Repository path: wiki/arms/ARM-atelier-cycle.md · rendered 2026-09-09

ARM-atelier-cycle — the second long work; translation as the point again

CLOSED resolved at S120, 8 visits. «Köyhää kansaa» is translated whole (13,517 Finnish → 20,276 English, 571 paragraphs to 571, D1–D128, V1–V28, five errata, the book collated entire) and the craft report is written: wiki/findings/essays/ES-20260806-craft-report-koyhaa-kansaa.md. Its central claim is measured, not recollected — RS-20260806d-first-span-again.

Deliverable: sustained translation practice (charter §3 Track 1, "the long work"; wiki/tracks.md §Deliverables, T1). Constituted 2026-08-01 (S082) in the Tom-directed refocus, as the successor ARM-longwork named when it closed at S052: a second work, chosen for more than one unframed proverb, so that U6 could actually be settled.

Question

What does a long work teach that short pieces cannot — sustained voice, cumulative characterization, motif and terminology consistency, fatigue of method — on a second work, in a second language, now that the first (Verga, 11,474 words, six visits) is done?

The atelier has not been the point of a session since S057's revision study; the ledger recorded "prose translated in order to be measured" fifteen times in fifty sessions. This arm is where translation is the point.

Done when

The chosen work is translated whole, in spans on cadence, each span filed as a first-class artifact with its frozen log; a closing craft report states what the length made visible (ARM-longwork's closing shape), including honest nulls. Where the evaluation ladder wants a regime pair on this material, a span may be produced under two regimes as craft — the pair is banked for ARM-first-judgment's successors, and the session's writing is still about the translating.

Cadence (the budget for a serial work)

One visit at least every three sessions once started — under-visiting is this arm's alarm, exactly as over-budget is every other arm's (ARM-longwork precedent, S039/S042). Expected span count declared at step 1 when the work is chosen; 4–6 visits is the norm the estimate should justify.

Steps

Blocks / blocked by

Blocks nothing; blocked by nothing. Runs interleaved with the evaluation ladder by ordinary rotation — after ARM-tierD-run's two consecutive sessions, T1 will be near the top of the selection again.

Constraints

• Lead translation is free and never ledgered; panel contrast subjects only when a design wants one, budgeted normally.
• The log freeze before any evaluation is designed; never primed with a stored published rendering of its own source (charter A4).
• The subject rule's sentence, answered at birth: this arm teaches how sustained literary prose is translated — the craft the framework is answerable to.

Log

• S085 — step 1 done, and the work is chosen. Minna Canth (1844–1897), «Köyhää kansaa» (1886), Finnish → English. Free source verified by fetch (Gutenberg 13976, PD, stored whole at 573 paragraphs / 13,525 Finnish words). Woman author — the arm's named gap, met by a second author after Lagerlöf. New language family for the corpus: Uralic, which is the largest breadth gain available and the reason this was preferred to candidates with better comparators. Contamination declared none and NOT MEASURED, because no English rendering of this work, or of any Canth prose, exists on three documented search routes — a different position from S084's unmeasurable declaration, and contamination.md argues the difference rather than assuming it. Span plan: EIGHT spans, which exceeds this arm's 4–6 norm, and the overrun is declared now rather than discovered at span 5 — span 1 measured Finnish→English expansion at 1.50 (against the 1886 Swedish translation's 1.34 over the whole novella), projecting ~20,300 finished English words, 1.8× Verga. The arm may prefer to revise the cadence instead of the span size; the arithmetic is on R05-v1/translation.md so the call can be made. Span 1 frozen (paragraphs 2–75; 1,795 Finnish → 2,699 English) with a 23-decision log and register.md at ten bound rules. Study limb: RS-20260801h-source-set-or-pair.

