visible section titled Delivery Check.

Include:

• Status — READY or NOT READY
• Gate version — the version recorded for research-verification-gate in manifest/skills.yaml (unversioned when absent), so past verdicts stay interpretable as the gate evolves
• Confirmed — the key checks that passed
• Gaps — anything still blocking delivery
• Residual uncertainty — unfinished / disputed / negative-space items
• Next step — deliver now or fix specific gaps first
• Style — style_profile_ref, policy_hash, active overlays, active_requirement_ids, and whether style_applied is supported
• Formal status (when formal claims appear) —
  • opengauss_run: completed | failed | not_used
  • lean_check_status / placeholder / trust-base from strict gate
  • claim_support_status from deep-research ladder
  • statement_relation_status / review_status
  • OpenGauss success alone → NOT READY for “proved C”

Use the checklist in references/checklist.md.

Guardrails

• do not silently downgrade a blocker into a caveat
• if material scope is unchecked, require incomplete analysis
• treat undisclosed truncation as unchecked scope, not as a formatting nit: a conclusion drawn from a silently partial source is unsupported, whatever the source appeared to say
• do not treat multi-LLM LGTM as proved or supported without different-family or machine-checkable support
• keep the gate short and concrete
• staged fail-visible rollout: while enforcement is warn-only, log every downgraded blocker or skipped check (in the run ledger when one exists) with its reason; a later fail-closed stage may make NOT READY a hard stop, and the override log is itself research evidence

Recommended templates

When this skill is involved, consider these workflow templates (install via the workflow-templates artifact profile, or --with-deps to pull backing skills):

• autonomous-research-loop-runbook -- Bounded autonomous research-loop runbook with four stop conditions, single-path solving, mandatory cross-agent verification, fresh-agent backtracking, and five-lane broker-routed heavy-compute offload with per-lane safety gates.
• cross-agent-adversarial-review -- Producer-never-confirmer adversarial review of a paper, proof, or code artifact across agent families with a fresh-agent confirmation gate.