Verify Report Skill

Claims verification is the quality gate that separates evidence-backed research from plausible-sounding prose. This skill runs in a fresh context window — separate from the research pipeline — so that claims extraction, source verification, and the review loop get the full attention they deserve without competing for context with research data.

When this skill loads:

1. If a cogni-research project exists in the working directory → auto-detect and proceed (Mode A)
2. If the user provides a path to a markdown file → standalone verification (Mode B)
3. If neither → ask: "Which report should I verify? Provide a project path or markdown file."

Prerequisites

• cogni-claims plugin installed (required — this skill has no fallback)
• For Mode A: a cogni-research project with at least output/draft-v1.md or output/report.md
• For Mode B: a markdown file with inline citations containing source URLs

Workflow

Phase 0: Locate and Load Report

Mode A — cogni-research project

A project directory contains entity directories (00-sub-questions/, 01-contexts/, etc.) and .metadata/project-config.json. This skill loads only the draft and source entities — not sub-questions or contexts — keeping the context window lean for verification work.

1. Find the project directory. Search strategy:
   • Check if cwd contains .metadata/project-config.json
   • Scan subdirectories one level deep for .metadata/project-config.json
   • If multiple found, list them and ask the user to choose
   • If none found, ask the user for the path
2. Read .metadata/project-config.json for report type, topic, language
3. Read .metadata/execution-log.json for phase completion state
4. Locate the report to verify, in priority order:
   • Highest output/draft-v{N}.md (latest draft, even if not finalized)
   • output/report.md (finalized report, for re-verification after user edits)
5. Load source entities from 02-sources/data/ to build source lookup (URL → entity). This is the only research data loaded — sub-questions and contexts are NOT loaded
6. Check for existing verification state → proceed to Phase 0.5

Mode B — standalone markdown file

When the user provides a path to a markdown file outside a cogni-research project:

1. Create a lightweight verification workspace: .verify-report/{slug}/ as a sibling to the markdown file
2. Within the workspace, create: 03-report-claims/data/, .metadata/, cogni-claims/
3. No source entities exist — claim-extractor will extract source URLs directly from inline citations
4. Set standalone_mode = true in .metadata/project-config.json

Read: references/standalone-mode.md for citation detection patterns and workspace layout.

Phase 0.5: Resumability Check

If previous verification artifacts exist, present the user with options rather than silently re-running or silently resuming:

No prior verification (no 03-report-claims/data/ content, no cogni-claims/claims.json): → Proceed to Phase 1.

Claims extracted but not submitted (03-report-claims/data/ has entities, cogni-claims/claims.json absent): → Resume from Phase 2 (submission).

Verification incomplete (cogni-claims/claims.json exists with status: pending claims): → Resume from Phase 2 mid-point (verification).

Previous verification complete (cogni-claims/claims.json exists with completed results):

Previous verification found

• Claims verified: N | Confirmed: N | Deviations: N
• Verified at: {timestamp}

Options:

• re-verify — clear previous results, re-extract claims from current draft, full verification
• inspect — show previous results without re-running (opens cogni-claims dashboard)
• continue — keep existing results, proceed to review loop

Handle the user's choice:

• re-verify: clear 03-report-claims/data/ contents and cogni-claims/claims.json, proceed to Phase 1
• inspect: Skill(cogni-claims:claims, mode=dashboard, working_dir=<project_path>), then stop
• continue: proceed to Phase 3 (interactive review) with existing verification data

Phase 1: Claim Extraction

Spawn the claim-extractor agent to identify verifiable factual claims in the draft:

Task(claim-extractor,
  PROJECT_PATH=<project_path>,
  DRAFT_PATH=<resolved draft path>,
  DRAFT_VERSION=<N>)

The claim-extractor creates report-claim entities in 03-report-claims/data/, each linking a factual statement to its cited source URL. It prioritizes statistical claims, attribution claims, causal claims, and definitional claims — in that order.

After extraction, report to the user:

Extracted N verifiable claims from the draft (N skipped — unsourced or general knowledge).

Phase 2: Claims Submission + Verification

2a: Submit claims to cogni-claims

Read: references/claims-integration.md for the cogni-claims submission protocol.

Collect report-claim entities from 03-report-claims/data/ and submit as a batch:

Skill(cogni-claims:claims, mode=submit,
  working_dir=<project_path>,
  claims=[...extracted claims from report-claim entities...],
  submitted_by="cogni-research/verify-report")

2b: Verify claims against sources

Skill(cogni-claims:claims, mode=verify,
  working_dir=<project_path>)

cogni-claims dispatches claim-verifier agents (one per unique source URL) that fetch each source, compare claims against actual content, and detect 5 deviation types: misquotation, unsupported_conclusion, selective_omission, data_staleness, source_contradiction.

2c: Update report-claim entities

After verification, read cogni-claims/claims.json and update report-claim entities:

• Set verification_status to verified/deviated/source_unavailable
• Set deviation_type and deviation_severity if deviated
• Set claims_submission_id to the cogni-claims claim ID

Phase 3: Interactive Claims Review

Present verification results to the user before proceeding to automated review. This ensures the user has visibility into what was verified and can steer corrections.

