5 of 5 seats affirm compliance; no seat finds a violation
Opus 5 (Anthropic)compliesread reasoning
This is an infrastructure change (schema, pipeline, protocols, UI) that adds a 'provisional' review state; it adds no content records and the tooling found no citation identifiers in added content lines, so there is no fabrication surface here. The new tier does not loosen §3.15's checks: `sourceExists()` admits a record only when a resolver-confirmed identifier or an answering URL carries metadata matching the record's title, a failing identifier blocks outright ("a failing identifier is a failing citation"), and 403/404/429/connection failures still block as before; the second reader is still asked to judge everything judgeable without the text (atomicity, relevance, independence), and what it refuses is refused. Honesty of labelling is preserved and arguably improved: a provisional source is forced to `verification: unverified` by the loader, `reviewState: provisional` must carry a block naming what showed existence, why the text is unread, the route and the run, and the loader rejects a block on any non-provisional/non-rejected record. Weightlessness is enforced structurally, not just in prose — `verifiedEvidence` excludes provisional records from the strongest-for/against selection, packet claim counts split `evidence` from `provisionalEvidence`, and edition-v9/check-v2 instruct grading them as absent, with UI badges on the evidence card, claim margin and source page. The remaining risk I note (a review-note-level concern, not a violation) is that an evidence record's verbatim quote enters unchecked; that is disclosed as provisional and marked unverified per §3.8, and re-verify promotes or refuses it. AGENTS.md is untouched and the DECISIONS entries are recorded founder directions, which §3.15 expressly permits.
GPT-5.6 Sol (OpenAI)compliescites §3.6, §3.8, §3.13, §3.14, §3.15read reasoning
The new tier is explicitly labelled `provisional`, forces its source to remain `unverified`, records the existence basis, retrieval failure, route, date, and admitting run, and does not represent unread material as verified. Provisional evidence is excluded from evidence counts, strongest-evidence selection, editions, and blind checks, while its public cards state that it carries no weight; this preserves the source/evidence distinction and calibrated uncertainty. Admission requires resolving metadata or a responding document title to match the source title, and a failed identifier blocks admission, so the real-citation check is not bypassed. Re-verification uses the existing second-reader path, records promotion or refusal in dispositions and append-only history, and leaves still-unread records provisional rather than silently upgrading them; no content citations, confidential material, or AGENTS.md amendment are added here.
Gemini 3.8 Flash (Google)compliescites §3.8, §3.14, §3.15, §4read reasoning
The change introduces a provisional intake tier and re-verification workflow for records whose sources demonstrably exist (via title-matching resolvers or verified HTTP 200 responses) but whose text cannot yet be extracted. All provisional sources are strictly enforced as 'unverified', provisional evidence and claims carry explicit provenance blocks and 'provisional' reviewState, and all assessment protocols (check-v2, edition-v9) and UI views explicitly exclude provisional records from evidentiary weight. No citations are fabricated, citation resolution checks remain fail-closed against failing identifiers, and the changes to AGENTS.md are zero. The research agenda lifecycle additions and plan gate further tighten rather than weaken epistemic standards.
Grok 4.5 (xAI)compliesread reasoning
The change implements founder-directed provisional admission (DECISIONS 2026-09-17) without editing AGENTS.md. Identifiers must still resolve with title-matching metadata (sourceExists; failing IDs block), the second reader still judges atomicity/relevance/independence without text, and provisional records are labelled, excluded from evidence counts, graded as absent in edition-v9/check-v2, and barred from strongest-evidence selection via verifiedEvidence—so fabrication and weight checks are not loosened. Re-verify promotes or refuses from materialised outcomes with stamped history; the research lifecycle and plan-v1 gate tighten agenda quality rather than weaken §3. No content citations were added, no confidence material is published, and provenance stamps (provisional block, statusBy/runId) remain honest.
GLM 5.3 Flash (Z.ai, via Venice)compliescites §3.8, §3.15, §3.13, §4read reasoning
The change does not edit AGENTS.md and does not weaken the fabrication check — it preserves and actually tightens it: a record enters provisionally only when an identifier resolves with metadata matching the record's title or a URL answers with a matching document title (sourceExists), a failing identifier blocks regardless of page title, a 403/404/429 or connection failure keeps the record blocked, and the second reader still judges everything judgeable without the text (atomicity, relevance, independence), refusing on those grounds. The weightlessness guard is enforced structurally, not just by convention: the loader throws if a provisional record lacks its block or a provisional source is not `unverified`; the packet excludes provisional records from evidence counts; check-v2 and edition-v9 grade them as absent; the case page selects strongest evidence from verifiedEvidence only; and every surface (badges, source page, margin notes) displays them as provisional and awaiting the text. This is consistent with §3.8's demand that unverified locators be labeled rather than silently treated as read, and with §3.13's calibrated-uncertainty ethos. The reverify verb promotes or refuses such records through the same second-reader checks, and the new plan gate and research lifecycle add checks rather than remove any. The founder direction of 2026-09-17 is recorded in DECISIONS.md with the session channel, satisfying the operating-record requirements of §3.15, and the diff's own review-note history (#322, #325-#327) shows prior objections were answered in code.
judged against constitution 12f4ae9ea8 · panel-v3 · verbatim record · pull request