Verified Repair — history

The dated record of how Verified Repair got here, moved out of the spec on 2026-10-21 because a reader does not need it to use the spec today. The spec keeps every section spec-format.md requires and its Deleted: tombstones; everything smaller is in git log -- compositions/verified-repair.md.

Decisions

Entries that record a round, a maintainer’s ruling or the formal model’s drafting rather than a turn the pattern took.

  • 2026-10-10 — Round 1: a repair builds its own tree, a round counts by pass, and the maintainer rules through an action. Chose: [Repair] builds its tree from the head and the repository takes it only on that head; a verdict names its pass and a round counts one verdict per pass; the settled triage and the round’s own commit decide whether a round was clean; the maintainer’s ruling is a record [Rule] writes, read where a round starts and where a claim is patched; the council nonce is the pairing datum, and the intent event id rides the trailer beside it. Over: a shared working copy restored on failure, three verdicts of any kind closing a round, and a ruling no action could record. Because: the first round’s three passes found the shared copy, the count and the ruling broken from two or three directions each, and every cure that deleted state beat a cure that guarded it.

  • 2026-10-10 — Round 2: the earliest record counts, and a ruling carries what it admits. Chose: the earliest sound verdict of each reader pass, and the earliest triage record of each role, as the ones that count, with one actor to one record of a finding; a maintainer’s ruling that carries the round and the patch whole, compared byte for byte; the round budget counted over frozen rounds, a page’s standing ruling spent by the freeze it admits; a commit excluded from every check once any record a check reads for it is purged; [Repair]’s staleness and settlement answered before any gate escalates. Over: a lock around the journal’s read-then-append, a ruling that names a patch id, and a budget over settled rounds. Because: a read-then-append the substrate cannot make conditional is settled by an order the journal already keeps, and a ruling on a name is a ruling on whatever later carries the name.

  • 2026-10-10 — Round 3: the journal’s order decides the budget, and a settled round is closed. Chose: a round’s position in the journal’s order, not [Freeze]’s read of a count, as what puts a round past the budget, its verdicts refused; a settled round closed to later reports and triages; every count scoped to the findings counted verdicts carry, a rejected finding counting for nothing; a finding named by its verdict and ordinal; the round’s records narrowed to the shapes they take; [Repair]’s caller held to the writer. Over: a lock around [Freeze], and a round whose settledness later records could undo. Because: the same race the second round closed for verdicts reappeared at [Freeze], and the same order closes it; and a fact a commit rested on must not change after the commit.

  • 2026-10-10 — Round 4: a settled round stays settled, and the status is a claim. Chose: a round settles at the first record in the journal’s order that settles it, with later triage records counting for nothing and [Repair] waiting out the protocol’s settle bound; the claim families widened to every Term their rules name and to the status line, so a page’s grounding is ruled by the maintainer; the target and the records sites declared, and the last gate line checked against the round’s count; the store’s codes returned bare, an undeclared refusal left to the protocol’s unknown arm. Over: a lock around [Triage], a claim gate keyed to three label families, and a wrapped store code. Because: every race this page met had the same shape and the same cure, and the cheapest way to weaken a claim was to change the words it is written in.

  • 2026-10-10 — Round 5: a council commit is any commit with the trailer, and the page’s own records land like the protocol’s. Chose: the checks’ population named by the trailer alone, so a forged trailer is found rather than defined away; the repository’s pre-commit codes closed to three; the page’s own writes ordered after its input checks, a step-4 record answered as written; the council’s journal reads held to the visibility the protocol guarantees its own; a clean round read across every settled round of its snapshot; the round budget written into the round it governs. Over: a population defined by its own validity, an open code set, and a budget read from today’s disclosure. Because: a check that cannot fail is not a check, and every number a later check reads has to be the one the action read.

  • 2026-10-10 — Round 6: a clean round is judged when its repair opens, and what a check reads is fixed when the action reads it. Chose: a round’s cleanness judged against its blocking siblings at its repair’s intent, never revised by a sibling that settles later; a freeze refused on a page the head has moved; a report’s position read only once the round’s record is past the settle bound; the settling record aged as the journal read yields it; every council record under the act kind’s retention policy; the claim families read over the snapshot and the repair tree, with the Intent and the Ledger’s status lines among them. Over: a clean status any later record could revise, and a claim gate read over the snapshot alone. Because: a fact a commit or a count rests on has to stop moving at the moment it is used, and a claim can be weakened by what a patch adds as easily as by what it changes.