1. Read {PROJECT_PATH}/cogni-claims/claims.json for verification results
2. Summarize results into a compact overview
3. For each deviated claim: include the claim statement, what the source actually says, deviation type, and severity
4. Present to the user via AskUserQuestion:

Claims Verification Results

Verified: N | Confirmed: N | Deviations: N | Sources unavailable: N

Deviations found:

1. [claim statement] — [deviation_type] ([severity]): [explanation]
2. ...

Options:

• proceed — pass deviations to reviewer + revisor for automated correction
• fix: 1, 3 — flag specific claims for mandatory correction
• drop: 2 — remove specific claims from the report entirely
• accept — mark report as verified, finalize without revision
• inspect: 2 — open cogni-claims inspect mode for claim 2 (detailed source comparison in browser)

How would you like to proceed?

5. Process user response:
   • proceed → continue to Phase 4 with all deviations as reviewer input
   • fix: N, M → add flagged claims to a mandatory-fix list passed to the reviewer
   • drop: N → add to a drop list; revisor will remove these claims from the report
   • accept → skip Phase 4, proceed directly to Phase 5 (finalization)
   • inspect: N → Skill(cogni-claims:claims, mode=inspect, claim_id=<id>), then re-present options
6. Store user decisions in .metadata/user-claims-review.json:

{
  "reviewed_at": "<ISO timestamp>",
  "total_claims": 25,
  "confirmed": 20,
  "deviated": 4,
  "source_unavailable": 1,
  "user_action": "proceed|fix|drop|accept",
  "fix_claims": ["claim-id-1", "claim-id-3"],
  "drop_claims": ["claim-id-2"]
}

Phase 4: Review + Revision Loop

This phase runs only if the user chose proceed or fix in Phase 3. The reviewer receives the draft + claims data + user decisions — no research data, keeping context focused.

Read: references/review-criteria.md for the scoring rubric (shared with research-report).

Maximum 3 iterations. Each iteration:

4a: Review

Task(reviewer,
  PROJECT_PATH=<project_path>,
  DRAFT_PATH=<current draft path>,
  CLAIMS_DASHBOARD=<project_path>/cogni-claims/claims.json,
  USER_CLAIMS_REVIEW=<project_path>/.metadata/user-claims-review.json,
  REVIEW_ITERATION=N,
  OUTPUT_LANGUAGE=<output_language>)

The reviewer scores the draft on 5 structural dimensions (completeness, coherence, source diversity, depth, clarity) and multiplies by the claims verification rate. It flags high/critical deviations as mandatory fixes and applies user override decisions.

4b: Revise (if verdict="revise" and iteration < 3)

Task(revisor,
  PROJECT_PATH=<project_path>,
  DRAFT_PATH=<current draft path>,
  VERDICT_PATH=".metadata/review-verdicts/v{N}.json",
  NEW_DRAFT_VERSION=<N+1>,
  OUTPUT_LANGUAGE=<output_language>,
  MARKET=<market>)

After revision:

1. Clear 03-report-claims/data/ (revised claims may differ)
2. Re-run Phase 1 (claim extraction on new draft)
3. Re-run Phase 2 (re-submit and re-verify)
4. Re-run Phase 3 (present new results to user)
5. Continue to next review iteration

4c: Accept

When verdict="accept" or iteration reaches 3: proceed to Phase 5.

Phase 5: Finalization

1. Copy the accepted draft to output/report.md (overwrite if exists)
2. Update .metadata/execution-log.json:
   • phase_5_review.claims_verification: actual verification stats (not "deferred" or "skipped")
   • phase_5_review.verification_rate: N.NN
   • phase_5_review.review_iterations: N
   • phase_5_review.final_score: N.NN
   • phase_5_review.verified_by: "verify-report"
3. Report summary to user:

Verification Complete

• Claims: N extracted, N confirmed, N deviated (N fixed), N sources unavailable
• Verification rate: X.XX
• Review iterations: N, final score: X.XX
• Verified report: output/report.md

Error Recovery

Scenario	Recovery
cogni-claims not installed	Cannot proceed — this skill requires cogni-claims. Tell the user to install cogni-claims plugin
All source URLs unreachable	Report results with source_unavailable count. Suggest user run /claims cobrowse for interactive recovery, or check URLs manually
Claim extraction produces 0 claims	The draft may lack inline citations. Suggest re-running research-report writer with citation requirements
Review loop reaches max (3)	Accept current draft with quality warning
Project directory not found	Ask user for explicit path

References Index

Reference	Read When
references/claims-integration.md	Phase 2 — cogni-claims submission + verification protocol
references/standalone-mode.md	Phase 0 Mode B — standalone markdown verification
references/review-criteria.md	Phase 4 — understanding review scoring (shared with research-report)

Note: references/review-criteria.md is a symlink to ../research-report/references/review-criteria.md — both skills share the same scoring rubric.