{
 "round": "two-frontiers-cove-r1",
 "gate": "hostile-read",
 "utc": "20260906T181649Z",
 "verdict": "READ — the page may NOT release as read: seventeen must-fixes, fourteen of them defects rather than sentences, and the puff-piece thesis is NOT proven — the page cheats with absences, not with numbers",
 "detail": "it read draft v1 as the contestants' family's own advertisement and could not prove it — the deck is a rival's line, the author's family holds no chair, the one comparative that would put its arm 1.886 clear of its nearest rival is registered nowhere and printed nowhere, the numeric-token diff against fills.json comes back clean on every figure, no percentage survives, and the publication-law screen returns nothing but the allowlisted tavern and a pinned model tag; what it did find is a page that omits — the section headed \"what the panel could not agree on\" tells a reader 0 splits above six divided cells, four of them the chief rival's; six of the fifteen registered probes wrote no receipt at all, among them THE OUTSIDE READ, the one instrument this design has for answering exactly this accusation, and G-QUOTE, the gate §7.8 makes the enforcement of fabricated-never-the-beat, which the act-one prose then breaks by handing the night's warmest compliment to a cell three seats called a fabrication; not one of the 32 reply shas the page prints exists anywhere in the shipped kit, and the page promises that join once per act; the establishing section carries five asymmetries and not the sixth, A2's, which sits on the author's own row — its transport answered inside a prompt window about 1.6× the other three arms'; every author's pick prints its verdict and, against §11.3, none prints its figures, which is what would have shown a reader that GLM 5.3 out-scored the author's arm on three of the six asks while the author's arm took the beat in all three acts; the prose calls its own family's line the truest thing on the page; the balances are stamped to the second where the receipt records an operator's ten-minute eyeball read three hours earlier; the bill drops §9's one stated convention, prints no total, and its two metered rows sum past the ceiling the same paragraph names; and A6's broken clause, a duplicated recusal sentence, nine hard-wrapped bullet fragments and the retired twelve's-means slogan print beside their own replacements as a version log inside a page registered fun and warm",
 "evidence": {
  "findings_raised": 42,
  "must_fixes": 17,
  "must_fixes_landed": 17
 },
 "class_counts": {
  "A": 19,
  "B": 8,
  "C": 8,
  "D": 1,
  "E": 6
 },
 "class_basis": {
  "A": "a claim beyond its receipt — accuracy and internal consistency",
  "B": "a §9-class figure violation — denominators, orderings, mis-scoped units, numbers typed by hand",
  "C": "framing that favours the author's own arm",
  "D": "the publication law, including the written-persona law; the screen itself is clean",
  "E": "voice and the cold-scroll law — working notes, version log, broken render, unintroduced subjects"
 },
 "must_fixes_named": [
  {
   "name": "\"Where the panel could not agree: 0 splits\" on a page printing six divided cells, four of them the chief rival's, and with no denominator",
   "section": "What the panel caught, and what it couldn't agree on"
  },
  {
   "name": "six of fifteen registered probes wrote no receipt — the outside read, the quote gate, the publication screen, the calibration gate, the panel gate and the closing quota probe — and the page prints no probe ledger",
   "section": "How to check our work — and see it live / prereg §8"
  },
  {
   "name": "not one of the 32 reply shas the page prints appears anywhere in the kit, and the join is promised once per act",
   "section": "the three act folds / The kit"
  },
  {
   "name": "\"every arm faced the same six\" — the local seat faced five, by the page's own recusal table and every one of its cards",
   "section": "What changed since the open call"
  },
  {
   "name": "no author's pick prints its figures (§11.3), and the author's arm holds the beat in all three acts while a rival out-scored it on three of six asks",
   "section": "the three acts' pick blocks"
  },
  {
   "name": "three cross-arm superlatives — \"the truest thing on the page\" for the author's own family — and a SPLIT fabrication cell pull-quoted as the night's closest thing to the ancestor's best line",
   "section": "Act one / Act three / prereg §7 forbidden 1 and 8"
  },
  {
   "name": "A2's sixth asymmetry, on the author's own row — a prompt window about 1.6× the other three arms' — registered for the establishing section and printed nowhere",
   "section": "Who's at the chair tonight"
  },
  {
