When a many-paper formal framework must change its mind — a computational defect found in a deposited paper, a conditional whose antecedent a later audit moves, a pricing sharpened by a later result — who edits? The implicit answer this paper personifies as the Global Editor: the authority with retroactive write access to the corpus's history, for whom consistency is maintained by silently updating the record until it agrees with itself. Its practice-side shadow is familiar — the updated preprint with no change log, the silently corrected equation, the claim whose history is unrecoverable across versions. This series has built against it piecemeal: deposited texts are DOI-immutable, registers are frozen, upgrades are annotation-only, in-place new versions carry itemized notes re-verified at gates. But piecemeal discipline is precedent, not law, and the place precedent predictably breaks is under load: Paper 64's audition left a live adoption-triggered branch whose execution would touch frozen conditionals across a dozen papers at once. Without a cascade calculus the honest options degenerate to "rewrite everything" (the Editor) or "annotate nothing" (unannotated staleness accumulating in the record). This paper — the closing panel of the nineteenth grouping, consuming Paper 63's Audition Charter unmodified and branch-complete by construction: nothing here consumes 64's realized landing — replaces that precedent with an explicit, checkable discipline. The registered controlling phrase: The corpus may revise its standings, but it never rewrites its past. The revision typology fixes four acts, none of which changes content-at-address: annotate (add-only marginal content on a frozen register — 48's extension rule, refined) ; re-version (two declared forms — a full version bump with an itemized note and re-run gates, and a status pass; at deposit status S3, no act can replace the deposited artifact: a re-version act at S3 creates a new deposited successor artifact and a forward-supersession relation; it never mutates the prior deposited object) ; retract (a leaf's downstream-consumption status withdrawn; the text untouched; the retraction notice itself a record) ; and retain-as-conditional (the falsified-antecedent case, in the canonical format: L: A=>C becomes L': Retained as A=>C with AntecedentStatus (A) recorded — the proposition body unchanged, only the consumption right changes). The act-selection function sigmaₐct (leaf type, defect type, deposit status) is total and decidable, with the D1/D3 boundary registered — D1 concerns error at the time of assertion; D3 concerns later change in the status of an antecedent while the original derivation remains correct under its stated condition. The per-leaf rules and the revision authority layer Aʳev (roles x leaf types, every cell a right or an explicit bar) close with the paper's flagship display: for every role r and leaf l, (r, rewrite, l) is not in Aʳev — no role has rewrite authority over any frozen leaf. The annotation-cascade calculus makes propagation mechanical: nodes are (paper, leaf) ; edges are declared consumptions read from each paper's "what is consumed" section and index-card hooks, typed by mode (record / record-input / template / display) ; each mode fixes the consumer's obligation; traversal is breadth-first over reverse edges with a visited set per revision event, so the finite declared graph terminates unconditionally; a cycle containing a record or record-input edge flags a registration defect, while display-only cycles are reported and carry no cascade obligation; an undeclared consumption is itself a typed defect whose route degrades incompleteness to late notification, never silent inconsistency. The non-retro invariant — no act changes content-at-address — is proven preserved act-by-act. The discipline is then checked, not validated: applied to the corpus's own executed revision history — eleven in-place re-versions, five DOI-insertion passes, the add-only annotation formulas, the deviation-note routings, one gate-time repo-standard correction — sigmaₐct returns, for every instance, the act that was actually performed; the two acts history never exercised (retract; retain-as-conditional) are supported instead by independent constraints and displayed negative cases. The application table ships both templates schematically — the adoption-triggered-branch rows with acts pre-assigned per 63's census formulas, the declined-branch rows over 47's successor-protocol seat — instantiable by whichever landing 64's branch later realizes, which is what branch-completeness means. The ceiling: no physics anywhere; no audit re-run; 64 a status line; the Replication row — the community-level mechanism by which independent readers reproduce, contest, or extend results — explicitly not taken: this paper only specifies how the corpus records a defect or revision once a typed defect is declared. AuditionStatus (revision) = scoped at typing grade; (j5) open by seat. The grouping closes at the panel level with mechanism, audition, and law all delivered; the belief-revision, scholarly-record, and versioning literatures (AGM; Hansson; COPE; DataCite; Nosek) are engaged as comparison seats — what none ships is an executable per-leaf mechanism with a cascade calculus and a non-retro invariant, and that mechanism is the contribution.
Tomoyuki Uchida (Mon,) studied this question.