No AGENTS.md edit and no changes under content/, so no citations were added (verification tool confirms none) and no published assessment or provenance was altered — the synthetic test even asserts `loadCase(dir).contentHash` is unchanged after a run. The new `research-case.ts` path is an explicitly manual CLI producing an intake artifact labeled 'Unverified AI research report' and 'unverified working report', with docs stating no automatic adoption, no schedule, and that proposals still pass source checks and the normal publication gate — consistent with §3.15's ban on presenting unratified drafts as ratified. The prompt instructions actively restate the constitution's rules (primary sources first, exact locators only when inspected, never invent provenance, distinguish source statement from inference, avoid forced binaries), so the tooling reinforces rather than weakens §3.7–§3.9. The relaxations in `openai-response.mjs` (search-call count check changed from `=== 1` to `some(...)`) and `metered-model.mjs` (multi-pass reservation) are budget/transport mechanics, not epistemic checks, and the reservation is made strictly more conservative ($18.88 held, settled by actual usage; arithmetic in config/README and the test both check out at $1.88/$1.83). New schema validation in ai-policy.mjs adds constraints (approved model, tariff, tool-count ceiling) rather than removing them. Phase-1's ban on live AI calls concerns the site itself; this is an offline script alongside the existing metered workers.

The new research report is consistently described and stored as unverified working material, not as canonical evidence or a ratified assessment. Its prompt requires primary sources where available, honest access labels, exact locators only when inspected, separation of source statements from inference, competing explanations, and no invented provenance. The workflow preserves request, response, model, runId, prompt version, timestamps, citations, failures, and reconsideration reasons in append-only intake history before any source verification or publication review. It does not modify case content, automatically promote citations, or bypass multi-model publication gates. Hosted research remains manually invoked, bounded by explicit tool, output, spend, and timeout limits, with fail-closed policy validation and no weakening of the existing publication checks.

The PR introduces an offline research investigation script and associated configuration, metered OpenAI transport calls, and unit tests. It does not modify AGENTS.md, nor does it alter any canonical published content under content/ or present AI-generated working reports as ratified editorial conclusions (§3.15, §4). Output from research runs is stored explicitly as unverified working material in intake history, preserving strict provenance and requiring standard downstream verification and publication gating before any claim or evidence reaches the site ledger.

The diff adds only a manual, metered research-report path that writes unverified working material into existing intake history and local .research-runs receipts; it does not edit content/cases, publish assessments, or present drafts as ratified. Instructions and stamps explicitly require separate source checks, forbid invented provenance, label reports unverified, and rest unchanged/failed runs unless a reasoned --reconsider is given. Shared budget limits, fail-closed tariff/policy gates, and the multi-model publication gate are preserved rather than weakened; AGENTS.md is untouched. No citation identifiers were added under content/, and the PR’s docs/DECISIONS entry records founder-directed scope without reclassifying checks. This is tooling and accounting for a bounded trial, not a unilateral epistemic publication.

The change adds a manual, script-invoked research tool (not the site itself, consistent with §4's 'no live AI calls in the site'), and every output is explicitly labeled an 'unverified working report' that never publishes, adopts citations as evidence, or alters canonical case content — the test even asserts contentHash is unchanged. Provenance is well handled: requests are persisted before sending with runId, promptVersion, inputHash, and model stamps in the append-only intake history, satisfying §3.14 and §3.15's gated-AI requirements; the publication gate and separate source checks are preserved rather than weakened. The metered-model change broadens webSearchCalls from 1 to a policy-capped 8, but it is a guarded capability extension with proportionally scaled reservations ($18.88 = 9×200k input + 20k output + 8 tool fees, arithmetic verified) within unchanged $150/$25 limits, and new schema validation blocks unapproved models/tool counts — not a weakening of §3 checks. No content/ citations are added (verification section is empty), and the o3-deep-research tariff is a post-training fact I cannot independently confirm, but the pricing is internally consistent and sourced/checked-at dated, so under temporal humility I do not treat it as fabrication.