<!-- Generated by scripts/generate-plugin-artifacts.py; do not edit. -->

Codex wrapper for /verify

Follow the OloLand command instructions below when the user asks for /verify or the equivalent workflow in Codex.

/verify (deal-scoped, corpus-based)

Runs OloLand's verification layer over an artifact you drafted anywhere — including in Claude Cowork or another tool — checked against an OloLand deal's ingested data room. For each figure it retrieves the supporting evidence from the corpus itself, so it catches numbers the author cited to the wrong source, or that the data room contradicts. Get back a per-figure verdict + a content-hashed provenance pack you can defend at IC.

It requires a deal with an ingested corpus. If you don't have one, this command walks you through creating it (create_deal + upload your documents) — that is the path from a Cowork draft to a verifiable OloLand deal.

Usage

/verify

Then provide the artifact and the deal to verify it against (see Inputs).

Inputs

Ask the user for these if missing:

1. artifact_text (required) — the drafted memo / analysis / answer to verify. Accept any of: text pasted inline, a path to a local file (read it with your file tool and pass the contents), or the body of something the user generated elsewhere (e.g. a Cowork IC memo).
2. deal_id (optional, required for verification) — the OloLand deal whose ingested data room backs the verification. If the user has no deal, do NOT block or repeatedly ask for one — proceed to the funnel (see Execution step 1) so they get the create_deal path.

Optional:

• rubric — override the grader rubric (defaults to the institutional DD rubric).
• mode (default strict) — strict fails closed: if a provider call (grader / NLI) fails or times out, the affected figures come back UNVERIFIED, never a silent pass. advisory lets such cases pass through.

Execution

1. No deal yet? If the user has no deal_id, call mcp__ololand__grade_external_artifact with just the artifact_text — it returns status: "needs_deal". Relay its guidance: create a deal and upload the data room with mcp__ololand__create_deal, then re-run /verify with the returned deal_id. (This is the intended onboarding path — do not fall back to pasting evidence.)
2. Verify. Call mcp__ololand__grade_external_artifact with artifact_text, deal_id, mode (and rubric if supplied).
3. Handle the status field:

• ok — render the verdict (see Output).
• needs_corpus — the deal exists but has no ingested documents. Tell the user to upload the data room (ingest), then re-run.
• search_error — a retryable corpus-search failure (not a missing corpus); tell the user to retry.

1. Do not re-compute or second-guess the per-figure statuses — they come from the deterministic engines, the NLI check, and the independent grader.

What it checks (per figure)

• Corpus number reconciliation — each $ / % figure must reconcile (5% rel / 0.5pt abs) against the chunks retrieved for that figure from the deal corpus. A 38% whose data-room source says 11% fails.
• NLI derivation — the retrieved source must entail the claim (entail / neutral / contradict), so a guessed number that happens to appear is caught — not just a digit match.
• Independent grader — a Gemini grader (a genuinely different model from the one that drafted the artifact) judges figure support and attribution.
• Fail-closed (strict mode) — provider degradation surfaces UNVERIFIED, never a green pass; a figure with no supporting corpus evidence is UNVERIFIED (we report we could not confirm it, we do not silently pass it).

Output structure

Render a single verdict block:

/verify verdict — overall: <PASS | FAIL | UNVERIFIED>   (deal: <deal_id>, mode: <strict|advisory>)

Per figure:
  <value>  source [<file>, p.<n>]   <PASS|FAIL|UNVERIFIED>
     reason: <one line>
     checks: number_matches=<bool>  nli=<entail|neutral|contradict|—>

Coverage: <marker_count> corpus chunks used
Grader (<model>): <passed|failed|—>  <rationale>
Degraded: <list, if any provider could not be reached>

Provenance: run_id=<…>  sha256=<…>

Then list every FAIL and UNVERIFIED figure verbatim (value + source file/page + reason) — those are the deal-killing review items. For each FAIL, recommend the user check the figure against the data room (the corpus source that should support it) or remove/qualify the claim. Surface the provenance.run_id + content_sha256 as the artifact the user can hand to an IC chair or LP.

When to use this

• Before an IC memo, analysis, or board section drafted outside OloLand (e.g. in Cowork) goes to committee — verify it against the deal's ingested data room.
• To spot-check a single number that "looks off" in any drafted text, against the corpus.
• As a pre-send gate on AI-generated financial prose produced in any tool.

When NOT to use this

• For prose with no $ / % figures — there is nothing to reconcile.
• When you have no deal and no intention of ingesting one — verification is grounded in the corpus; without it there is nothing to verify against (the command will return needs_deal).

Honest bounds

• This is precision over the corpus: it checks that the figures the memo states are supported by the data room. It does not yet flag material risks the memo omitted (a recall / completeness pass is a separate, forthcoming capability). A PASS means the stated figures reconcile — not that the memo is complete.
• Deterministic number reconciliation covers dollar and percentage figures. Multiples (8.5x), ratios, and bare counts are not deterministically reconciled — they rely on the grader. Do not represent the result as "every number verified".

Companion: the deal-scoped verifier-stack in ololand-dd runs the four atomic-claim verifiers + cross-document reconciliation against the same ingested corpus.