cubetribe/codex_godmode_on/godmode-paperwork

GodMode Paperwork

Operate an evidence-first document pipeline. Treat OCR and model interpretation as probabilistic inputs; make hashes, source anchors, validation rules, review decisions, and archive construction reproducible and auditable.

Non-negotiable boundaries

• Keep every case root outside Git worktrees and plugin caches.
• Preserve originals byte-for-byte. Never move, overwrite, rename in place, delete, or submit them.
• Run the bundled CLI offline. Never add network, cloud, installer, or autonomous filing behavior.
• Stop on integrity failures. Return REVIEW when evidence is incomplete or uncertain.
• Require explicit human approval for critical claims and release of a review-bearing archive.
• Describe results as prepared or validated evidence, not legal advice, tax advice, compliance certification, or guaranteed truth.
• Treat the named local binaries and trained language files as an external trust boundary. The CLI records their resolved paths, versions, hashes, and argument contracts but cannot prove that a binary is benign.
• Treat audit hashes and declared-human approvals as internal consistency evidence, not signatures or protection against an attacker who can rewrite the whole case.
• Require Python 3.11 or newer. Native Windows behavior remains deferred; use a validated Unix-like environment for this release.

Run the workflow

Set the CLI path once:

PAPERWORK="<skill-directory>/scripts/paperwork.py"

1. Inspect local capabilities without installing anything:

``bash python3 "$PAPERWORK" --json doctor --languages deu,eng --require pdf-native,pdf-render,ocr ``

1. Create a case in a deliberately chosen non-Git directory:

On macOS, do not use /tmp: it resolves through a symlink and the strict case-path policy rejects it. Use /private/tmp only for disposable tests.

``bash python3 "$PAPERWORK" intake \ --case /secure/cases/example \ --case-id example-2026 \ --input /path/to/document.pdf \ --assurance EXTRACTION \ --languages deu,eng ``

1. Extract existing PDF text before OCR:

``bash python3 "$PAPERWORK" extract --case /secure/cases/example ``

1. Route only unresolved pages to local OCR:

``bash python3 "$PAPERWORK" route --case /secure/cases/example ``

1. Resolve a page only after route places it in VISUAL_REVIEW. This applies to both EXTRACTION and CASEWORK:

``bash python3 "$PAPERWORK" resolve-page --case /secure/cases/example --input /secure/input/page-resolution.json ``

1. For CASEWORK, write requirements, claims, and approvals only through the schema-checked commands. Keep each input JSON outside the case root:

``bash python3 "$PAPERWORK" requirements --case /secure/cases/example --input /secure/input/requirements.json python3 "$PAPERWORK" claim --case /secure/cases/example --input /secure/input/claim.json python3 "$PAPERWORK" approve --case /secure/cases/example --input /secure/input/approval.json ``

Never edit records/*.jsonl, requirements.json, page routes, or validation state manually.

1. Validate the selected assurance level, then verify integrity:

``bash python3 "$PAPERWORK" validate --case /secure/cases/example python3 "$PAPERWORK" verify --case /secure/cases/example ``

1. Pack only after validation. A plain tar archive requires an explicit acknowledgement:

``bash python3 "$PAPERWORK" pack \ --case /secure/cases/example \ --output /secure/exports/example.paperwork.tar \ --contents evidence \ --actor operator \ --acknowledge-plain-archive python3 "$PAPERWORK" verify --pack /secure/exports/example.paperwork.tar ``

Any successful mutation invalidates the prior validation. Validate again immediately before packing. Pack refuses a state, audit head, or audit count that differs from the latest validation.

In 2.10, both evidence and full packs retain every state-bound case artifact so an extracted case remains independently verifiable. The labels are reserved for a future versioned subset contract.

Use --json for machine-readable results and inspect exit codes before continuing: 0 is PASS, 10 is REVIEW, 2 is CLI usage error, 20-26 are controlled failures, and 70 is an internal error.

Read references progressively

• Read evidence-contract.md before adding claims, requirements, approvals, or downstream automation.
• Read ocr-escalation.md when native extraction is incomplete or OCR tools are missing.
• Read safety-and-review.md before handling sensitive material, closing review findings, or packing a case.

The eight JSON Schemas under assets/schemas/ distinguish persisted records from mutation inputs: approval input and stored approval are separate, visual page resolution has its own input schema, and claim anchors encode method-specific bounding-box rules. For OCR anchors the CLI derives the normalized box from the unique contiguous TSV word sequence; a supplied box must match exactly. The CLI remains the enforcement authority and uses only Python's standard library.