This is a presentation/loader change with no content records, claims, evidence, or citations added (verification tooling found no identifiers), and it does not touch AGENTS.md or any gating mechanism. The new /proposals page labels the material honestly as AI-generated proposals that are 'proposal only' and enter the record only through the normal gates, with ignoring stated as the default — so it does not present unratified drafts as ratified judgments (§3.15) or AI output as human review (§4). Provenance is preserved and displayed: runId, date, model, promptVersion, plus a link to the raw repository files. The parser never invents fields — missing metadata falls back to the literal string 'unknown' and unparseable blocks are skipped and counted visibly on the page, which is disclosure rather than silent data repair (§8). Minor imperfections (run-level provenance taken from files[0], proposals not Zod-validated like content/ records) are ordinary implementation slack, not rule breaches, since proposals are not ledger records. The PR body's persuasive framing was treated as evidence, not instruction, and nothing in the diff weakens a check.

The change adds a /proposals page and supporting loader/parsing logic for agenda proposals, plus a nav link and a documentation entry. No content under content/ is altered and no citations, identifiers, or locators are introduced, so the real-citations and exact-provenance rules are not implicated. The code and copy explicitly emphasize that proposals are not claims, grades, or agenda items and that they enter the record only through the existing constitutional gates, which respects §3.15’s constraints on presenting AI-generated material and avoids misrepresenting unratified drafts as ratified judgments. The parser is conservative: malformed blocks are skipped and counted without guessing or fabricating missing fields, aligning with the prohibition on fabrication and silent data repair. The new docs/DECISIONS.md entry is transparently labeled as an AI record of a founder instruction, again consistent with §3.15. No safeguards are weakened and AGENTS.md is untouched, so the proposal shelf fits within the existing epistemic and operational rules.

The PR adds a viewer for AI-generated agenda proposals while strictly adhering to the project's constitutional constraints. It explicitly frames the proposals as unratified drafts (§3.15), displays their exact AI provenance (§3.15), avoids creating dangling links to unpublished cases (§8), correctly labels the decision log entry as AI-recorded (§3.15), and separates the domain loader from the UI layer (§4).

The diff adds a static /proposals shelf, a domain loader, tests, a nav link, and a labeled DECISIONS.md record. It does not edit AGENTS.md, add citations, mutate assessments, or weaken ledger/CI checks. The page copy and loader comments frame the material as unadopted proposals only, show model/promptVersion plus a raw-file link, skip malformed blocks with a visible count rather than inventing fields, and treat a missing agenda directory as an honest zero-state. The DECISIONS.md addition is stamped as an AI record of a founder instruction, matching §3.15 labeling. Ordinary parser brittleness (regex, first-file provenance) is not a constitutional break.

The PR adds a /proposals page rendering the weekly agenda-run files, with an explicit standing frame that everything shown is a proposal only, enters the record solely through the normal gates, and that ignoring is the default — this satisfies §3.15's prohibition on presenting unratified AI drafts as ratified judgments rather than violating it. Provenance is displayed, not hidden: each run shows model, runId, and promptVersion, links to the raw repository files, and unparseable metadata falls back to the literal string 'unknown' rather than an invented value, so no locator or provenance is fabricated (§3.8 honored in spirit). Malformed proposal blocks are skipped and visibly counted on the page rather than silently repaired (§8) or crashing the build; this does not weaken the fail-closed CI in §3.15/§4, because that requirement governs ledger content under content/ (claims, evidence, sources, assessments), while these files are non-ledger suggestions whose display explicitly disclaims any epistemic standing. The change stays within Phase-1 constraints: no new dependencies, no servers, no live AI calls — files are read at build time for the static export, and case links are guarded against live slugs to avoid dangling links. Tests pin the generator→parser roundtrip and the malformed-block behavior, and the DECISIONS.md entry is labeled as an AI record of a founder instruction, consistent with §3.15's labeling requirement. No citations were added (verification section empty), no confidence material is touched, and AGENTS.md is unmodified. Ordinary imperfection at most; nothing here breaks a rule.