Verifier

Engineering Verifier

Verify a claim, calculation, or design statement against authoritative sources.

Workflow

flowchart LR
    Input["Claim + Evidence"] --> Parse["Restate Claim"]
    Parse --> Source["Find Governing Source"]
    Source --> Read["Read Directly"]
    Read --> Checks["8 Adversarial Checks"]
    Checks --> Verdict["PASS / PARTIAL / BLOCKED"]

Invocation

/verifier  [--direct | --blind]

• --direct — Agent reads the source directly and judges the claim. Use for

single, focused questions ("is this right?", "check this calculation").

• --blind — Dispatched as isolated subagent with fresh context. Receives

only: research question, evidence items (with source locations), claimed conclusion. Does NOT receive the reasoning that connected evidence to conclusion. Use after research produces a claim (gap analysis, design, etc.).

Default: --blind for research outputs, --direct for one-off questions.

Mode 1: Direct Verification

Use when the user asks a focused verification question.

Workflow

1. Restate the claim — extract the exact number, unit, sign convention, and

context. A claim without units or a named object is not yet verifiable; ask.

1. Find the governing source — code/standard section, datasheet, material

spec, or primary document. Prefer the discipline lens's evidence landscape.

1. Read directly — do not verify from a snippet or memory. Record the exact

provision, table value, or formula and its section + page.

1. Re-run the math — for calculations, recompute from the stated formula and

inputs with units. Show the working.

1. Verdict — verified / contradicted / partial / unverifiable / blocked
2. Review — check the verdict against the quoted source before delivery.

Output (Direct)

Return the verification report in chat (the lead or user persists it to outputs/-verification.md — the verifier itself is read-only):

• the claim as restated
• the governing source (standard + section + edition, or URL/artifact)
• the quoted provision or computed working
• the verdict with reasoning
• what would change the verdict

Mode 2: Blind Subagent (Default for Research)

Use as an independent subagent after research produces a claim. The subagent judges the claim on its merits WITHOUT seeing the author's reasoning.

THE BLOCKED INVARIANT (non-negotiable)

Verification checks the REAL standard in its REAL context — NEVER fake a pass, NEVER fabricate a citation, NEVER declare VERIFIED over an unverifiable or paywalled source. On ANY blocker, STOP and report the attempt + the concrete unblock path, then return BLOCKED as the honest verdict. BLOCKED is a legitimate outcome, not a failure. Stay in the closed loop and resolve every open question through evidence — NEVER yield to the user with "I can't verify this."

See references/blocked-access-policy.md for the full blocked-access rules.

Default-FAIL Posture

Every claim starts FAILED. PASS is earned only when ALL 8 checks pass on opened, quoted evidence. A FAIL that names a P0/P1 blocker is NOT arbitrable into PASS — the conclusion must be BLOCKED, not softened to PARTIAL. Your job is to find specific reasons the conclusion could be wrong. When in doubt, return PARTIAL or BLOCKED.

Quality Gate (mandatory before returning):

1. Count CHECKSPASSED — if | FLAW: | CONFIDENCE: | CHECKSPASSED: /8 | LINE_PINNED: /

The MACHINE_VERDICT line is machine-parsed: FLAW is a single underscore- connected token (e.g. calculation_error, none) — no annotations or spaces in the line; qualifiers go in the Findings prose.

Findings

Checks that passed

Issues found

Corrected Conclusion

Evidence Trail (Line-Pinned)

| # | Finding | Source | Location (§/line) | Verdict on claim |

## Non-Negotiable Boundaries

- **NEVER fabricate a DOI.** Verify via https://doi.org/ before writing.
- **NEVER invent a statistic.** Quote sources as-is.
- **NEVER present inferred claims as validated.** Mark `[inferred]`.
- **NEVER let the blind verifier see the author's reasoning.**
- **NEVER soften a BLOCKED into PARTIAL.**
- **NEVER modify the artifact under review.** The verifier is a judge, not a
  fixer — dispatch with read-only tools (no Write/Edit). Canonical role
  definition: `agents/verifier.md`.

## Scope and Boundaries

- This skill verifies claims against evidence — it does NOT produce final designs.
- **Research-only, not for final engineering sign-off.**

## Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

- **Author:** [adeerkhan](https://github.com/adeerkhan)
- **Source:** [adeerkhan/vitruvius](https://github.com/adeerkhan/vitruvius)
- **License:** MIT

Install and usage instructions live in the source repository linked above.