   "name": "the bill has no calls cell, no total, its two metered rows sum past the ceiling the same paragraph names, and the author's arm's figure is on an unregistered basis 2.6× below §9's",
   "section": "The bill"
  },
  {
   "name": "an operator's ten-minute eyeball balance read printed as a to-the-second stamp three hours later",
   "section": "The bill — the two metered balances"
  },
  {
   "name": "\"two readings of the same six questions\" — the open call judged four of the six, so the two axes pool different instruments",
   "section": "The rows on their own axis"
  },
  {
   "name": "four working-note artifacts render as published prose, including a retired slogan printing beside its own replacement and a five-sentence list rendering as nine bullet fragments",
   "section": "Seven seats, six carried / What this page does not say"
  },
  {
   "name": "Finn, Sam and Eleanor are named as written personas once, in the frame, against a law that says every time they appear",
   "section": "Three visitors, written down / all three acts"
  },
  {
   "name": "the per-sample cards print the ask's mean and the ask's cell count under a sample heading",
   "section": "the three acts' card blocks"
  },
  {
   "name": "\"all 42 are kept\" against \"all 36 of 36 anchor cells are kept\", and \"seven different minds\" over a six-seat calibration table",
   "section": "The calibration / What this page does not say"
  },
  {
   "name": "\"6 of 38 cells of every pair's denominator\" — five of the fifteen pairs have 30",
   "section": "The agreement matrix"
  },
  {
   "name": "A8's registered per-reply matched deed tokens are not printed, on the one axis A8 says the column can be wrong about",
   "section": "Act two — the deed relay"
  },
  {
   "name": "six §10 kit contents absent with nothing naming them, and the shipped index's four withheld classes carry no sha and no refusal class while its build stamp is null",
   "section": "How to check our work — The kit / release/kit-cove/index.json"
  }
 ],
 "version_read": 1,
 "critique": "article/cove/panel-v1/CRITIQUE-hostile.md",
 "critique_sha256": "c07d8bef6cf260935e46847ce9eb5020b14596e384419ef1dc3e483b3d4e01c0",
 "read_against": {
  "page": "article/cove/pour/three-new-voices-at-the-narrators-chair-v1.md",
  "page_lines": 939,
  "page_md5": "f890f4b7717d99dfe7e7b9abc4d90e7c",
  "also_read": [
   "prereg/PREREG-COVE-LEG.md with amendments A1-A8",
   "article/cove/fills.json (30 fills, 27 rendered on the page)",
   "release/kit-cove/ — index.json, README.md, stripped.json, carriage.json, auditions.json, scores.json, sheet-manifest.json, asks.json, seats.json, arms.json, ladder.json, counting-rules.json and all 42 verdict arrays",
   "results/cove/scores/SCORE-two-frontiers-cove-r1.json",
   "prereg/receipts-cove/ — 39 files",
   "plan/STORY-BRIEF-COVE-2026-09-06.md §3-§4",
   "plan/LENS-REPORT-COVE-strategy.md, -security.md, -measurement.md"
  ]
 },
 "checks_that_passed": [
  "the numeric-token diff against fills.json: four unsourced tokens on the whole page, and all four are the canon year twice, a model name and the licence version — §7's no-number-typed-by-hand holds",
  "no percentage anywhere on the page, while scores.json carries two the page declines to quote",
  "publication law clean: the gate's own alternation returns the allowlisted tavern name and one judging seat's pinned model tag, whose date suffix the port shape catches, and nothing else on the page or in the 46 kit files",
  "no persona name reaches the kit outside a judge's own note, asserted over every field of every kit file",
  "all 46 listed kit files exist at their listed sha256 and byte count; stripped.json itemises all 77 removals with reasons",
  "both registered end-paragraphs publish verbatim against their A7 templates, five pairs, with the sixth pair's absence disclosed in full — the one place a puff piece would have cheated and did not",
  "the arithmetic that can be checked, checks: both of the hosted arm's bill lines to four decimals, the metered seat's line, the caps to 5, 39 non-clean rows against 39 of 184, 38 cells per seat for 228, 48/48/48/40 and 6/6/6/5, and no google key on the local arm's leave-one-family-out"
 ],
 "not_relitigated": "the panel composition, the promoted spares, the absent head-to-head leg, the third contestant, the caps, the fence ruling and the tie band's value — all ruled before the run",
 "model_calls_made": 0,
 "no_name": "this file names no person, no box, no path and no port: an operator is 'an operator', and every port-shaped literal is described rather than printed",
 "verify_pass": {
  "version": 4,
  "utc": "20260906T201852Z",
  "must_fixes_verified_landed": 2,
  "of": 2,
  "new_must_fixes": 0,
