The diff touches only pipeline code (`src/pipeline/answer.ts`, `src/pipeline/reverify.ts`), its tests, and an append-only DECISIONS.md entry; AGENTS.md is untouched and no content/ records or citations are added (verification tool found no identifiers). The change tightens rather than loosens the checks: records relinked to the parts of a split claim are no longer left weightless-but-lingering until a seven-day re-verify cadence, but are read against the parts by the second reader in the same sitting, promoted only with the links the passage actually bears, refused if it bears on none, or left `provisional` when the text will not come — which is exactly the §3.5/§3.2 discipline (bearing is read, not inherited) and preserves the provisional-carries-no-weight rule. Failure handling rolls back the ledger writes of that pass alone and the account states the outcome either way, so provenance and history entries record what was read, by whom (`aletheia answer (<model> second reader), on #<pr>`), and why. The DECISIONS.md entry is appended, dated, and stamped 'Supervised-by: founder (direction in session, 2026-09-20)', consistent in form with the surrounding entries; nothing presents an AI act as a human conclusion. No fabricated locator, no confidential material, no reclassification to dodge review.

The change makes relinked evidence provisional until its source passage is read against each newly split claim, then keeps only the links the reader judges supported, satisfying explicit evidence direction and source-versus-inference handling. Unread records remain provisional and therefore weightless, while records bearing on no part are refused with a recorded reason. The implementation adds actor-stamped history, run-account reporting, and rollback tests rather than silently overwriting or losing the earlier assessment. It adds no citations or confidential material, does not edit AGENTS.md, and does not visibly weaken or bypass the ratification and validation checks.

The change implements `readRelinked` within the answer pipeline to immediately evaluate evidence records relinked after a claim split against the resulting parts, rather than leaving them provisional until a later cadence pass. This directly upholds §3.5 by ensuring evidence direction and relevance are read and verified against atomic sub-claims rather than inherited. All modifications preserve provenance stamps, rollback semantics on write failure, and transparent recording in history without touching AGENTS.md or weakening verification gates.

Pipeline-only change: settleRecords reports relinked/relinkedTo, and readRelinked re-reads those live evidence records against the split parts in the same sitting before the edition, promoting only links the second reader finds, refusing none, or leaving provisional when text will not come. That enforces §3.5 (bearing read, not inherited) rather than weakening checks; history/actor stamps and rollback-on-failure preserve provenance. DECISIONS.md gains a founder-supervised entry; no AGENTS.md edit, no content/ citations, no confidence material, no fabricated records. Tests cover promote/refuse/unread/rollback and the answer ordering.

This is a pipeline-code change with no content citations (verification confirms none added), no AGENTS.md edit — the DECISIONS.md append is the sanctioned decisions log, dated and channel-stamped like its neighbors. Far from weakening §3.15's checks, it strengthens them: records left weightless-provisional after a claim split are now actually read (source fetched, second reader judges which parts each passage bears on) before the edition, and the provisional rule's substance is preserved — if the text will not come (HTTP 503, torn socket), the record stays provisional and is reported as 'Still unread', never promoted on unread text; refusals carry reasons on the record, history entries say what was read and why, and failed passes roll back their ledger writes (tests assert byte-identical restoration). Evidence direction per part is judged explicitly rather than inherited, honoring §3.5, and promoted records keep AI stamps (ai_extracted) rather than implying human review per §4. The edit to verification workflow timing (same sitting rather than a seven-day-cadence later pass) does not dodge any check — the check is that unread text carries no weight, and that is enforced throughout; on #372 the split parts remain held on their anchors alone pending the re-telling, as the body states. I saw no fabrication, no confidence material, and no unratified-draft-as-ratified presentation in anything visible.