The diff touches nothing under content/, adds no citations, and leaves AGENTS.md alone. The core change memoizes loadAllCases and loadArbiterRecords only when ALETHEIA_CONTENT_FROZEN=1, which only the package.json build script sets, so pipeline verbs and tests still read fresh (frozen.test.ts pins both the default and the flag's single use). No §3.15 check is weakened. readOnce still runs the full validating load once per process, and a load that throws is never cached, because deepFreeze(load()) throws before held.set, so malformed records and dangling IDs still fail the build closed. Deep-freezing adds a failure mode (a page that mutates a record throws) rather than removing one. The StandingPanel change shows the claim-by-claim tally that was previously withheld, labelled with how many seats' verdicts were given on an earlier version; that is more disclosure, and the standing computation is untouched. The DECISIONS.md entry records what changed and why, with the house founder-direction stamp (date and channel), and nothing visible shows that stamp or the #421 seat attributions misstating who did what. Two non-blocking nits. First, the new method-page sentence says unresolved and presently untestable 'stand on the same step', while the constitution says they are 'one step from either and from each other'; adding 'one step from each other' would mirror it exactly and avoid implying an unresolved-versus-untestable pair counts as exact. Second, the guard test scans only top-level files in .github/workflows, scripts and src/pipeline, so 'every verb' is overstated if verbs sit in subdirectories.

The build-only cache does not weaken fail-closed validation: cases and arbiter records are still parsed and schema-validated on their first read, and the cached results are deep-frozen against cross-page mutation. The flag defaults off, and the added test restricts its repository use to the static build, preventing pipeline verbs from judging stale ledger state. The standing panel now expressly labels how many verdicts concern an earlier case version, while the method text retains the distinction between unresolved and presently untestable. The change edits neither AGENTS.md nor content records, adds no citations or confidential material, and records the implementation in the decision history.

The change introduces an in-memory cache and deep freeze for build-time content loading (`ALETHEIA_CONTENT_FROZEN=1`) without weakening pipeline verification or test runs, which remain fresh reads by default. The UI and documentation updates in `StandingPanel.tsx` and `app/method/page.tsx` accurately reflect the constitutional rules on panel alignment and the graded scale placement of 'unresolved' and 'presently untestable'. No citations or content are fabricated, `AGENTS.md` is untouched, and decision provenance is properly recorded in `docs/DECISIONS.md`.

The diff is a build-only content cache (ALETHEIA_CONTENT_FROZEN gated to package.json build, deep-freeze, tests forbidding the flag elsewhere), plus two UI notes that better match §3.15 (same-step unresolved/untestable wording; claim tally shown with a stale-version count rather than hidden). No content/, citations, assessments, AGENTS.md, confidence material, or check-weakening appear; DECISIONS/MAINTENANCE only document the change.

This is a build-performance change plus two display corrections, with no content records, citations, or provenance-bearing material added, and the citation verifier found no identifiers in the added lines. It does not touch AGENTS.md and does not weaken any §3.15 check: validation still runs fail-closed on every load (the Zod parse is inside the readOnce closure), the freeze defaults off so pipeline verbs keep reading fresh content (avoiding a stale ledger judged as current), and a new test pins that only the build script sets the flag. The two #421 fixes move toward transparency rather than away from it: the method page now correctly states the constitution's §3.15 scale (unresolved and presently untestable on one step), and the claim-by-claim tally is now shown with an explicit label disclosing how many seat verdicts were given on an earlier version of the case, instead of being silently withheld — disagreement displayed rather than hidden. The DECISIONS.md entry records the decision with date and channel as the constitution's operating record contemplates, matching the file's established pattern; nothing in the visible diff fabricates a source, locator, or attribution. The output-equivalence claim (byte-identical builds) is asserted but plausible given the caching is read-only and deep-frozen, and even if the performance numbers were optimistic that would be ordinary imperfection, not an epistemic violation.