• S090 — step 2 done, and the arm's cadence alarm was what selected it. Last worked S085; this is S090, so the declared cadence (one visit at least every three sessions) was overrun by two sessions, and check_balance.py structurally cannot see it — the tool's own self-test asserts that a budget: cadence arm never fires a budget violation. That, not the track count, is the reason this arm was taken over the tool's selection. Translation limb: span 2 frozen at 4a87026 before any comparator was opened — paragraphs 76–133, 1,782 Finnish → 2,758 English words (factor 1.55, against span 1's 1.50), 58 source paragraphs to 58 English paragraphs, 23-decision log D24–D46. Study limb: RS-20260802d-class-line-carry — 13 loci where span 2 crosses the class line, scored against Hertzberg's authorized 1886 Swedish on two axes, capability and choice, plus a 9-call contrast probe (E-20260802d). The finding: span 1 found Swedish's proximity dissolving three English problems, all grammatical. Span 2 finds one running the other way — and it is the device Canth uses at both gentry scenes. «mutta ruotsiksi, ettei Mari ymmärtänyt» depends on Swedish being a switch away from the language of the book; Hertzberg keeps the words and loses the structure, because in his target the marked language is the medium. English is the better-placed target there. Proximity buys grammar (4 address loci English cannot carry) and costs position (1 locus Swedish cannot). Two errata, and the second was forced by the comparator rather than by the probe. V10 fell because D11 wrote a rule for a site outside its span on the false ground that the formula "recurs" — it occurs once in the novella. V14 fell because it asserted no compensation existed for the child→parent honorific: 6 of 6 cold machine cells marked nothing, and Hertzberg marked it with a kin term. New standing rule V17: no register rule is written for a site outside the span being translated. Costs $0.108764428 of a $0.67 declared worst case; critic NEEDS-AMENDMENT, five findings, three BLOCKING, all five accepted, and A2/A3 are what made V14 refutable. Verifier 61 checks, 0 failures, three mutation tests all caught. One check failed first and the defect was in the check, recorded rather than silently repaired.

• S095 — step 3 done, and a decision taken in the morning was reverted by measurement in the afternoon. Selected by the tool (T1 at 4) and by the cadence, which was overrun by two sessions for the second visit running. Translation limb: span 3 frozen at 1218096 before the Swedish was opened — ¶134–206, 1,795 Finnish → 2,778 English (factor 1.548, against 1.50 and 1.55), 73 source paragraphs to 73 English paragraphs, log D47–D64. Mari, refused a loan, meets a woman poorer than herself on the road and carries her child to the bishop's; then the gentry come down into the poor room while she is asleep on the floor. All four §Unresolved questions settled, and two of the answers were not the ones the questions expected. The direction reversal does not reverse the address forms — it abolishes them: in the whole visit the two ladies speak three times, all three to a four-year-old boy, and never address Mari at all, while Mari uses no second-person form of any kind. V13 gets no second clause; its site did not recur. And V14a's licensed kin-term compensation was unavailable at ¶175 because Canth has already spent the vocative — a compensation competes for the same slot as the source's own marking, reached before the comparator was opened and then confirmed by it. Study limb: RS-20260803b-honorific-carry / E-20260803b, 18 cells, 3 arms differing by one string. The compensation worked — +1.167 standing, exact permutation p = 0.0368, 3 of 3 seats, 5 of 6 quoting it, control unmoved — and the naturalness tripwire frozen before dispatch reverted it anyway (0.833 against 0.50). Erratum 4; V18 retired the day it was written. The no-information arm is what the session will be remembered for: seats shown no text at all rated the girl's deference at 3.667, above both English versions. P − N = −2.167 — the honorific loss is worth about 2.2 scale points, and the best compensation English offered recovers half of it. The first time this project has priced a grammatical loss rather than declaring one. The comparator settles the capability question. Hertzberg uses Swedish ni eight times in this span and then writes the singular «Stig upp!» at ¶175. Twenty-one second-person sites, Swedish carries twenty, and the one it loses is the only one it chose to lose. Costs $0.263880726 of a $0.94 revised worst case. Critic TWO passes — NEEDS-REDESIGN (nine findings, five BLOCKING) then NEEDS-AMENDMENT (nine, three BLOCKING) — all eighteen accepted; the first pass established that no outcome of the design could have changed the translation, and the amendment that fixed it is what produced Erratum 4. Verifier 31 checks, 0 failures, two mutation tests, both caught. One defect recorded rather than repaired: the second critic dispatch overwrote the first one's stored body.