  • 2026-10-10 — Round 7: the repository guards the trailer, and a verdict carries the position it was counted at. Chose: a repository that refuses a council trailer from anyone but the writer or with a nonce already used, and never moves its head off its history; a verdict recorded with the round position it was counted at, so a later purge cannot recount it; a claim gate over sites, not rules; a reader handed the permitted ranges alone; a site that runs to the next label. Over: uniqueness asserted by a Term, and a position recomputed from whatever the journal still holds. Because: a property a check relies on has to be one the store enforces or one a record froze.

  • 2026-10-10 — Round 8: an intent names its round by the council’s own key, and a third triage runs only where two parted. Chose: a round key minted at [Freeze]’s seam in place of the event id and the base commit the substrate and the repository mint, with the probe reading by trailer alone; a council nonce fresh for every call; the permitted ranges read at the base commit; a third triage admitted only on a parted finding, its brief blind as the second’s is. Over: stating why the protocol’s ban on minted fields did not reach this intent, and a third triage any triager could add. Because: the cheapest way to honour a constituent’s rule is to stop needing what it forbids, and a tie-breaker that can overrule an agreement is a third vote, not a tie-breaker.

  • 2026-10-10 — Round 9: the triager who disposes a cure carries the patch, and a commit’s intent follows its ancestor’s. Chose: the settled triage carrying the cure’s patch whole, [Repair] refusing any other; every triage rule reading the same counted triages; a council commit whose trailer names an intent older than its ancestor’s caught by sequence numbers a purge keeps. Over: a patch the writer assembles beside the triage, and an exclusion that let a forged trailer naming a purged intent pass every check. Because: a fragment no triager saw is a stowaway whatever the diff check says, and the one fact a purge leaves standing is the order the journal wrote things in.

  • 2026-10-10 — Round 10: a ruling names the round it was given against, and a purged record bears on what its neighbours bracket. Chose: a page-alone ruling recorded with the page’s latest round and standing only while that round is latest; a purged record’s reach settled by the sequence numbers a purge keeps; a purged round answered as purged; a triage written only once admitted. Over: a ruling read as “later than the latest round”, which a late copy satisfies twice, and an exclusion that needed content the purge had destroyed. Because: the fresh share fell to two in eight, and what remained were the draft’s own assumptions about time and identity.

  • 2026-10-11 — Round 11: a check that meets a purge says unverifiable, and closing a debt is the maintainer’s. Chose: a check of this page that reads a purged record answering unverifiable, carrying purged-horizon, as Chain of Custody’s checks already do; Term bears on deleted; every council action’s own record written only after all its refusals; removing a Ledger open line gated to a ruling. Over: an exclusion that had to know, after the purge, what the purged record had named. Because: the count fell to four, and the cheapest cure for the last purge finding was to stop guessing what a destroyed record said.

  • 2026-10-17 — Round 12: triage reads only what the round counts. Chose: [Triage] admitting a finding only of a counted verdict or a formal engine’s sound verdict, so a verdict that lost the race to its pass is sorted by no one and answered by no patch. Over: a dedupe of raced verdicts by finding site, and a countedness test at [Repair] alone. Because: the first round the automated council ran counted one pass of three, two free readers having read past their ranges, and the pass that counted found a raced verdict reaching the commit path on one triage, past the double sort a deciding count is owed.

  • 2026-10-17 — The maintainer’s ruling on round 12: a lapsed reader’s findings are cured like any other. Chose: Invariant 6.1 and Check 6.1 narrowed to a formal engine’s malformed verdict, the case their WHY covers; a reader’s malformed verdict counting for nothing while the findings it carries are counted findings, admitted to triage, sorted twice where a count decides, and cured; and the lapsed passes’ four foundational findings cured in place — a step-4 refusal’s event id answered by [Freeze] alone, as its round id; one counted verdict per verifier as well as per pass; the yield’s reply keyed on the unknown arm’s two cases; the repository read Invariant 7.3 walks declared (Capability requirement 26). Over: routing a lapsed pass’s findings to the next round, as round 3 and this round’s first commit did, and admitting them to triage outside the count, the side door this round’s Linus finding had just closed for a raced verdict. Because: the maintainer ruled that the entry A lapsed pass does not count in pressure-testing.md already says what a lapse costs, the pass’s count and not its findings; the same ruling closed the four open lines the cures answer and narrowed 2026-10-10-n, whose first clause the second cure carries (Action wiring 63).

