The diff adds a maintenance-only step that writes model-generated *proposals* to `proposals/agenda/<runId>/`, never to `content/`, so no claim, evidence, source, or assessment record is created or altered. Every proposal file is headed 'PROPOSALS ONLY' and stamped with model, runId, promptVersion, and date, explicitly invoking §3.15 — no unratified draft is dressed as a ratified judgment (§3.15) and provenance stamping is preserved (§3.14). Validation is fail-closed: proposals must anchor to IDs actually present in the case files (dangling anchors rejected, matching §8's no-dangling-IDs discipline), malformed replies are dropped with recorded reasons, and a model-call failure skips the case rather than emitting anything. No citations, DOIs, or locators are added, so §3.7/§3.8 are not engaged (verification tooling found no identifiers). Nothing weakens an existing gate: the diff only appends a step to the workflow and a digest line; the classifier/arbiter/merge policy are untouched, so §3.15's anti-check-weakening clause is respected. AGENTS.md is not modified, and the DECISIONS.md entry honestly labels itself as an AI record of a founder instruction. The Phase-1 ban on live AI calls concerns the site itself; this is an offline CI script alongside existing triage/reassess model steps, and it does no web research.

The change adds an automated weekly agenda-proposal step that is explicitly scoped to "proposals only" and writes to a separate proposals/agenda/ directory, never modifying published content, respecting §3.15’s requirement that AI drafts not be presented as ratified. The rendered proposal files contain a clear disclaimer that nothing inside is a real claim or agenda item until adopted through existing gates, and they record model, runId, and promptVersion, aligning with append-only, labeled AI records (§3.14, §3.15). The PROPOSAL_SYSTEM prompt reinforces atomicity, asking for one proposition with a clear truth condition for new claims (§3.2), and pushes toward decisive, computable items, which is compatible with the evidence-led ethos. The validator is fail-closed and enforces anchoring to existing IDs, limiting at most three proposals per case and rejecting malformed or dangling references, which supports independence tracking and avoids hidden duplication (§3.10). No citations, locators, or confidential materials are fabricated or altered; the DECISIONS.md entry is a labeled AI editorial record of a founder directive, as §3.15 requires. Tests cover the validation and rendering logic, and the workflow wiring only surfaces a digest in the maintenance PR body, without granting new publication powers or weakening existing checks.

The change correctly isolates the generative AI call to the offline maintenance pipeline rather than the site itself (§4), and enforces fail-closed mechanical validity by explicitly rejecting dangling existing IDs (§8). Furthermore, it perfectly satisfies §3.15 by prepending generated proposals with a clear unratified draft warning containing runId/model metadata, and correctly labels the newly added `docs/DECISIONS.md` entry as an AI-generated record of a founder instruction. (The reserved human-only edit applies to `AGENTS.md`, which was untouched).

The diff adds a maintenance-only propose-agenda step that writes labeled draft files under proposals/agenda/, not under content/, and does not edit AGENTS.md. Proposal files are stamped with model/runId/date/promptVersion and headed as non-claims pending existing gates, which matches §3.15’s ban on presenting unratified drafts as ratified judgments and does not weaken ratification, CI, or freeze discipline. Validation is fail-closed (real existing IDs, capped kinds, empty answers allowed) and no citations, locators, or confidential material are introduced. The new DECISIONS.md note is labeled as an AI record of a founder instruction rather than a silent constitution change. Live model calls stay in the existing GitHub Actions maintenance job, not on the static site, so phase-1 site constraints are intact.

The change adds a maintenance step whose output is explicitly quarantined: proposals land under proposals/agenda/, each file headed with a nothing-here-is-real-until-adopted notice and stamped with model/runId/promptVersion/date per §3.15, and the DECISIONS.md entry is labeled as an AI record of a founder instruction. Nothing touches content/, no citations are added (the verification section is empty), no published assessment is altered, and the validator is fail-closed — proposals must anchor to real existing IDs, malformed replies are dropped with recorded reasons, and empty output is treated as success. Adoption is explicitly routed through the existing gates (founder judgment, freeze discipline, classifier, arbiter), so no check is weakened and no consequential change publishes on one model's judgment; the workflow merely appends a digest section to the existing weekly maintenance PR. AGENTS.md is untouched. The design — model proposes, gates decide, taste stays human — is consistent with the constitution's declared AI-operation experiment rather than a dodge around it.