• S100 — step 4 done, and the text the arm has been translating from turned out to be the wrong text. Selected by the tool (T1 at 4) and by the cadence, overrun by one for the third visit running. The gate first, because the arm had declared it a requirement. register.md §Settled-5 wanted a second printing before span 4. The three e-texts the project could reach are one transcription of one posthumous reprint (Gutenberg 13976 = Projekti Lönnrot 99 = Finnish Wikisource, all Tapio Riikonen from Otava 1917, all reproducing the same corruptions character for character). The 1886 Edlund FIRST EDITION is free, scanned, and one search away at Doria. It reads Ei at ¶94 and ¶197 where the working text reads EL and el — both project renderings vindicated — and it carries ¶228's unbalanced quotation too, so that one is Canth's problem and not a printer's. Note (bif). Then the collation, and it is the session's finding. Span 4 diffed in full: 8 substantive variants in 1,800 words. Four are the reprint standardising the eastern-dialect layer (uhhaa→uhkaa, mahallaan→vatsallaan, lepoaa→lepoa, tyyneemmällä→tyynemmällä) — the exact layer V3 declares a loss the translation carries by syntax alone. Two are reprint readings that are not construable (vastamaan→katastamaan, kuule→kuulee). Three are the first edition's own slips, where 1917 is right. Neither printing is clean, so V21's policy is eclectic: 1886 is the copy-text, 1917 followed where 1886 slips. collation.md. Translation limb: span 4 frozen before Hertzberg was opened and before the study limb was written — ¶207–295, 1,800 Finnish → 2,747 English words (factor 1.526), 89 source paragraphs to 89 English, log D65–D83. The neighbours come in, the baby is dosed with Hoffmann's drops, and five people with no bed between them work out what eighty pennies a day on the new railway will mean. The title phrase «köyhää kansaa» occurs exactly once in the novella and it is Lopo who says it, here (D75); the English title is fixed as Poor People so the sentence carries it. Study limb: RS-20260803g-distance-not-deference. Span 3's registered prediction held at 15 sites out of 15 — every second-person form between acquainted adults of the poor class in span 4 is T-singular, not one te. Then the whole-text distribution broke the project's own vocabulary: te is used mutually by landlord and tenant, by servant and pauper, by pauper and beggar-woman, so between adults it does not mark rank at all. V13's and V14a's "pronominal honorific" is wrong as a description; te is the non-intimate form (V23), rank is marked by the third person and by vocatives, and the honorific reading survives at 4 of 30 sites, all a daughter to her mother, licensed by asymmetry-inside-intimacy. And the register's count of those four was wrong by half. V18a said "both known sites"; a whole-text search returned four, one of them in span 1, already translated and behind the span-by-span counter. Note (bie). The same search found a third verb in a family closed twice at two (D74). Costs $0.097558200 of a $0.42 declared worst case, one adversarial seat of two. P1 returned zero characters of content on both dispatches and was withdrawn under the design's own FC4 rather than replaced after the first verdict was in hand. P3's three findings were all accepted, all corrections against the lead — including the one that re-worded this arm's own headline claim. The key-usage cross-check did NOT close, and the residual was a fourth request the shell had killed mid-flight and OpenRouter had billed anyway: note (bid), and the delta is the ledger's actual.

• S105 — step 5 done, and the gate found more than the span did. Selected by the tool (T1 at 4, the most neglected arm) and by the cadence, overrun by two sessions for the fourth visit running — S085, S090, S095, S100, S105, a clean five-session rhythm against a declared three. The gate, and it is half the session. register.md §Unresolved 1 required ¶2–206 collated before any new prose. 5,372 words, 62 filtered rows, 16 substantive divergences — and nine of the sixty-two dissolved on the page image, three of which would have been published as real variants and one of which (murjottivat for muljottivat, sulked for goggled) would have changed an English word. Note (bim). Erratum 5 changes three sentences (¶73, ¶105, ¶163); two more renderings the collation vindicated (¶102, ¶134). A third corruption of the EL/el letter-substitution family at ¶134, and the first of the three to fall inside a clause the translator had had to guess at — the guess was right. And the reprint's editorial hand is not one-directional: it standardises the colloquial layer at five sites, runs the other way at three, and adds a word at two. collation.md §§7–9. Translation limb: span 5 frozen at 65e5f8a before the Swedish was opened — ¶296–375, 2,041 Finnish → 2,982 English (factor 1.461, the lowest of five), 80 source paragraphs to 80 English, log D84–D98. Mari's prayer turns into a psychosis, she strikes her husband in the dark, and he lies awake till morning deciding it is only tiredness. The first span translated from the copy-text rather than collated afterwards, which is what V21 was for. Two craft results in the log. D85: the first edition answers «Oletko sinä kipeä?» with «Ei» at ¶320 and «En» at ¶348 where the reprint prints En twice — a colloquial agreement contrast, and the first site in five spans where V3's "carried by syntax only" clause actually paid, rendered "No." against "I'm not." D88 / V24: «Täytyi olla hiljaa» is a zero person with no subject at all, and the English bare imperative is subjectless too — "Keep still ... keep still...", the refrain kept without V15's generic adding a subject the source does not have. D89 discharges D17's four-span-old suspicion about tuuditti uskoa: it occurs twice in both printings, so the slip hypothesis is refuted, and Hertzberg independently read it as vaggade träget. Study limb: RS-20260804e-whose-mind / E-20260804f, 16 loci × 3 blind seats. The attribution survives whole — 47 of 48 cells, 30 of 30 on free indirect discourse — and the lead's registered expectation that it would fail at the switch between minds was wrong at 9 of 9. The finding is the mechanism, and the registered mechanism prediction failed too: English supplies a person-index Finnish hän does not have, and the seats did not cite it — 9 of 30 quotes, against a registered ≥ 20. 21 of 30 cite the diction, the generic and the modality — the stratum V15, V20, D90 and D91 are about. At ¶365 the extract contains "Mari" in the third person and not one seat quoted it. The Swedish comparator closes it: Hertzberg carries the clitic with ju and väl, English has neither and buys the function back with a body knew, surely and Why, — and at ¶297 the two translators kept opposite halves of one construction. Pre-run critic NEEDS-AMENDMENT, nine findings, four BLOCKING, all nine accepted — F1 showed the primary was reachable on a proper-name cue and it was withdrawn and replaced; F4 forced a ground-truth cross-check the design had asserted and not run, and the assertion was false at ¶357. Verifier 24 checks, 0 failures, three mutation tests, all caught. Costs $0.132372 of a $0.75 declared worst case, 3 of 3 seats first dispatch, zero retries. The key-usage cross-check closes to 1e-9 — after a second reading; the first snapshot lagged by two seats' billing, which is note (bil) and is very likely what S104 saw.