  • 2026-10-17 — Round 13: refusals that co-fire are ordered. Chose: the three pairs of refusal rules the round’s readers found co-firing with nothing ordering them classed foundational under the method’s test, since two conforming builds answer different codes on one call, and cured with the finders’ orders — [Triage] answers Action wiring 36, and then 35, only after its other refusals (Action wiring 64, 65); [Rule] answers Action wiring 38 after its primitive policies (Action wiring 66); [Freeze] answers Action wiring 57 after Primitive policy 2 (Action wiring 67) — and the same shape elsewhere on the page routed as 2026-10-17-p. Over: keeping the shape refining, as 2026-10-10-n has since round 1, and one precedence for the whole page, a direction left to the maintainer. Because: two of the round’s readers, each handed only the reading set, found the shape apart on three actions, and the method’s test counts a divergence the page’s silence produces even where the refusal writes nothing.

  • 2026-10-17 — Round 14: the envelope is the kind’s largest record, a purge hides payloads and not order, and a purged round leaves the indexes. Chose: the commit envelope as the true maximum of the records the kind writes, the sweep’s and the maintainer’s closings beside the committed commit id, with a council.commit.escalated’s store candidates held to one commit by Capability requirement 19; Invariant 7.2 and Check 7.2 over a purged record’s payload, so the sequence numbers Check 7.3 reads, which a purge keeps, still catch a forged trailer; the round, triage and ruling indexes rebuilt from what the journal still holds, a purged element dropping out, with [Repair] answering purged before unsettled (Action wiring 68); and a verdict recorded with its round id. Over: an envelope of one commit id, which Recoverable Invocation External check 7 fails, a blanket unverifiable that silenced the page’s own Check 7.3, and a past-horizon reading left to each build. Because: the round’s three readers, on the reading set alone, found the page sizing its act below what its constituent requires, contradicting its own check across a purge, and leaving a purged round’s answer to whether an index kept it; round 13 had sent the last down to refining, missing the round-closed answer a kept round gives.

  • 2026-10-17 — Round 15: the settling record is read from the journal, and the payload cap covers the intent. Chose: Term settled round reading the round’s journal records, as Action wiring 54 names them, where it had borrowed the name of the page records a round owes; Capability requirement 23’s caps set above the council.commit intent too, every patch whole included. Over: splitting a round’s cures across repairs, which 2026-10-17-aa proposed and Action wiring 15 refuses once the first lands, and leaving the intent to the invocation open’s sizing alone. Because: a round’s cures land in one commit or not at all, and the last gate line every repair carries grows each round, so a cap that leaves out the intent closes the page to repair as it ages; and a Term’s name used in another sense is read as the Term.

  • 2026-10-17 — Round 16: a commit’s intent is compared with every council-commit ancestor’s, and two more refusal pairs are ordered. Chose: Invariant 7.3 and Check 7.3 comparing a council commit’s intent with the intent of any council-commit ancestor, where they had named the nearest one, which a person’s merge leaves undefined; Action wiring 61 deferred with the rest of [Repair]’s checks until settlement is answered (Action wiring 46), and Action wiring 6’s escalation until [Report]’s primitive policies have answered (Action wiring 51). Over: the shortest parent path the Linus finder proposed, under which a merge back to an older commit lets a forged trailer naming a purged intent pass, and an existence test on the nearest ancestor, which leaves a merge’s nearest unread. Because: Check 7.3 is the one check a purge cannot blind, and a check two auditors read two ways on one history catches a forgery for one of them only; a comparison with every ancestor needs no Term for nearest and is met by the first council commit, which has none.

  • 2026-10-17 — The formal model drafted: a malformed verdict written after the settling record carries no counted finding. Chose: three TLA+ models derived from the page, of the commit path, the round’s settling and the round budget, with eleven twins and nine probes; Term counted finding reading a reader’s malformed verdict only where the verdict is recorded no later than the round’s settling record, with the WHY of Verdict and of Invariant 6 saying so; and Invariant 2 resting on Capability requirement 13 and not on Recoverable Invocation’s one writer per act key. Over: one model of the whole loop, whose state space the harness cannot hold, and every malformed verdict’s findings kept counted, which leaves Check 5.1 failing on a council that kept every rule. Because: the settling model found that [Report] refuses already-counted and round-closed only from what its read shows, so a malformed verdict can be written after the record at which the round settles; its finding was then a counted finding no counted triage could reach, a triage recorded after the settling record counting for nothing, so a zero count that decided the round rested on no triage of it. The verdict still reaches the maintainer through its escalation (Action wiring 6). The commit-path model holds Invariant 2 with no critical section at all, two council commits being in flight at once with no commit fence.

