Why the old citations cannot settle it

The End is less avoidable project work while preserving a defensible account. The earlier steering assistant invalidates a whole packet after an edit. This essai tests a narrower review and exposes a condition under which a citation-based shortcut fails.

A finite result

A complete two-option comparison recommends R. The truth of “R is least-cost admissible” depends on R, every other admissible option, the admission rule and the completeness of the option set. Keeping the original documents byte-identical preserves none of the latter three by itself. Adding a cheaper admissible B changes the answer. Adding an unauthorised cheaper X leaves it alone. A new signed possession note can make the available window conflicted without deleting the old note.

Our calculation represents truth as supported, defeated or unresolved. It compares the original claim with a fresh evaluation over the current typed packet. Unknown context never counts as retention. It can preserve a sentence while replacing the support calculation behind it. It also exposes two old overclaims that were unsupported before any edit. A change reviewer must know whether the starting review was sound.

The precise reuse condition

For a deterministic claim rule f over a complete typed input tuple d, equality of the complete tuples implies equality of f(d). More usefully, if d changes, recomputing f can justify the same sentence again. For a minimum-choice claim the tuple includes the competitor set and its scope, not just the winning item’s attributes. In this model the competing inputs can be enumerated. For arbitrary prose the corresponding dependency account has not been supplied, so the model returns unresolved.

This result does not require an AI to infer the dependencies. A careful person can make the same six-row table. The construction shows exactly what a computational assistant would need to maintain. A dependency proposed by a language model would itself need checking.

A fair work comparison

ApproachWork doneWhere it can fail
Manual change reviewRead the changes and the paper in context; judge each consequential sentence.A skilled reader may notice the new comparison or scope problem without maintaining a formal account. Their time and misses have not been observed here.
Whole packetWithdraw every earlier review after a packet edit.Safe triage can repeatedly send unchanged claims back to a reviewer. The reviewer must still identify pre-existing overclaims.
Changed old citationsCheck whether a cited record changed.New options, an unlinked contradiction and a changed use can defeat the conclusion while the old cited records persist.
Recalculate declared meaningInterpret and maintain typed evidence, its scope and version relation, then recalculate every mapped rule.Incomplete extraction or an unknown rule defeats the assurance. Keeping a sentence requires fresh support, even when no prose changes.

If preparing the typed account costs P units of effort, its upkeep costs M per review and it saves R units of human checking per review, it pays back only after N reviews with N(R−M) > P. If R ≤ M it has no positive repetition threshold. No values for P, M or R have been measured. Sentence counts are available as a limited burden proxy.

How the method changed during this attempt

The tempting first design marked a claim stale only when a cited source changed. The independently specified challenge included a new option and a new conflicting source. Both can leave all original citations untouched. The revised construction therefore evaluates the complete declared decision context and retains unmapped claims as unresolved. Its reader selects frozen versions, so the analytical result can be inspected without implying that free-text authorship, live source extraction or organisational review has been delivered.

Sources and verification

Buneman, Khanna and Tan distinguish the origin of a value from the input contributing to its existence. That distinction motivates asking what supports the conclusion beyond its quoted passage. Their result concerns database queries. Our steering sentences, option choices and manual extraction are a separate construction. Why and Where.

W3C PROV-DM distinguishes entities, activities, derivation and revision. It supplies a useful vocabulary for keeping the old paper and the new support distinct. A provenance relation alone does not establish that the sentence is warranted. The why-provenance paper also separates this question from view maintenance, which must account for additions; the new-option counterexample makes that boundary material here. PROV-DM.

Tests compare all four independently specified challenge changes and an ordinary direct calculation across 288 combinations of option price, budget, window, attendee authority, extra option, conflict and search completeness. The independent challenge is preserved as supplied, with its expected labels. A subsequent adversarial review revealed that a window in night shifts cannot be compared with uninterrupted-day estimates. This is now an explicit unresolved case. Four cases directly expose the citation-only rule’s false retentions. Ordinary browser verification checks the ordinary entrance, changed evidence, replay link, reload and reset. These establish a finite construction. Practical benefit remains unmeasured.

Calculation · Reproduced outputs · Independent challenge · Repository checks