  "page_read": {
   "file": "article/cove/pour/three-new-voices-at-the-narrators-chair-v4.md",
   "lines": 1047,
   "md5": "5da939e109cdc37a1544e71f716abab4",
   "settled": "unchanged across three minutes and two re-reads"
  },
  "kit_read": {
   "path": "release/kit-cove",
   "files_on_disk": 112,
   "files_manifested": 110,
   "built_utc": "2026-09-06T20:13:05Z",
   "withheld_rows": 6
  },
  "commits": [
   "dee8a99",
   "5fcfaed",
   "c82a73b"
  ],
  "notes": [
   "delta verify of v4: the TWO must-fixes the v3 pass raised are both LANDED and nothing else moved. diff v3 v4 is SIX hunks and six changed lines, every one expected: the byline (5), the clock sentence (828), the clock fill's own source list (830, which gained the second arm's cells_with_a_figure because the rewritten sentence needs it to say '8 and 8 of them'), the probe ledger's G-PEN row of record and its re-run tally (877, 888), and the kit stamp (997). Line count identical at 1,047; no act, card, table, figure or limits sentence moved",
   "NEW 1 LANDED — the clock lead names no arm: 'the quickest and the slowest of the 3 of 4 published arms, each the median of its own scored calls (8 and 8 of them)... one is a harness-wall... and the other a client-wall. Which arm is which, and every arm's own figure, is in the table beneath, in roster order and never sorted by it.' A scan of that line for every arm display name and arm id returns no ranked arm; the only arm-shaped phrase left is 'the local seat', in the EMPTY clause, which is a disclosure of an absence and not a position. The page-wide ordering-word sweep now returns three hits, none an author comparative: the de-named superlative, and 'walked in ahead of' twice inside quoted model replies",
   "NEW 1 figures re-derived from latency_by_arm: published_extremes.fastest.median_ms 9,333 -> 9.3 s and slowest.median_ms 33,515 -> 33.5 s, both cells_with_a_figure 8, against by_arm medians 9,333 / 22,104 / 33,515 / None, so '3 of 4 published' is the count. The scorer's field DOES name the arms and the page declines to, pointing one table down where all three figures sit unranked — the ordering leaves the prose and no data leaves the page",
   "NEW 1 residual superlative ACCEPTED and ruled: 'the quickest and the slowest of the 3 of 4 published arms' names a range's endpoints without naming who holds them, over a stated denominator, with every arm's figure unranked beneath — the same construction the anchor spread table already uses with its lowest and highest columns. §7 bans orderings AMONG ARMS; an unattributed range plus an unsorted table is not one",
   "NEW 2 LANDED — release/pour-draft41.sh lines 171-182 carry the assertion and name the finding in a comment ('the hostile reader's ordering finding on draft v3'): the steps run 1b scan, 1c rebuild and snapshot the file list, 1d fill, 1e rebuild the kit that ships with the round's screen and the hub's gate, 1f the assertion, which refuses unless the shipping kit equals the scanned kit plus the fill artifacts and the pen receipts",
   "NEW 2 verified on the round it governs: g-pen-cove.8.json is stamped 20:13:03Z, PASS over 311 files with all six shipping counts 0, and records kit_scanned files 109 with its own index sha; the shipping manifest reads 110. I identified the delta — exactly one file, receipts/g-pen-cove.8.json, the scan's own receipt, the one file it is logically impossible for a scan to have read. 110 minus 1 equals 109 equals the scan's own count; nothing else entered and nothing left. A full re-hash of all 110 manifested rows returns 0 listed-but-absent and 0 sha or byte mismatches",
   "THE FENCE DISPOSITION IS ACCEPTED, and I withdraw the port half of my own finding as a false positive of my own pattern rather than a miss by any gate. Measured, all four rules against both strings: the cove leg's own pen rule matches only the estate's actual port range and returns no match on either; public_copy's broader rule and kit_build's derivative both carry a lookbehind whose own comment says why a colon preceded by a letter or digit is an identifier and not an address, and both return no match; the hub gate's transcription at harness/forbidden_literals.txt has NO generic port-shape alternative at all, only specific numbers, and neither of the two four-digit values is among them. Only my own broader grep matched. So the pen receipt's port count of zero over a set that included the fence is correct, not blind",
   "AND I RECOMMEND AGAINST DESCRIBING THEM: scrubbing either literal would change fence-served.md's bytes, break the registered sha the page's five-line-diff and G-FENCE claims rest on, and put the kit's copy out of step with what the judges actually read — a real loss to close a match no gate makes. The fence is verified untouched, byte-identical to sheet-manifest.json's registered sha at 11,615 bytes",