  • 2026-10-18 — Round 17: a formal engine’s findings are triaged before the round settles, [Triage]’s round-closed is ordered, and a ruling needs a round. Chose: Action wiring 69 and Term settled round holding the round unsettled until every finding a formal engine’s sound verdict of the round carries has a counted triage, the Verdict WHY saying where that triage is owed; Action wiring 46 deferring 24, 25 and 62 with the rest of [Repair]’s checks; [Triage]’s round-closed evaluated after Primitive policy and before [Triage]’s other refusals (Action wiring 70, 71); Primitive policy 22 refusing a ruling that names a page alone where the journal holds no round of the page. Over: the finder’s order for [Triage], round-closed before Primitive policy, under which a journal read that refuses leaves Primitive policy 11 unreachable and the call with no answer; the Verdict WHY narrowed to fit the rules as they stood, which leaves an engine’s counterexample with no owner once the round settles; and [Report] recording a malformed verdict on a settled round, which GRID proposed for a gap the text does not have, since [Report] escalates every malformed verdict it records. Because: an engine’s counterexample is the formal layer speaking about the page, and a round that settles past it leaves the counterexample to nobody; an order between refusals is the answer a caller reads; and a rule that records a field which does not exist has no conforming build. An engine verdict recorded after the settling record, in the race [Report]’s read leaves open, stays without an owner and is routed (2026-10-18-b).

  • 2026-10-19 — The maintainer’s ruling on 2026-10-17-p: each action evaluates its refusals in one stated order. Chose: a refusal order for each of the five actions, every refusal rule the action answers ahead of its own write listed earliest first (Term freeze, report, triage, ruling and repair refusal order), and one rule per action holding each rule back until every earlier one has answered no refusal (Action wiring 72 through 76); the pairwise orders the rounds had cured one at a time (Action wiring 46, 50, 59, 64 through 68, 70 and 71) folded into the five with every one of them kept; an action’s input ahead of its reads, a read’s failure ahead of the rules that need it, purged ahead of absent, [Report]’s too-young ahead of its budget-spent; 2026-10-17-ad and 2026-10-10-n’s first clause closed with it. Over: one precedence for the whole page, and curing pairs as readers found them, which five rounds did without reaching the rest. Because: the maintainer ruled it after round 17: a caller reads the code, and an order stated for each action leaves no two refusals answering one call; every order an earlier round cured is kept, so no answer it fixed changes, and a refusal writes nothing, so no state any of the three models checks changes.

  • 2026-10-19 — Round 18: a last gate states its count where [Repair] reads it, and the council’s records come only through its actions. Chose: Term stated count, the numeral after the words foundational count in the last gate line’s leading entry, a clean round’s stating foundational count 0, with Action wiring 53 answering mismatch where the count is blank or differs from the round’s; Capability requirement 27, the deployment refusing a record_action under the composition’s own action references from any caller but its actions. Over: a count field added to the Ledger’s grammar, which spec-format owns and the linter parses; and the reservation the eos reader proposed, Audit Trail’s widened to council. or the references kept to the writer’s actor reference, which changes a constituent’s contract outside a round on its page (Composes 7) and would refuse the records [Triage], [Report] and [Rule] write under their callers’ own references (Primitive policy 12). Because: the eos reader reported both as refining, and the blind second triage, placing every finding the count could include, raised both: one build reads a clean entry as zero where another finds no count and refuses, and every index this page reads is rebuilt from records that any holder of a journal credential could append beside the actions, so a second triage with no first, or a verdict about other bytes, would settle or count a round while every rule here held; a page whose counts decide outcomes has to say where a count is written and who may write what it is read from.

