The diff touches only the loader's integrity checks, a pipeline cascade, tests, and an append-only DECISIONS.md entry; AGENTS.md is untouched and no content/ records or citations are added (the verification section finds no identifiers). The relaxation is narrow and principled rather than a dodge of §3.15's fail-closed checks: `requireClaim` still demands that the referenced claim exist, so dangling IDs remain errors everywhere, and only frozen, append-only artifacts (tombstoned evidence with reviewState "rejected", assessment-run overlays stamped to a ledger state, predecessor editions or a head edition whose basis hash no longer matches the current ledger) may point at a claim rejected since. Live evidence, research items, images, and the current edition on the current ledger stay strict, and when no ledger hash is available the head edition defaults to strict — a fail-closed default, not a loosened one. This serves §3.2 (a compound claim can actually be replaced by anchored parts) and §3.14 (histories and overlays keep saying what was true when written) rather than weakening them; the reverify addition extends the same cascade to research items and images, keeping live records clean. The DECISIONS.md entry records the change with date, channel, and the triggering run id, consistent with §3.15's stamping requirement; I have no positive evidence any of that provenance is invented. Tests cover both the frozen and strict branches, so the behavior change is exercised.

The change makes a coherent distinction between append-only historical records and live records: tombstones, prior assessment overlays, and editions tied to earlier ledger states may retain references to claims rejected later, while current evidence, research items, images, and a current-ledger edition remain subject to the live-claim check. That preserves reconstructable histories under §3.14 and permits compound claims to be replaced by atomic parts under §3.2 without making earlier records unloadable. Unknown claim IDs still fail in every context, and the settlement cascade now removes or redirects refused-claim links in live research and image records. The tests specifically cover both the historical exception and the fail-closed behavior for current records, so this is a narrowing of an overbroad integrity rule rather than an apparent dodge of §3.15.

The change distinguishes frozen historical records (append-only assessment overlays, tombstones, and past editions) from live records when validating claim references. This directly supports §3.14 (append-only history and visible revision histories) and §3.2 (allowing compound claims to be replaced by atomic parts without breaking historical runs that assessed the original compound claim). Live records and the current edition on the current ledger remain strictly bound to live claims, and unknown IDs remain disallowed everywhere.

The change correctly distinguishes frozen historical records (tombstones, assessment runs stamped to a past ledger, editions whose basis no longer matches the current ledger) from live ones, so a claim refused under §3.2 can be replaced by atomic parts without making the case unloadable. Live evidence, research, images, and the current edition on the current ledger remain under the strict no-rejected-claim rule; unknown IDs stay errors everywhere. The settlement cascade now relinks research and images the same way evidence already was. DECISIONS.md records a founder-supervised direction; AGENTS.md is untouched; no citations or confidence material are involved. This enables atomic-claim replacement and append-only history rather than weakening §3.15 checks.

The change resolves a real internal conflict in the checks rather than dodging them: assessment overlays are constitutionally append-only (§3.14, §4), so a historical assessment run or tombstone cannot be rewritten to drop a reference to a claim later rejected under §3.2 — under the old loader rule, no assessed or featured compound claim could ever be split, making §3.2 unenforceable precisely where it matters. The diff preserves every fail-closed property: an id no claim carries is still an error in every edition and every record, and live evidence, research items, images, and the current edition on the current ledger are still forbidden from referencing rejected claims; only frozen records (tombstones, stamped overlays, editions on a moved ledger) are excused, which matches their semantic role as records of what was live when written. The tests cover both the strict and frozen branches, including the case that an unknown id always fails regardless of frozenness, so this is not a blanket relaxation. No citations are added (verification confirms none), no confidence material is touched, AGENTS.md is not edited, and the DECISIONS.md entry documents the change with founder supervision, consistent with the operating-record rules.