This is an operations/cost-control change: no files under content/ are touched, no citations, evidence records, assessments, or published judgments are added or altered, and the citation tooling found no identifiers in added content. AGENTS.md is untouched, so §3.15's reservation to the founder is not engaged. The panel machinery is preserved rather than weakened — scripts/lib/vendors.mjs keeps all five independent seats and their models, no concurrence threshold or publication gate is changed, budget exhaustion fails closed (stops paid work) rather than allowing unratified publication, and a $30 monthly reserve is explicitly held for review calls. Provenance discipline is if anything strengthened: llm.mjs now rejects silent cross-vendor substitution and keeps the refusal-fallback that stamps the model that actually answered, and metered-model.mjs verifies the returned model id before settling. The new inbox files are labeled AI-prepared discovery leads with explicit runId/model/promptVersion stamps and repeated instructions to establish object identity, rights, and source statements before any claim reaches the ledger — they assert no findings. The operator budget proxy binds to 127.0.0.1 on the Actions runner and is documented as not part of the static export, so §4's 'no servers' rule for the site is not breached; the self-authorized bootstrap policy is narrowly scoped (valid only before config/ai.json first appears in main) and disclosed in DECISIONS.md rather than hidden.

The five Birdmen entries are explicitly labeled AI-prepared discovery leads rather than verified findings, and they direct later workers to establish object identity, source chains, rights, dependencies, and source-versus-inference distinctions before publication. They include model, date, runId, promptVersion, and authorship provenance, while adding no Claim, Evidence, Source, or Assessment record that falsely carries these links as verified evidence. The budget implementation reserves liability before each paid call, retains uncertain reservations, records receipts publicly, and fails closed when accounting or tariffs are unavailable; the diff does not alter panel identities, concurrence thresholds, or publication gates. No AGENTS.md amendment, confidential-material disclosure, fabricated locator, silently revised assessment, or positive internal evidence of a false citation is present.

The pull request introduces shared cost controls, switches the main authoring client to Astra, meters API requests (including the operator and review seats), and adds five public research leads to the inbox directory. No files under content/ or the core constitution AGENTS.md are modified. The five-vendor independent review panel and fail-closed publication checks under §3.15 remain completely intact, with review budget explicitly protected and metering errors set to fail closed. Provenance is cleanly recorded across decisions and inbox files, and all unit tests and schema validations pass.

The diff is complete and does not touch AGENTS.md, published case ledgers under content/, assessments, or concurrence/publication gates. It adds shared metered spend controls, model policy in config/ai.json, workflow wiring, tests, and docs; exhaustion stops paid work without lowering review independence or thresholds (§3.15). Five inbox items are founder-directed link leads for deep-memory, explicitly labeled AI-prepared discovery context with provenance stamps, not ratified findings or confidential material. Citation tooling found no new content/ identifiers to verify; no positive sign of fabricated locators, silent judgment edits, or check-weakening.

The diff does not touch AGENTS.md, and it explicitly preserves the check architecture: all five independent panel seats keep their models and concurrence rules, no publication or ratification threshold changes, and failures (missing usage, exhausted allowance, unavailable accounting) stop paid work fail-closed rather than bypassing gates — consistent with §3.15's prohibition on weakening checks. The Birdmen inbox files are labeled AI-prepared discovery leads with explicit AI provenance stamps ('Eugene supplied the direction; the wording is AI-written'), presented as unverified intake rather than ratified findings or human-reviewed conclusions, satisfying §3.9/§3.15 honesty requirements; they add no citations to content/ (the citation verifier found none), and the URLs in config are tariff-source metadata, not evidence citations. The bootstrap mechanism is time-limited, publicly recorded on a dedicated branch, cannot resurrect itself after model policy appears in main history (tested), and governs financial accounting rather than editorial ratification; the founder-authorization claims in the PR body are recorded as AI records of founder instruction, not passed off as human ratification. Model names like 'gpt-6-astra' postdate my knowledge, but per temporal humility unfamiliarity is not evidence of fabrication, and the citation tooling raised no failures. Nothing here fabricates provenance, publishes unratified drafts, exposes confidence material, or dodges a check, so ordinary infrastructure imperfection would not warrant 'violates'.