   "NEW SHOULD-FIX: the assertion's baseline is a post-scan rebuild snapshot rather than the scan receipt's own kit_scanned index sha. In this round the two coincide exactly (109 either way) and the only file the rebuild can add is the receipt just written, so nothing is loose today; the stronger form compares the shipping index against the receipt's own record and names the allowance per file. Worth a line, not a re-pour",
   "the three should-fixes of the v3 pass stand unchanged, as expected of a fold that carries two must-fixes and nothing else: subject-verb agreement in two new fills, counted() as the last bare function name in the forbidden-claims list, and line 26 dating the nineteenth townsperson as 'since added' where the story brief records that soul in the campaign file five days BEFORE the seal and absent from the bundles. None is false; none blocks",
   "publication law clean on the page: two hits, the allowlisted tavern name and one judging seat's pinned model tag whose colon-and-date suffix the port shape catches, and a scan for every form of the nautical box-name token returns exactly that one tavern phrase — so the page still needs no new allowlist entry. 0 percentages on 1,047 lines; an operator keeps its indefinite article throughout; three unsourced numeric tokens against fills.json and all three legitimate prose (the canon year, a model name, the licence version) — the kit's file count is a sourced fill again, where the first v3 byte-state had it typed by hand",
   "CLEAR TO RELEASE on this lens: 2 of 2 landed, nothing else moved, no new must-fix",
   "verify report: article/cove/panel-v3/VERIFY-hostile.md, section 'V4 DELTA' (appended, not a new file)"
  ],
  "verify_report": "article/cove/panel-v3/VERIFY-hostile.md",
  "verify_report_sha256": "0d120782c8e6e4dde9b549d9db8f33d30761a94917b64e2e9ef76f121e7bf762",
  "history": [
   {
    "version": 3,
    "utc": "20260906T200835Z",
    "must_fixes_verified_landed": 17,
    "of": 17,
    "new_must_fixes": 2,
    "page_read": {
     "file": "article/cove/pour/three-new-voices-at-the-narrators-chair-v3.md",
     "lines": 1047,
     "md5": "c5b9cc17d90539eef2c9ef9f530b0699",
     "note": "v3 was re-poured DURING this pass — md5 352d481e… at 19:52:46Z, c5b9cc17… at 19:59:33Z, on commits 14e91a1 and 5712d38. Every verdict below was re-run against the second byte-state; nothing here is asserted of the first."
    },
    "kit_read": {
     "path": "release/kit-cove",
     "files_on_disk": 111,
     "files_manifested": 109,
     "built_utc": "2026-09-06T19:59:30Z",
     "withheld_rows": 6
    },
    "notes": [
     "17 of 17 LANDED, each against a quoted v3 sentence or a kit field, and every figure whose grouping the fold changed re-derived under my own arithmetic rather than read off the page",
     "the three split groupings re-derive exactly from the scorer: by REPLY 4 of 32, by ASK 2 of 24, by SCENARIO 0 of 12 against canon_tally.split_count 0 — and the page now prints all three with the paragraph saying why the scorer's own field reads zero, which is a better answer than the one I proposed",
     "the join my third finding said the kit could not make, I executed: page sha 10239757… -> replies.json (arm, ask, sample) -> sheet-keys.json (letter H on sheet S1-ask-A) -> verdicts/kimi-k3/… -> the exact judge note the page prints in its canon-findings table",
     "the picks went past the fix: all six print their figures with the unit labelled, all six re-derive from the scorer, all six are clean at reply level so G-QUOTE's PASS is genuine, the candidate lists print what was stepped over and why, and the author's own arm moved from 3 of 6 picks to 1 of 6 with a line naming GLM 5.3 as the higher family-mean on both act-two asks",
     "the bill's arithmetic re-derives: 0.5183 + 0.2469 = 0.7652 for the metered total with the over-estimate sentence, and 12 x 0.12992 = 1.5590 for the registered basis printed beside the transport's own — the registered rate applied to the calls actually made, with both call counts on the row",
     "the anchor reads 36 of the 36 that exist against the 42 a seven-seat panel would have filed, and the calibration says six minds; the agreement matrix's split re-derives twice, from score.agreement.pairs[].cells (38: 10, 30: 5) and off the page's own column (10 rows at 38, 5 at 30)",
     "the deed cells' matched tokens are the scorer's own, in the scorer's order, on all four arms; the reasoning-leak table's eight cells match reasoning_quarantine.reasoning_leaked_chars exactly",
