Discipline utilities — executable scope and limitations

Python 3.10+; standard library only. The existing four baseline utilities remain in place. The two additions are local decision-support tools, not an agent runtime, credential system or deployment service.

1. Advisory role router

python3 Tools/discipline_router.py --task "The contact form says sent but no email reaches the mailbox" --intent repair --output /tmp/inquiry-route.json --prompt-output /tmp/inquiry-context.txt
python3 Tools/discipline_router.py --task "Review the DMARC policy" --role R17

Run from the extracted library root. Choose fresh output filenames: the tool refuses to overwrite existing outputs or its input task file. --task-file reads a UTF-8 task file instead of inline text. --intent is audit, repair, publish or release; every mode is advisory/read-only. --role supplies an explicit accountable lead. --max-prompt-roles accepts 1–8 and defaults to 4.

The proposed route includes a lead, consulted roles, matched terms, reading paths and limitations. The context pack includes a bounded set of standalone prompts and names other consultees without claiming they executed. The rules are deterministic phrase/term matching, not semantic diagnosis; they do not understand every synonym, negation or causal relationship. Review and override where necessary. Security-exposure, availability and known critical-path phrases have priority. No shell commands, network requests or model calls are executed.

The complete task is written into the route/context. Treat those outputs as private when task text includes sensitive details; do not upload them to a public directory. The embedded instruction boundary reduces ambiguity but is not a formal proof that an external model resists prompt injection.

2. Evidence-record consistency gate

python3 Tools/evidence_gate.py --contract /path/to/approved-contract.json --record /path/to/acceptance-record.json --evidence-root /path/to/private-evidence

Use actual authorized paths. Optional --output writes a new JSON result and refuses overwrite. Exit 0 means structural validation passed, 1 means the submitted evidence/contract is blocked or inconsistent, and 2 means invalid invocation or unreadable/malformed input. Duplicate JSON keys and nonobject top-level files are rejected.

The contract requires schema version 1; kind release/publication; meaningful contract/project/build/environment identifiers; an approval declaration and timezone-aware approval time; a minimum independence level; and nonempty required checks. Each check defines ID, owner role, requirement, explicit N/A permission and minimum evidence count. Release contracts require at least a declared separate operator.

The record must match the exact contract SHA-256 and identity fields, include timestamped builder/verifier identity/run declarations and declare actual independence. Separate reviews require distinct runs; separate operators require distinct actor declarations and owner review. Publication records additionally require owner review of the article. These are consistency checks on declarations, not identity authentication.

Each PASS needs method, observed result, verifier execution/run, observation time and sufficient distinct evidence paths. Artifacts are relative to the supplied evidence root, cannot traverse parents or escape through symlinks, and carry SHA-256, build/environment, run ID and capture time. Evidence must be captured after contract approval and no later than the check observation; the observation cannot be later than the record. Future or timezone-less timestamps fail. The utility checks actual local file bytes, with a 256 MB artifact limit and 5 MB JSON limit.

FAIL or BLOCKED required results block validation. NOT_APPLICABLE requires both contract permission, a specific reason and an owner-approved exception declaration. Missing, extra or duplicate check IDs, reused paths, mismatched builds/runs, empty requirements and placeholder values fail. The template pairs deliberately remain blocked.

A pass always includes authorization_granted: false, external_truth_verified: false and identity_or_independence_authenticated: false. A human must inspect the evidence, verify the real approval and reviewer arrangement, and make any actual deployment/publication decision. Metadata and hashes cannot prove a website worked or a reviewer was independent. A falsified but self-consistent record can still pass structural checks.

Evidence artifact shape

A PASS check's evidence item contains path, sha256, build_id, environment, run_id and captured_at. Use an actual local relative path to an authorized artifact and its real digest; do not insert a fabricated value. The synthetic examples used by unit tests are generated in temporary directories and never represent Joseph's systems.

Testing and integrity

python3 -m unittest discover -s Tools/tests -v
python3 Tools/library_audit.py .

The current test suite contains 84 tests: the 24 baseline tests plus 60 new router/gate tests. Read Reports/discipline-unit-tests.txt and the current validation report for the executed results and boundaries. The modified integrity auditor also covers Publishing/ and validates all 26 charter/prompt/brief paths, source IDs, handoff IDs and baseline original hashes. Neither command is a live website or account audit.