The diff does not touch AGENTS.md, adds nothing under content/, cites nothing and carries no run records, so the question is whether it loosens a §3.15 check. It tightens them instead: it adds a mechanical reader's budget, puts markerErrors into the loader, and widens extractPlateRefs so plates the old regex missed are now validated. It also puts the check and comparison seats under the spend caps while leaving the constitutional panel uncapped, and gates pure re-tellings on a comparison whose order is balanced and blind. The chain.test `loaderErrors` filter strips only the new budget findings, not any older loader rule. The comparison is labelled 'a preference among AI readers, not a human review', verdicts stay with the blind check, and every edition still faces the panel in its PR, so nothing presents an AI or unratified record as human or ratified; the verb's measured judgment-change line strengthens §3.14. Protocol v14's plain-words rule keeps the report, unread and provisional distinctions ('plain wording changes the vocabulary, never what is claimed to be known'). A global length ceiling is not the per-case pin §7 retired, and the comparison builds the competition of drafts §7 names. The DECISIONS entry's word counts, costs, tallies, run ids and over-budget case count agree across the PR body, the log and v14, and the founder direction carries its date and channel; I see no positive evidence of invented provenance (readerBudget.ts's 2,177→7,514 versus 2,178→7,502 reads as two word counters). Fixes I would like, as ordinary imperfections rather than breaches: comparisonInWords' 'undecided' branch says an edition replaced its predecessor 'because the evidence had changed' even when it went out to answer panel dissent or an objection on an unmoved ledger, and should add 'or the panel had to be answered'. The method page's 'without being told which is which' should also admit that the WRITTEN FROM marker identifies the candidate whenever the incumbent is stale, and should mention the answers-the-panel branch.

`judgmentChanges` detects component changes only by comparing the sorted multiset of component states, while discarding the association between each component and its state. Thus swapping two components from A=contradicted/B=supported to A=supported/B=contradicted would be reported as no component-verdict change, and `judgmentChangesSentence` would append that false assertion to the edition rationale. That makes the visible assessment history incomplete or incorrect, contrary to §3.14 and §3.15's requirement that readers can reconstruct what changed and how it was checked. The provenance narrative is also internally inconsistent: `readerBudget.ts` says the same article grew from 2,177 to 7,514 words, while the added DECISIONS entry says 2,178 to 7,502; these figures should be reconciled and the component comparison must preserve component identity before the measured history can be trusted.

This change introduces the reader budget floor and candidate comparison step in the editorial pipeline without weakening the constitutional checks or the blind panel. AGENTS.md is untouched, and decisions log documentation records the founder direction from 2026-09-29. The UI and schemas clearly attribute the comparison as a preference among AI readers rather than human review or epistemic ratification. No content under content/ is added, no citations are fabricated, and all automated checks and domain tests pass.

The diff implements a reader budget and blind multi-seat comparison for edition candidates, with no AGENTS.md edit and no content/ ledger, citation, or case-record fabrication. Comparison outcomes, reasons, and judgment deltas are stamped and displayed as AI preferences with honest labels, preserving §3.14–§3.15 provenance and multi-model checks rather than weakening them. Budget enforcement, marker refusal, and append-only decisions logging strengthen mechanical validity without bypassing concurrence or presenting unratified drafts as ratified. Omitted-files and citation-verification sections are empty; RUN ACCOUNT carries no runs, matching the PR’s claim that editions follow separately.

The diff touches no AGENTS.md and adds no citations or case content, so neither the reserved-amendment rule nor the fabrication rules are in play; the citation verifier found no added identifiers. The change adds checks rather than weakening any: the reader's budget is a new fail-closed mechanical gate, the comparison adds a preference test before a re-telling can replace an incumbent (and a moved-ledger candidate still goes out, with the preference recorded, rather than being suppressed), and the blind check and constitutional panel are left intact and explicitly named as what would catch a 'shorter, plainer and wrong' candidate. AI involvement is displayed honestly throughout — the method page and ComparisonPanel label the preference as 'a preference among AI readers, not a human review', and the code comments and tests enforce that the wording never claims 'better', 'human', or 'reviewed'. Provenance is kept: run records, spend rows for the two discarded runs (with the unrecorded cost of lost calls admitted rather than hidden), rationale appended with verb-measured judgment changes so a changed assessment records what changed (§3.14), and a test holds the protocol's stated figures to the code so a promptVersion always points at the text that ran. The union-merge attribute for the append-only decisions log preserves both appended entries and does not weaken provenance. Nothing here presents an unratified draft as ratified: the two standing test editions are explicitly held unpublished pending this gate.