Ledger

The gate records the Ledger’s last gate: line carried behind its latest until 2026-10-21, newest first:

  • 2026-10-18 — Final Critique 17, the first round with the formal model present, three fresh readers on free models (GRID DeepSeek V4.1 Flash, EOS LongCat 2.5 Preview, Linus MiMo V2.6 Flash), each handed only the round’s reading set, three passes on one text (round 16’s cures and the formal model’s, frozen at a0ab44a), no lapse — foundational count 4 after triage (5 reports and 1 refining report raised, 1 rejected), two in fresh text: Action wiring 46 leaving 62, 24 and 25 unordered against staleness and settlement, and Term settled round letting a round settle with a formal engine’s finding untriaged, which the Verdict WHY says is triaged; and two in older text: [Triage]’s round-closed ordered against none of [Triage]’s other refusals, and a ruling naming a page alone admitted on a page with no round, which Action wiring 39 then records with a latest round that does not exist; 16 refining reports, 7 rhetorical reports; not clean; every foundational finding cured in place, none in a rule the models check; the council’s sixth round on the page without grounding, so the page goes to the maintainer
  • 2026-10-17 — Final Critique 16, three fresh readers on free models (GRID and EOS DeepSeek V4.1 Flash, Linus GLM-5.3 Flash), each handed only the round’s reading set, three passes on one text (round 15’s cures, frozen at 9d00a79), the Linus reader reading three process files outside the set after its context was compacted, none of them corpus text, so its verdict is malformed and counts for nothing and its findings stand — foundational count 3 after triage (13 reports, 11 after merging duplicates and one carrying two pairs, 9 sent down to refining), all in older text: Invariant 7.3 and Check 7.3 comparing a commit’s intent with its nearest council-commit ancestor’s, which a person’s merge leaves undefined and one reading of which lets a forged trailer naming a purged intent pass, and two refusal pairs nothing ordered, Action wiring 31 against 61 on [Repair] and Primitive policy against Action wiring 6’s escalation on [Report]; 19 refining reports, 6 rhetorical reports; not clean; every foundational finding cured in place; the council’s fifth round on the page without grounding, so the page goes to the maintainer
  • 2026-10-17 — Final Critique 15, three fresh readers on free models (GRID and EOS DeepSeek V4.1 Flash, Linus GLM-5.3 Flash), each handed only the round’s reading set, three passes on one text (round 14’s cures, frozen at 130a2bc), no read outside the set — the Linus reader lost the spec under review after its context was compacted and reviewed Audit Trail, so its verdict is malformed and counts for nothing — foundational count 2 after triage (7 reports against this page, 5 sent down to refining; 2 more against Audit Trail, escalated), both in older text: Term settled round reading the journal through a Term that names page records, and Capability requirement 23’s caps leaving out the intent a round’s one repair carries; 16 refining reports, 3 rhetorical reports; not clean; every foundational finding cured in place
  • 2026-10-17 — Final Critique 14, three fresh readers on free models (GRID and EOS DeepSeek V4.1 Flash, Linus GLM-5.3 Flash), each handed only the round’s reading set, three passes on one text (round 13’s cures, frozen at dda2c77), no lapse — foundational count 3 after triage (10 reports, 6 sent down to refining and 1 rejected), all in older text but the Invariant 7 WHY, which round 12 rewrote: the commit envelope sized below the true maximum Recoverable Invocation requires, Invariant 7.2 answering unverifiable where Check 7.3 reads through every purge, and the three indexes with no past-horizon classification, which round 13 had sent down to refining; 12 refining reports, 6 rhetorical reports; not clean; every foundational finding cured in place
  • 2026-10-17 — Final Critique 13, three fresh readers on free models (GRID and EOS DeepSeek V4.1 Flash, Linus GLM-5.3 Flash), each handed only the round’s reading set, three passes on one text (round 12’s cures, frozen at f40662b), no lapse — foundational count 3 after triage (10 reports, 5 sent down to refining and 2 rejected), all in older text but Action wiring 35, which round 12 rewrote: three pairs of refusal rules that co-fire with nothing ordering them, on [Triage], [Rule] and [Freeze], the shape 2026-10-10-n had carried as refining, raised under the method’s test; 11 refining reports, 3 rhetorical reports; not clean; every foundational finding cured in place
