Use this skill to verify CLAIM-LEVEL grounding of a documentation page (or set of pages) against the source code. Activate when you have specific pages to check for factual accuracy -- not when sweeping a whole corpus (use docs-corpus-audit for that) and not when triaging a PR diff (use docs-sync for that). Trigger nouns: "is this doc accurate", "verify the page against the code", "fact-check this section", "any claims that drifted from source", "fact-checking", "grounding audit", "drift hunt", "claim verification". Returns per-claim verdicts (GROUNDED | PARTIAL | CONTRADICTED | UNSUPPORTED) with file:line evidence citations. Catches paragraph-level inaccuracies that page-level audit averages over -- e.g. a paragraph with 5 claims where 4 are grounded and 1 is fabricated. Does NOT modify files (returns advisory only); does NOT re-architect the docs; does NOT triage PRs.
87%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Passed
No findings from the security scan
The canonical home for this skill is docs-grounding-verifier in microsoft/apm
CLAIM-LEVEL grounding verification. Adapts the RAGAS faithfulness-eval pattern (proven in RAG literature) to docs/code instead of generated- answers/retrieved-context. Source code is the ground truth; docs paragraphs are the candidate text under audit.
python-architect persona doc-writer persona
This skill is a SIBLING of docs-corpus-audit and docs-sync. The
boundary is load-bearing:
| Skill | Trigger | Scope | Granularity |
|---|---|---|---|
| docs-sync | PR opened/synchronized | PR diff only | Page-level |
| docs-corpus-audit | Maintainer asks for whole-corpus pass | Entire corpus | Page-level |
| docs-grounding-verifier | Verify specific pages factually | 1..N pages | CLAIM-level |
docs-corpus-audit invokes this skill in its VERIFY phase on the
highest-risk pages of each wave. docs-sync can invoke it on the
specific pages in a PR diff. The skill is also runnable standalone.
docs-corpus-audit.docs-sync.editorial-owner persona directly.PARENT
-> [Stage 1: EXTRACT claims, fan-out PANEL]
per page -> LLM extracts atomic factual claims as JSON
script: scripts/extract-claims.py
-> [Stage 2: RETRIEVE evidence, deterministic S7]
per claim -> grep over src/ via keywords + hints
script: scripts/retrieve-evidence.sh (NO LLM)
-> [Stage 3: JUDGE grounding, adversarial A7]
per (claim, evidence) -> LLM rules GROUNDED|PARTIAL|CONTRADICTED|UNSUPPORTED
asset: assets/judge-prompt.md
-> [Stage 4: SYNTHESIZE]
aggregate ungrounded -> doc-writer for fix
re-verify after fix (A8 ALIGNMENT LOOP)Stage 2 is the load-bearing design choice: evidence retrieval is DETERMINISTIC (grep + AST hints), not LLM. The judge in Stage 3 can only rule on evidence it actually receives -- it cannot hallucinate support that the retriever did not find. This is the structural guard against the failure mode "the LLM convinces itself the docs match the code."
Input: list of page paths to verify (1..N). If a risk_class is
attached (e.g. "high-stakes"), prefer it; otherwise treat all as equal.
Out-of-scope:
docs/src/content/docs/ or
packages/apm-guide/.apm/skills/apm-usage/.For each page, dispatch ONE claim-extractor agent:
scripts/extract-claims.py <page> produces the
prompt and embeds the page content.{"page", "claims":[{"id","text","section","keywords", "expected_source_areas"}]} capped at 15 claims per page.Parallel safe; no shared state between extractors.
For each claim, pipe to scripts/retrieve-evidence.sh:
{"claim_id","claim_text","evidence":[...], "evidence_count"}.Sequential is fine (grep is fast). No LLM. Diagnostics on stderr, data on stdout.
For each (claim, evidence) tuple, dispatch ONE grounding-judge agent:
assets/judge-prompt.md.judge-prompt.md.Batching across claims-of-one-page into a single judge call is fine (prompt with all tuples at once). Across pages, fan out.
Aggregate verdicts. Materialize the report:
{
"summary": {
"pages_verified": N,
"claims_total": N,
"grounded": N, "partial": N, "contradicted": N, "unsupported": N,
"grounding_rate": N/total
},
"actionable": [
{"page", "claim", "verdict", "evidence_cited", "fix_suggestion"}
]
}CONTRADICTED and PARTIAL are doc-writer work items. UNSUPPORTED is
split: if retrieval_fix_suggestion is plausible, retry retrieval
with the suggested keywords; if still empty, treat as CONTRADICTED.
Hand actionable items to doc-writer (one subagent per page). After edits, RE-RUN the pipeline on the same pages. The grounding_rate must MONOTONICALLY INCREASE between iterations or the loop has diverged -- stop and escalate to the operator.
evals/) passes on the trigger evals and
the content evals before the skill is treated as production-ready.scripts/extract-claims.py -- Stage 1 prompt builder. --help, --schema.scripts/retrieve-evidence.sh -- Stage 2 retriever. Deterministic. --help.scripts/verify-page.sh -- end-to-end orchestrator. --help.assets/judge-prompt.md -- Stage 3 adversarial judge prompt.evals/trigger-evals.json -- 20 dispatch queries (10 should, 10 shouldn't).evals/content-evals.json -- seeded-drift recall scenarios.evals/run-evals.sh -- the eval-runner that turns JSON into metrics.references/.microsoft/apm
f574a62
Canonical home
since Nov 16, 2026
If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.