Constitution

Taste, compiled. This document is the standard the audit population (gatewatch, ratchetwatch, tastewatch — see .warren/triggers.yaml) measures merged work against. Auditors cite articles by number when filing findings. The quality gates in CI enforce what they can; these articles cover what executable gates cannot — and every article aspires to become a gate.

Provenance: distilled 2026-06-11 from CLAUDE.md conventions, mulch records, and a forensic audit of merged history (the PR #270 title/diff mismatch, the silent 91.62% → 90.82% coverage decay inside ratchet slack, the v0.7.1 supervisor-entry production break, four files grandfathered at birth on 2026-06-06, and 8 releases in 10 days carrying mostly comment fixes).

Article I — Merges are truthful

A PR’s title and description describe its diff. A feature-titled PR contains the feature. A merge is never rationalized past a red gate (“this failure is unrelated” is a finding, not a waiver — fix the gate or fix the change). If work is blocked, the PR says it is blocked and carries no feature title.

Article II — Ratchets only tighten

Coverage floors rise to track actuals; slack above ~0.75pt is debt. Budget and grandfather exceptions carry a tracker reference in the diff that adds them. Nothing is grandfathered at birth: a new file written over the size limit is decomposed before merge, not exempted at write time. Automated budget raises (autoheal) stay within their caps and are reviewed in aggregate, not ignored individually.

Article III — Releases mean something

A published release contains at least one consumer-observable change. Internal hygiene (comment fixes, doc drift, test tightening) batches into the next real release rather than minting its own. Patrol plans release conditionally, not ceremonially.

Article IV — Tests verify behavior

Coverage gains come from meaningful assertions, not test theater. Aggregation, classification, and parsing logic ships with adversarial cases — malformed input, boundary values, the case that broke last time — not happy-path only. A test that asserts a mock was called is documentation, not verification.

Article V — Comments state constraints

Per CLAUDE.md: a comment exists to state a constraint the code cannot show. Institutional memory belongs in mulch records, not in 200-word JSON $comment essays. Narration of the obvious is noise and gets removed, not preserved.

Article VI — Refactors prove runtime paths

Any file move or rename verifies references that tests do not import: Dockerfile entrypoints, supervisor spawn paths, workflow YAML, config strings, docs. The check is a repo-wide search for the old path, including non-source files. “All gates green” is not “deploy works.”

Article VII — Identity is consistent

Agent-authored commits use canonical co-author identities. One agent, one spelling. An identity that appears once and never again is a finding.

Article VIII — Evidence or it didn’t happen

Findings cite commit SHAs, file paths, and line ranges. A seed filed without evidence gets closed without action. Auditors that cannot show their work do not file.

Article IX — The constitution outranks the population

Changes to this file, to auditor prompts (the agent entries for gatewatch / ratchetwatch / tastewatch — today agents/*.md, historically rendered under .canopy/; the .canopy/ references here are struck in the phase-C canopy sweep, warren-6fcd), or to .warren/triggers.yaml audit entries require explicit human review — they must not ride an auto-merged PR. Any auditor that observes a merged change to these files without human approval files a priority-1 finding citing this article. The population does not rewrite its own mandate.

Executable form: the “Article IX check” step in .github/workflows/auto-merge.yml refuses to enable auto-merge on any PR touching this file, .warren/triggers.yaml, .canopy/, or that workflow itself. The auditors still verify (the gate can be deleted; the deletion is itself a protected change).

Amendments

Amend by PR touching this file, flagged for human merge (Article IX). Tastewatch’s weekly digest may propose amendments; it may not apply them. When an article becomes fully enforceable as an executable gate, note the gate here and retire the manual check.

---

Audit Warden boundary (warden architecture)

The Audit Warden consolidates findings from gatewatch, ratchetwatch, and tastewatch into a weekly digest. This section records the constitutional constraints.

Retired (warren-e7e7 / pl-3a79): the warden was formerly a standing Leveret conversation bound to a long-lived meta-Plot, fed over POST /conversations/:id/messages and re-woken via POST /conversations/:id/re-wake. The conversations subsystem is being removed; the warden now operates entirely over seeds issues. All /conversations* references below have been struck.

Ingestion rule

Audit findings are recorded only as seeds issues via the sd CLI (labeled audit plus the auditor’s own label). No auditor writes directly to the DB or calls a new endpoint. The seed is the durable, auditable record — this preserves Article IX: auditors cannot modify their own mandate, and every finding is as reviewable as any other seed.

Digest cadence

• warden-digest cron: Sunday 05:00 America/Los_Angeles — the digest driver triages the week’s accumulated open audit-labeled seeds and files one consolidated digest seed (and, where mechanical, a plan).
• tastewatch-digest cron: Sunday 04:00 America/Los_Angeles (60 min before warden-digest) — tastewatch files its taste digest seed first so warden-digest can synthesize it alongside the other auditors’.

Both trigger entries are in .warren/triggers.yaml and are protected by Article IX (require human review to change).

Autonomy-promotion recommendations

When tastewatch’s precision table shows consistent, high-precision findings from an auditor, tastewatch recommends autonomy promotion in its digest. These recommendations are advisory — they are recorded in the digest seed for a human to route as a proposed seed or amendment, and any resulting change to .canopy/ or .warren/triggers.yaml requires explicit human review (Article IX). No auditor may grant itself or another auditor autonomous dispatch authority.