  • 2026-10-17 — Final Critique 12, the automated council’s first round, three fresh readers on free models (GRID and EOS DeepSeek V4.1 Flash, Linus GLM-5.3 Flash), three passes on one text (round 11’s cures, frozen at 5b7ea86) — the structural reader lapsed, a search printing Ledger and Decisions lines, and the EOS reader lapsed unreported, reading method sections past its ranges, so under Verdict 5 both verdicts are malformed and count for nothing, and their findings stand — foundational count 5 after triage (10 reports, 4 sent down to refining and 1 rejected), 1 from the sound Linus pass and 4 from the lapsed passes, all in older text but Check 7.3, which round 10 rewrote; 19 refining reports, 3 rhetorical reports; not clean; every foundational finding cured in place, the lapsed passes’ four in a second commit on the maintainer’s ruling that a lapsed reader’s findings are cured like any other, Invariant 6.1 narrowed to a formal engine’s malformed verdict
  • 2026-10-11 — Final Critique 11, three fresh readers, three passes on one text (round 10’s cures, frozen at 844317f), no lapse — 4 foundational after deduplication (5 reports), two in text rounds 9 and 10 wrote and two older; 51 refining reports, 13 rhetorical reports; not clean; every foundational finding cured in place
  • 2026-10-10 — Final Critique 10, three fresh readers, three passes on one text (round 9’s cures, frozen at 4a1cdcd), no lapse — 8 foundational after deduplication (9 reports), two in text rounds 8 and 9 wrote and six older; 52 refining reports, 10 rhetorical reports; not clean; every foundational finding cured in place
  • 2026-10-10 — Final Critique 9, three fresh readers, three passes on one text (round 8’s cures, frozen at ac06b93), no lapse — 9 foundational after deduplication (10 reports), four in text rounds 7 and 8 wrote, five older; 49 refining reports, 10 rhetorical reports; not clean; every foundational finding cured in place, each report given its own triage row first
  • 2026-10-10 — Final Critique 8, three fresh readers, three passes on one text (round 7’s cures, frozen at e9edaa7) — the Linus reader lapsed in form only, cutting one search’s output, with no read outside its ranges, so its verdict counts — 9 foundational after deduplication (11 reports), four in text rounds 6 and 7 wrote and five older, one of the older a round-5 report the round-5 triage dropped (a third triage overturning two that agreed); 54 refining reports, 13 rhetorical reports; not clean; every foundational finding cured in place
  • 2026-10-10 — Final Critique 7, three fresh readers, three passes on one text (round 6’s cures, frozen at baede42), no lapse — 9 foundational after deduplication (11 reports), six in text rounds 5 and 6 wrote, three older; 50 refining reports, 16 rhetorical reports; not clean; every foundational finding cured in place
  • 2026-10-10 — Final Critique 6, three fresh readers, three passes on one text (round 5’s cures, frozen at c63ce69) — the structural reader lapsed in form only, cutting long lines and joining two reads, with no read outside its ranges, so Verdict 5 does not reach it and its verdict counts — 9 foundational after deduplication (13 reports), seven in text rounds 4 and 5 wrote, two older; 53 refining reports, 13 rhetorical reports; not clean; every foundational finding cured in place
  • 2026-10-10 — Final Critique 5, three fresh readers, three passes on one text (round 4’s cures, frozen at 1aa5a6e), no lapse — 10 foundational after deduplication (14 reports), six in text rounds 3 and 4 wrote, four older; 49 refining reports, 10 rhetorical reports; not clean; every foundational finding cured in place
  • 2026-10-10 — Final Critique 4, three fresh readers, three passes on one text (round 3’s cures, frozen at 8795298), no lapse — 8 foundational after deduplication (15 reports), six in text rounds 2 and 3 wrote, two older; 51 refining reports, 10 rhetorical reports; not clean; every foundational finding cured in place
  • 2026-10-10 — Final Critique 3, three fresh readers, three passes on one text (round 2’s cures, frozen at fa017b5) — 12 foundational after deduplication from the two sound passes (16 reports), six of them in text round 2’s cures wrote and six older; the Linus reader lapsed, reading the Status heading and its one line past its range and no Ledger line, so under Verdict 5 its verdict is malformed and counts for nothing: the four of its findings the sound passes also reported are cured on theirs, and its one other is routed, not cured (Invariant 6.1); 46 refining reports, 12 rhetorical reports; not clean; every foundational finding of a sound pass cured in place
  • 2026-10-10 — Final Critique 2, three fresh readers, three passes on one text (round 1’s cures, frozen at 660ecec) — 13 foundational after deduplication (25 reports), nine reached by two or three passes, all fresh, ten of them in text round 1’s cures wrote; 35 refining reports, 9 rhetorical reports; not clean, so no count decided anything; every foundational finding cured in place
  • 2026-10-10 — Final Critique 1, three fresh readers, three passes on one text (the first draft, frozen at 47128cc) — 12 foundational after deduplication, every one reported by two or three passes, all fresh; 32 refining reports, 9 rhetorical reports; not clean, so no count decided anything and no second triage was owed (Action wiring 11); every foundational finding cured in place
↑