verify-document – Documents & Office skill for AI agents

Document Verification

Inspect a document for forensic authenticity signals — not a fraud verdict, but a risk band with the evidence behind it. Uses the Stipple API (free anonymous tier, no signup).

When to use

• Before onboarding a tenant, contractor, or employee from uploaded documents
• Before paying an invoice that arrived by email
• Before relying on a bank statement, payslip, or certificate in any workflow
• Reviewing documents in due diligence, claims processing, or loan applications

Service availability check

The documented Stipple endpoints, schemas and allowances are historical examples, not a verified current service contract. Confirm the intended provider and current official API documentation, pricing and privacy terms before any request or upload. If the document-verification/screening API cannot be confirmed, stop the API workflow and report it unavailable; do not send data to a guessed endpoint or a repurposed domain.

Instructions

1. Get the document. URL or local file path (PDF, PNG, JPEG, BMP, TIFF).

2. Optionally check the cache first. If the user has the file's SHA-256, check whether it's already been inspected (free):

   curl "https://www.stipple.sh/v1/warrants/check?sha256=<hash>"
   

3. Run verification. POST the document:

   curl -X POST https://www.stipple.sh/v1/warrants \
     -F "[email protected]" \
     -H "Authorization: Bearer $STIPPLE_API_KEY"
   

Add ?fresh=true to force re-inspection of a previously cached document. Add ?deep=true for deep inspection (more thorough, more credits).

4. Interpret the response. Two independent axes — read both:

| Axis | Question it answers |

|---|---|

| risk_band | Does anything look tampered? (low / medium / high) |

| inspection_quality | Could the engine actually see enough to judge? (thorough / limited / poor) |

A clean phone photo of a real payslip is commonly low risk + limited quality — low coverage is not risk. Per-signal evidence includes: amount/words mismatch, font discontinuity in values, date anomalies, document label integrity, identifier checksums (ABN/ACN/TFN), table arithmetic.

5. Report honestly. This is a *signal with evidence*, not a verdict:

• "risk_band: LOW — nothing looks tampered"
• "inspection_quality: limited — couldn't inspect everything; low coverage is NOT fraud"
• Show the per-signal evidence for anything flagged

6. Pair with related checks. For identity documents, follow with a 100-point identity check (/v1/identity-check). For extraction, use extract-document-data.

Output format

risk_band:           LOW — Nothing looks tampered.
inspection_quality:  limited
recommended action:  review_before_action

evidence (signals):
  [pass] Amount words/figure mismatch: Spelled-out amounts agree with figures.
  [pass] Font discontinuity in value: Numeric values share the font of surrounding text.
  [skip] Identifier checksum: No checksummable identifier (ABN/ACN/TFN) present.

Limitations and Safety

• This workflow uploads documents to a hosted third party. Obtain explicit approval,

minimize personal, financial, identity, and confidential data, and confirm the

provider's current retention, residency, access, and deletion terms first.

• Forensic signals can miss sophisticated tampering or flag benign editing, scanning,

compression, and template artifacts. A low-risk result is not proof of authenticity,

and a high-risk result is not proof of fraud.

• Preserve the original bytes and use authoritative issuer verification plus a

qualified human reviewer before payment, onboarding, lending, employment,

disciplinary, compliance, or legal action.

Notes

• Document types the engine recognizes (payslips, invoices, bank statements) get type-specific checks; unrecognized types get generic checks only — say so in your report
• Historical content-hash caching claims do not establish current speed, retention or cost; verify the active service contract.
• This measures forensic signals, not authorship style. AI-text detectors cannot establish who wrote a document and should not be used as proof of misconduct.
• Historical anonymous/key allowances are unverified; do not assume a free tier or enter credentials at an unconfirmed provider domain.