• S110 — step 6 done, the collation gate closed for good, and the study limb returned a null that convicts its own instrument. Selected by the tool (T1 at 4, the most neglected arm) and by the cadence, overrun for the fifth visit running — S085, S090, S095, S100, S105, S110, a clean five-session rhythm against a declared three. The cadence has now been overrun every single visit; either the arm's declared three is wrong or the rotation cannot deliver it, and the craft report must say which. The gate, and it finishes the job. register.md §Unresolved 1 required ¶376–572 — 4,308 words, 32% of the novella, the last unchecked stretch. 68 surviving rows, 23 real print divergences, every one adjudicated against the page scan (fourteen scan pages read as images). A fifth demonstrated corruption in the reprint at ¶490 — «jutteli Heiskasen kanssa» set as Heiskanen, which kanssa cannot govern — found only because a diff row that looked like OCR noise was checked against the image anyway. The reprint also adds words at ¶502, ¶538 and ¶549 and removes jo at ¶554, so §4's account of its hand as one that removes is now wrong in both directions and §11 rewrites V21's reason. The whole novella is collated: 13,522 words, 59 divergences, 4.4 per 1,000. Translation limb: span 6 frozen at 9c41878 before the Swedish was opened and before the study limb was designed — ¶376–441, 1,732 Finnish → 2,516 English (factor 1.452, the lowest of six), 66 source paragraphs to 66 English, log D99–D112, thirteen copy-text readings applied before a word was written. Anni dies in the cradle while her mother watches without feeling it, and by evening Mari is tied hand and foot on the heap of clothes. Two craft findings in the log. D100: the first edition prints the poor people's herra in lower case where the reprint capitalises, and because this book's word for the class above them is herrat = the gentry, that looked like a voice device. A whole-novella census killed it — Jumala is capitalised at 31 of 31 and the two capitalised mid-sentence Herra tokens are inside poor characters' free indirect discourse too, so the split does not fall on who is speaking. The reading that lost is recorded because it was attractive. D103 / V25: Tiina Katri is a within-speaker control — sinä to Mari and te to Holpainen, one household, one day, both mutual — so te is non-kinship, not the "non-intimate" V23 called it. Study limb: RS-20260805b-address-or-action / E-20260805b, 3 arms × 3 blind seats. The four-category instrument returned the same three letters in 9 of 9 bodies and FC6 fires: H1 is withheld as a floor effect, exactly as both pre-run critic bodies predicted. The ordinal measure moved — 3 of 3 arm-1 seats put Tiina Katri–Mari above Tiina Katri–Holpainen, the ordering the Finnish pronouns draw — but that is three seats, p = 0.125, and it is not claimed. The finding is where they got it: 3 of 3 cited what Tiina Katri does, not one word she says. Which is why stripping the three "Mari" vocatives (arm 3) moved one seat, and why inserting "Mr Holpainen" (arm 2) moved no letter and no ranking — only what the seats quoted. The scene chosen because it contains the grammar is the scene whose action gives the grammar away, and that goes to the craft report. Pre-run critic NEEDS-AMENDMENT, six findings, four BLOCKING, all six accepted as A1–A6 — F4 corrected the design against its author, showing that §1's vocative check was a claim about the book stated as a claim about the scene, and that correction is what built arm 3. Costs $0.143528033 of a $1.40 revised worst case, 9 of 9 seat bodies first dispatch, zero retries; key-usage cross-check closes to 1e-9. Two defects recorded rather than repaired quietly: FC2 as frozen could not detect the ceiling it existed for (note (biu)), and a second billed critic body sat unread because a format guard rejected it — it held two findings the first did not (note (biv)).

