The diff touches only pipeline logic (src/pipeline/verify.ts), its test, and an append-only DECISIONS.md entry; AGENTS.md is untouched and no content/ records or citations are added (verification found no identifiers). The change adds a bounded second split round to the anchorless branch, mirroring the existing anchored path, and every second-round part is still judged for ofTheCompound and atomicity — a strengthening of §3.2 enforcement, not a loosening of any §3.15 check; parts refused as not of their parent are still refused and noted in the account. Provenance is improved rather than degraded: the origin ref now names the immediate parent ('; second round, of the part "..."') and the reader prompt states which statement is the compound and that it is itself a part of the original, which the test asserts. Relations are explicitly not carried to parts, and the compound's rejection reason still lists the surviving part ids, so the ledger remains reconstructible. The DECISIONS entries record the discarded run honestly — spend records stand, proposal and edition do not — with date and channel stamped. I see no fabricated locator, no confidence material, and no reclassification to dodge a check.

The change strengthens atomicity handling by giving an anchorless compound claim the same bounded second split already used for anchored claims, while refusing second-round outputs that remain compound. It also corrects provenance by judging each second-round output against its immediate parent and recording that parent verbatim in the claim origin. The tests cover the two-round limit, rejection of a stray proposition from elsewhere in the original claim, relinking, and verification-account notes. No citations or confidential content are added, AGENTS.md is untouched, and the added decisions entries record the date, channel, and stated supervision rather than weakening publication checks.

The change extends the two-round claim splitting logic to anchorless claims that are judged against citing evidence statements, mirroring existing behavior for anchored claims. It strictly upholds atomicity (§3.2) by splitting non-atomic sub-parts once more (and no further) rather than dropping valid propositions. Exact provenance (§3.8, §3.14) is preserved and improved by tracking and stamping the immediate parent wording in candidate origins and verification contexts. Automated pipeline tests are thoroughly updated, and no constitutional rules or verification gates are weakened.

The diff only touches pipeline split logic, its tests, and DECISIONS.md. It extends the existing second-round compound split from the anchored path to the anchorless path (depth < 2, then stop), and stamps second-round origins/context with the immediate parent so ofTheCompound is judged against that parent—tightening §3.2 handling rather than weakening §3.15 checks. AGENTS.md is untouched; no citations, confidence material, run records, or fabricated provenance appear. Founder-supervised decision entries match the code change and are allowed under §3.15.

This is a pipeline-code fix with tests and a decisions-log entry; it edits no AGENTS.md, adds no citations, and carries no run records, so the fabrication, confidence, and constitution tripwires are not in play. Substantively the change strengthens §3.2 rather than weakening checks: it extends the one-recursion split (once more and no further) to the anchorless branch that previously refused everything, and it improves §3.14/§3.8 provenance by recording the immediate parent in each part's origin so a part from elsewhere in the original compound cannot pass silently — the new test exercises exactly that refusal. The DECISIONS.md entries record a founder direction with date and channel ('in session, 2026-09-21'), which is the form §3.15 permits for the operating record, and the same supervised-by pattern already appears in the pre-existing context lines, so I have no positive evidence the attribution is invented; if it were, that would be a §3.15 provenance problem, but nothing in the shown material contradicts it. No dangling IDs, no fabricated locators, and the tests assert the account text the runs must produce, so I see no rule broken.