     "the six once-missing probe receipts all exist with the verdicts the ledger prints, and the outside read's raw reply genuinely ships: evidence.reply is a 14,693-character string in the shipped receipt, matching the reply_chars and the sha the page prints",
     "a full kit audit over all 109 manifested rows returns 0 listed-but-absent and 0 sha or byte mismatches; withheld carries 6 rows, each with a refusal class and a sha or the stated reason none exists; built_utc is stamped where it was null",
     "NEW MUST-FIX 1: the clock lead prints a named two-arm ordering — 'Claude Fable 5.1 at the quick end and GLM 5.3 at the slow one' — putting the author's own arm at the good end, against §7's forbidden 1 and the page's own restatement of it nine lines later; the paragraph discloses the class mismatch honestly, which makes it worse rather than better, because the same table's note says a harness-wall figure is published as a labelled upper bound and NEVER beside a serving figure (PLAN §8.2). The two medians re-derive correctly (9,333 ms and 33,515 ms); the sentence around them is the finding",
     "NEW MUST-FIX 2: the re-pour shipped the served fence to close my seventeenth finding, and those bytes carry two colon-and-four-digit literals the publication gate's port shape matches — a config value inside a code identifier and a source-line reference. Neither is a port and neither is a redaction failure: the file's sha256 is byte-identical to the registered value, so the strings were always in the published redacted copy; what changed is that the copy entered the kit's SHIPPING scope. The screen did not see it — g-pen-cove.7.json, the receipt the page's ledger names as G-PEN's row of record, is stamped 19:59:27Z over 310 files reporting zero port-shaped hits and never mentions the file, whose mtime is 19:59:29.97Z with the index at 19:59:30.53Z. §11 registers that the allowlist covers the kit and that G-SCREEN exists to find this out BEFORE the kit is built, and the pour's allowlist patch adds nothing covering either form. The durable fix is ordering: re-run the pen and screen gates AFTER the final build and let the ledger name that receipt",
     "RAISED AND CLOSED during the pass: at the first byte-state the page read '108 manifested files', flagged by my numeric-token diff as UNSOURCED against fills.json — a number typed by hand against §7's forbidden 6 — while the manifest carried a fence-served row with null bytes, null sha and state NOT-BUILT for a file that did not exist, so a reader counting got 107. Commit 14e91a1 closed both halves; the count now reads 109, the file exists at 11,615 bytes and its registered sha, and the re-run diff returns only 1888, 5.2 and 4.0, all legitimate prose",
     "SHOULD-FIX: subject-verb agreement in two new fills ('16 receipts was filed'; five bullets reading 'N earlier readings FAILED and is kept beside it'); counted() survives as the last bare function name in the forbidden-claims list; and line 26 still dates the nineteenth townsperson as 'since added' where the story brief records that soul in the campaign file five days BEFORE the seal and absent from the bundles — a bundle-composition fact, not a date. That last was NOTE 8 of the v1 read rather than a must-fix, so the fold was right not to carry it",
     "publication law: clean on both surfaces. The page returns two hits — the tavern's name and one judging seat's pinned model tag whose colon-and-date suffix the port shape catches — and a scan for every form of the nautical box-name token returns exactly one, the allowlisted tavern phrase, so the page still needs no new allowlist entry. The kit's 111 files return that same model tag in four files plus the two fence literals of NEW 2; every one of the 61 files added since v1 is otherwise clean. 0 percentages on 1,047 lines; an operator keeps its indefinite article throughout",
     "no measurement moved: every scorer figure I checked in the v1 read re-derives unchanged — 48/48/48/40 scoring cells and 6/6/6/5 families, 39 of 184 non-clean cells, 38 filed cells per seat for 228, pairs_within_band 1 of 5 with the sixth pair still unregistered and disclosed, and the local arm's leave-one-family-out still carrying no google key",
     "two commits landed mid-pass and both were fixes rather than regressions: 14e91a1 closed the finding above, and 5712d38 folded the two rows the v1 fold had DECLINED as the author's own prose, neither of which was mine. All 17 verdicts were re-run against the final bytes and none moved",
     "verify report: article/cove/panel-v3/VERIFY-hostile.md"
    ],
    "verify_report": "article/cove/panel-v3/VERIFY-hostile.md",
    "verify_report_sha256": "2c3f323b6bd196c8319425612656220fb1a4f53b3937386bee05be4da736f8b5"
   }
  ]
 }
}