• S115 — step 7 done, «Köyhää kansaa» is translated whole, and the plan was revised rather than broken. Selected by the tool (T1 at 4, and the most neglected arm on the board) and by the cadence, overrun for the sixth visit running — S085, S090, S095, S100, S105, S110, S115, a clean five-session rhythm against a declared three, without one exception in the arm's life. The span plan went from eight to seven, declared before translating (D113). Two reasons: the collation gate that took half of S105 and a third of S110 was discharged at span 6, so this span began on settled text and no gate was owed; and the remainder was 2,572 words, 26% above the largest previous span rather than twice it. Step 8 is now the craft report alone. Translation limb: span 7 frozen at a3b9a0b, before the study limb was designed and before the Swedish comparator was opened — ¶442–572, 2,572 Finnish → 3,796 English (factor 1.476), 131 source paragraphs to 131 English, log D113–D128, ten copy-text readings applied before a word was written. Doctor Wialén ties Mari with her own baby's swaddling-bands; the pastor and the doctor walk away arguing about euthanasia and the causes of poverty; Mari is carried a verst out of town and shut in a board box in the children's building; and the novella ends with Holpainen digging a railway cutting in trousers torn behind, hearing her. Whole work: 20,276 English words from 13,517 Finnish, 571 source paragraphs to 571 English with no split or merge anywhere. The step-1 projection of ~20,300 was right to 0.1% and its reasoning was wrong (D128): it assumed span 1's factor would hold, and the factor rose for two spans and then fell for three, so the error cancelled. Three findings the length made visible, and all three are craft-report material. (i) The address grid closes and V25 is falsified by it (D119, V28). Heiskanen says sinä to Holpainen and to Tiina Katri, while those two say te to each other — same class, same acquaintance, neither kinship nor gender separating them. Four cells of the grid are determined by the situation and the fifth is a property of the speaker; V23 and V25 were both hunting a rule in the one cell that has none. What English loses in this book's address system is precisely the part that is characterisation. (ii) The zero person has a third class and it is the doctor's (D120). ¶484 carries four agent-suppressing zero persons in one paragraph — his indictment of war never names who declares it — and they take English passives, never V15's generic, which would put a person where the source deliberately puts nobody. Whole-work count 53. (iii) V1, V16 and V15 were each found under-specified by a referent they had never been tested against (D117, D118, D121), which is the same failure three times and is the register's structural blind spot. Study limb: RS-20260805g / E-20260805g, 3 arms × 4 seats. The primary is withheld by its own pre-registered gate G3: EXCL is 0 in twelve of twelve bodies, in every arm, including the baseline. A four-to-six-sentence retelling has room for the plot and nothing else, and all twelve spent it on the plot — the RS-20260805b lesson arriving a second time in six sessions. What the run did establish disqualifies the instrument: every A1 seat reported the content of the untranslated Swedish, one of them naming the language with no warrant on its page and one translating it silently. D115's declared cost depends on the target reader not being able to read that line, and a panel of multilingual models cannot be made into a monolingual reader by any prompt. The rendering stands as frozen; its cost is now declared as unpriced rather than claimed. Pre-run critic THREE passes, all NEEDS-REDESIGN — nine, thirteen and nine findings, twenty-two then nine BLOCKING, all thirty-one accepted. Pass 1's F1 showed the primary measured orthography; pass 2's F4 rewrote the question by pointing out that the scene carries the same inference by four routes that are not the foreign line; pass 3's F4 showed the "uncued" primary was cued, because a single-pass model reads the whole prompt before generating — Part 2 was deleted and the confirmatory test withdrawn, before dispatch. No fourth pass, on a policy frozen before the third. Costs $0.146723450 of a $1.03 declared worst case, 14%; key-usage delta closes to the four-token hand diagnostic that found the second defect. Verifier 60 checks, 0 failures, five mutation tests all caught. Two defects recorded rather than repaired quietly: qwen3.7-max burned 30% of the run on six empty length bodies (note (bhf), a fourth slug) and was dropped rather than given a bigger cap; and call.py was throwing away the HTTP error body, so six mistral 400s were undiagnosable until one call was reproduced by hand.

