The change is pipeline/tooling only: a new `archiveItemOf()` helper, a branch in `retrieve()` that tries the Internet Archive's `_djvu.txt` OCR layer before the item page, a unit test with a fake fetch, and an append-only DECISIONS.md entry. No content/ records are added, so there are no citations, locators, or assessments to fabricate — the verification section found no identifiers, consistent with the diff. AGENTS.md is untouched and no §3.15 check is loosened; on the contrary, the OCR text is returned under the item page's key with `substitute: true` and an explicit `via` string naming the exact URL read, which preserves honest provenance rather than passing OCR off as the page image, and the fallback path leaves `via` undefined when only the page is read. The DECISIONS entry is appended, dated, and stamps the founder direction with channel and quote, as §3.15 permits for operating-record directions. The only wrinkle is a numeric mismatch of flavor text — the code comment says "eight blocked routes" while the PR body and log entry cite 176/139/37 pool counts — which is a documentation imprecision about internal ledger tallies, not an invented citation or provenance stamp, and I would want it reconciled in a later pass but it does not rise to a violation.

The retrieval preserves the Internet Archive item page as the source key while recording the exact OCR URL in `via` and marking the OCR as a `substitute`, which honestly distinguishes the source from the text representation used. Failure to obtain the OCR falls back to the item page rather than silently treating missing OCR as contradictory evidence. The tests cover details, download, and stream URL recognition, successful OCR substitution, and the fallback path. The change neither edits AGENTS.md nor weakens ratification, citation, verification, or provenance checks, and it adds no content citations requiring mechanical verification.

The change enhances retrieval of Internet Archive scans by fetching their plain-text OCR layer when available, correctly marking the result with `substitute: true` and recording explicit provenance in `via`. It modifies no content files or AGENTS.md, adheres to pipeline transparency and exact provenance requirements, and includes unit tests verifying fallback behavior. The append-only record in docs/DECISIONS.md properly documents the rationale and founder supervision.

The diff only extends retrieval to prefer Internet Archive _djvu.txt OCR under the item-page key with substitute:true (and falls back to the page), plus tests and an append-only DECISIONS.md entry. AGENTS.md is untouched; no content/ citations, confidence material, or run records are introduced. Marking OCR as a stand-in matches exact-provenance honesty rather than weakening checks. No positive evidence of fabrication or misattribution.

This is a retrieval-infrastructure change with no citations added, no AGENTS.md edit, and no run records, so the fabrication and constitution rules are not in play. The OCR-text route is handled honestly: the fetched text is keyed to the item page's URL, flagged `substitute: true`, and the `via` field discloses exactly that the Internet Archive OCR text was read in place of the item page — so a quote the OCR cannot support will be blocked rather than passed off as verified, consistent with §3.8's unverified-locator labeling and §3.9's source-statement-versus-inference distinction. The `_djvu.txt` URL pattern is the Archive's real convention, not an invented locator. The DECISIONS.md entry records a founder session direction in the form the constitution permits for operating-record directions, and contains no internally inconsistent provenance I can see; I cannot independently confirm the session quote, but temporal humility says unfamiliarity is not evidence of fabrication. The tests cover both the OCR path and the fallback with a fake fetch, and nothing here loosens a §3.15 check.