The going-forward record of how sources were found — the reproducible-search half of the trustworthiness initiative (trustworthiness r9, Tier 2). One entry per literature search that feeds acquisition. Append-only; grep ^## \[ for the index.

Required contents — the four-field policy (binding, every lit-search) [maintainer 2026-08-05]

Every literature-research / paper-search episode — via bin/openalex.py, web search, or any other channel — records ALL FOUR fields here (mirror in CLAUDE.md -> Bias machinery -> SEARCH-LEDGER POLICY):

  1. Query — the terms, the channel (openalex / web / other), and the filters/date used.
  2. Results — the hits (DOIs/titles), AND the zero-hit queries — a 0 is data, and it is attributed (genuine literature absence vs a tool artifact, e.g. title.search requiring every word in the title). A silent 0 read as “no evidence exists” is the exact failure this ledger prevents.
  3. Adjudication + reason — per candidate SELECTED / REJECTED / DEFERRED, each with the why (newer-supersedes, OA-vs-paywalled, design tier, too-new/uncited, off-question). The negative space — what was rejected and why — is the load-bearing content, not a courtesy; asymmetric scrutiny hides here, so it must be visible.
  4. Acquisition-requested — which candidates were sent to acquire, and their fetch status (fetched to inbox/ vs blocked -> maintainer-drop). Never downgrade to a weaker source on a fetch fail — pause and ask (the acquisition-pause-and-ask rule).

Mechanical vs judgment split: openalex.py auto-logs fields (1)+(2) (the raw query line). A web/other-channel search has NO auto-log — log it by hand. Fields (3)+(4) are [judgment-only]: write a curated ## [YYYY-MM-DD] SELECTION RECORD — <topic> block every session; the raw query lines never stand alone. The auto-log proves what ran; the SELECTION RECORD proves how it was judged.

Acquisition-ENGINE runs (A/B/C/D/E) are tracked the same way, in three tiers. The five acq_*.py engines call oa._get() directly, so they bypass the per-query _log; instead each ends its run with oa.log_engine_run(...) (default on; --no-log to skip), appending ONE - [date] acq-engine <X> [params] -> <counts> | top: <dois> pool-tsv=<path> line. That is fields (1)+(2) for an engine: the params, the anti-join-resolved pool counts (candidates / held-dropped / rejected-source soft-matches), and the top shortlist. The FULL candidate pool lives in the byte-diffable --write tsv (acq-<engine>.tsv), which the line names — that is where “all candidates” persist durably. Promotions/rejections are the same SELECTION RECORD block (fields 3+4): a candidate from the pool becomes PROMOTED (acquired) / DEFERRED / REJECTED there, adjudicated against the mechanical line. So: engine line = what ran; tsv = the full pool; SELECTION RECORD = the per-candidate verdict.

Scope + honesty (read before trusting this log):

  • Going-forward only — but the back-catalogue is PARTIALLY recovered. This log starts 2026-07-31; searches before it were not captured here (source selection grew breadth-first — the trust page’s historical-looseness caveat). Those past searches are partially recovered by mining the .claude session transcripts: search-log-recovered.md (generated by bin/recover_search_provenance.py) links 40/129 held sources to the acquisition search that named them (07-25..07-31). It does not pretend to a clean retroactive trail — 23 links are DOI-confident, the rest are name-candidates, and ~89 sources are unrecoverable (dropped via inbox/, pre-transcript, or unnamed). Heal what’s healable; disclose the residual.

  • ChatGPT-sourced research is recovered too. Some sources came from research the maintainer ran in ChatGPT (a question posed, a candidate reading list returned), then assessed + DOI-checked here. That trail is captured in search-log-chatgpt.md (from bin/parse_chatgpt_research.py over the archived transcript in provenance/): 4 research questions, 68 candidates, 10 held, 58 surfaced but not taken — an honest identification -> screening record, including the candidates rejected.

  • The STAGED ingest queue is provenance-retrofitted too. search-log-stage-provenance isn’t where it lives — see stage-provenance.md (from bin/stage_provenance.py): all 156 staged (un-ingested) papers attributed across the three channels — Claude web searches 108 (+11 both), ChatGPT 20, reading-list 16 (method/books), 12 matcher-miss. Confirms most sources came from Claude searches, a chunk from ChatGPT, the method base from the reading list — almost none invented.

  • The EXCLUSIONS are tracked with reasons. rejected-sources.md — candidates assessed and NOT acquired, each with a brief reason (surrogate / GLP-1-saturated / superseded / redundant / peripheral / advisory-not-SR / off-scope), recovered from the actual 2026-07-30 acquire-reject decisions.

  • The Claude-side decision REASONING is recovered too — at two grains. (1) The one 2026-07-30 keystone sweep: keystone-acquisition-review.md, the Claude analog of search-log-chatgpt.md — 7-domain research sweep, 4 ranked gaps, a Tier-A/B catalog where every candidate carries a design/tier + a “fills/unblocks” reason. (2) The scattered day-to-day searches embedded in ingest/weave/query flows across the 6 days: search-log-reasoning.md (from bin/search_reasoning.py) — the assistant text that FOLLOWED each search (what was found + why taken/skipped: “WGO gold-standard guideline, freely available”, “De-Regil Cochrane folate->NTD RR 0.28”, “umbrella review OA but JS-gated”). Operational-grade, honestly labeled; ~58 of ~150 carried reasoning, 11 substantive after filtering mechanics.

  • Per-source ingest rationale is recovered — from the curated log, not mined windows. source-reasoning.md (from bin/source_reasoning.py): for every held domain source, its real “why” from log.md (the authoritative one-entry-per-ingest record — “domain-opener”, “pairing with GRADE”, “the second guidance family arrives”, “KEEP; predictions 4/4”), with a keystone-review fallback. 81/99; the founding 07-25..28 set is fully covered (they were never a gap — the ingest log recorded their rationale at ingest time). The remaining ~18 are recent cluster-logged batches, a naming-match limit not a gap. NB: mining transcript windows for per-source rationale was tried and rejected — it surfaced usage-mentions, not decisions (would have dressed a nearby mention up as the reason); the curated log is the honest source.

The two research channels are symmetric, and the trail is complete: searches (search-log-recovered = Claude · search-log-chatgpt = ChatGPT), reasoning (keystone-acquisition-review = Claude · search-log-chatgpt + triage = ChatGPT), inclusions (stage-provenance), exclusions (rejected-sources). The identification -> screening -> selection trail the initiative wants, kept light and honest.

  • Not a systematic-review search. These are targeted searches under the acquisition rule (reviews / meta-analyses / umbrella reviews / guideline bodies first), not exhaustive PRISMA searches. A log entry documents what was done, not a claim of completeness.
  • This is a discipline, not a gate [judgment-only]. Appended during acquire / plan queue / process inbox. Lightweight by design — the point is a truthful trail, not bureaucracy.

Entry format

## [YYYY-MM-DD] <topic / decision question being sourced>
- Searched: <databases / bodies> — <terms or strategy>
- Filters: <date / language / study-design limits, or "none">
- Found / screened: <n> / <n>
- Included: <registry slug(s) staged or acquired>
- Excluded (notable): <source> — <reason>
- Notes: <keystone found? none found -> named gap? tier of what was taken?>
  • Included names the canonical registry slug so an entry reconciles with sources.md and the via selection-provenance axis (a search-sourced entry is via = pick unless it came from a keystone bibliography — see trustworthiness §2).
  • Notes should say when a search returned no keystone-tier source for a live sub-topic — that is a coverage gap (type-G), a finding in its own right, not a licence to grab a weaker primary study.