• S120 — step 8 done, the craft report is written, and THE ARM CLOSES resolved AT 8 VISITS. Selected by the tool (T1 at 4, and the board's most neglected arm at 4). The cadence was not overrun this time and could not have been: the arm's last visit was S115 and this is S120, the same five-session period as all seven previous visits — which is the craft report's §2 and note (bjp). A three-session cadence over a five-track rotation was unmeetable the day it was declared, and no session broke discipline. Translation limb — and it is a reviser's pass, which is what makes the study limb possible. Span 1 (¶2–75) re-rendered whole from the copy-text under the closed register, blind to R05-v1's span-1 English and to its log D1–D23, filed as T-koyhaa-kansaa-R05-v2 with a 16-decision log R1–R16, frozen at fdc8944 before the diff existed. 1,795 Finnish → 2,764 English (factor 1.54 against v1's 1.50), 74 source paragraphs to 74. The copy-text gate closed first: the seven collation.md §7 first-edition readings applied before a word was written. The design and its five named site predictions were frozen at 0912112 BEFORE a word was translated, which is what makes the predictions worth anything. Study limb: RS-20260806d-first-span-again / E-20260806d. 234 coded divergence sites; 13 attributable to something learned after span 1; 221 free variation. Eleven of the thirteen are sites the closed register would now render differently and no erratum ever caught — which measures, on a second work, the failure ARM-longwork named on Verga at S047: the register catches forward drift and is structurally blind backwards. (A) COPY-TEXT is ZERO: all seven first-edition readings changed not one English word, and the single span-1 site where the copy-text ever paid had already been retro-fitted by Erratum 5. The named-site score is 2 of 5, and it is 2 of 5 because the critic refused the lead's first scoring. N1 failed — D89 (span 5) fixed a rendering span 1 had already made. N3 and N4 are scored failed as written, not "held on substance". P4 came back at 1.00 on every countable rule — the mother 16/16, 86/86 double quotes, 74/74 paragraphs, husband 0/0 — and licenses nothing about learning, because agreement at rule-governed sites is the cheap direction and the design said so before the run. Two sites where the reviser's pass went BACKWARDS, named because the frozen scheme had no class for them: ¶70, where v1 kept Canth's happamet/karvas distinction and v2 collapsed it; and ¶6, where v2 supplied a subject the Finnish withholds — the very move it marked against v1 at ¶68. Pre-run critic NEEDS-AMENDMENT, eight findings, six BLOCKING, all eight accepted, run against the design AND the completed coding before a single seat call. Finding 1 — P1 is unfalsifiable by construction, because the predictor is the coder and (D) is the residual — governs the result page and forbids the sentence the arm wanted to write. Finding 2 found the coding protective; two of its three sites were taken, one overruled with a written reason, and the re-pass it demanded found a fourth. Attributable sites went 10 → 13, so P1 (≤ 12) FAILS, and it fails only because the finding was applied. Blind run: 3 seats × 28 items, 84 of 84 cells first dispatch, zero retries. FC1 passes (0 of 4 REPEAT majority S, unanimous F), FC2 passes (4 of 4 WRONG, unanimous S), FC4 fired before dispatch (EA n=2 < 3) so RT2 is not reported, and RT1 was declared NON-CONFIRMATORY before the run on the critic's finding 3. RT1 = 1.000 and is a description. Costs $0.109028580 of a $0.90 declared worst case, 12%; the critic was 34% of the spend and changed the result. Verifier 94 checks, 0 failures, five mutation tests, five caught. Key-usage delta closes to 0.000000000 after a pause — note (bil), third firing. Two defects recorded rather than repaired quietly: the diff's punctuation counter could not see a paragraph whose only difference was a lone em dash (found by a consistency check, not by the counter), and the first verifier build contained a vacuous mutation test that could not fire. The successor question this arm hands on (craft report §7): a third long work needs a first span translated twice by two different hands, because everything measured here is one translator's variance with itself, coded by the person who produced it.