This is a tooling/automation change with no content/ records, no citations added (verification section confirms none), and no edits to AGENTS.md (the only AGENTS.md write is a synthetic fixture inside a temp dir in discovery.test.ts). The new discovery adapter explicitly does not create Source/Claim/Evidence records: it queues leads into the existing source-reading queue, and docs and the PR body state repeatedly that search summaries are unverified leads and that ledger additions still pass the ordinary independent publication gate — so §3.15's gating is preserved, not weakened. Anti-fabrication is actively strengthened: searchSources admits only URLs from a completed web_search_call receipt or annotations, rejecting URLs appearing solely in model prose, and DiscoveryRunSchema refuses selections whose indexes do not name a distinct returned source; invalid selections are nulled rather than queued. Provenance and accounting are recorded (runId, model, promptVersion bounded-discovery-v1, basisHash, immutable append to the intake store; costs reserved with case/run/phase terms and an explicit, dated web-search tariff). Prompts forbid inferring absence from a search miss and require a counterevidence query, consistent with §3.11 and §3.13; the docs likewise remove the old self-graded 'saturation counter'. The docs/DECISIONS.md entry and PR body honestly disclose that live validation follows merge and that promptVersion is 'none' for the interactive session, which is candid rather than misleading. Phase-1's ban on live AI calls/autonomous web research applies to the site itself, not the maintenance scripts, and the static export is untouched.

The change adds a bounded discovery adapter rather than publishing claims or evidence, and explicitly routes all selected leads through the existing source-reading checks and independent publication gate. It freezes a narrow question, inclusion criteria, and disconfirmers before searching, requires counterevidence when searches are planned, and states that empty or failed searches are not negative evidence. URLs can be selected only from completed search-tool receipts, while prose-only URLs, invalid indexes, stale inputs, and failed calls are rejected or retained as labeled failures. Discovery records preserve model, response ID, hashes, run ID, protocol version, date, search actions, and selection rationale, providing appropriate provenance without representing leads as verified sources. The diff does not amend AGENTS.md, alter published assessments, expose identified confidential material, or weaken the ratification and validation gates.

The proposed change adds a bounded source discovery script and pipeline for the research intake workflow without altering any content files or published ledgers directly. In accordance with §3.12, the discovery process freezes the scope, question, inclusion criteria, and disconfirmers prior to executing any search queries, and mandates a counterevidence search in line with §3.11. Discovered search results are strictly treated as unverified leads rather than evidence, requiring URLs to originate in verified tool receipts and feeding into the existing intake queue rather than bypassing the reading and publication gates (§3.6, §3.15). No live AI calls or autonomous research are added to the static website runtime itself, respecting the Phase-1 boundary (§4).

The diff adds bounded discovery automation (scripts, config, workflow, docs, tests) that freezes one question with inclusion criteria and disconfirmers before at most two single-call searches, admits only URLs from completed search receipts (not model prose), and queues leads for the existing reader and independent publication gate. No content/, assessments, evidence, or AGENTS.md changes appear; search outputs are explicitly non-evidence with immutable intake receipts and no self-triggering. Hosted-search tariffs and shared-budget metering strengthen rather than weaken §3.15 checks, and citation tooling found no fabricated content identifiers.

The change adds a bounded discovery adapter, not a content change: no files under content/ are touched, no ledger additions occur in this PR, and every discovered URL must come from a completed web_search_call receipt (searchSources rejects prose-only URLs, uncompleted actions, and replayed outputs), so nothing is fabricated and §3.8's 'never invent a locator' is enforced in code. Search summaries are explicitly recorded as unverified leads that only the existing source reader and independent publication gate can convert into Source/Claim/Evidence records, preserving the gated multi-model process and append-only intake receipts (§3.14, §3.15); counterevidence queries are schema-required, and 'failed search ≠ absence' is encoded rather than asserted. The webSearch tariff is an added, metered cost term with fail-closed reservation/settlement and case/run/phase attribution — it strengthens rather than weakens the checks, and AGENTS.md is not edited, so §3.15's amendment reservation is not implicated. The only unverifiable items are two OpenAI documentation URLs in config/README.md, which fall outside the content/ citation-verification scope and are auxiliary tariff-source notes, not evidence citations; unfamiliarity with them is not positive evidence of fabrication, so they do not overturn an otherwise compliant, well-bounded change.