Repository path: workshop/experiments/E-20260806-carrier-or-knowledge/design.md · rendered 2026-09-09

E-20260806 — is a narrator's person carried by the five properties voice names, or by the one it omits?

Frozen 2026-08-06, S117, ARM-voice-persona step 2. No call has been dispatched. The three renderings and their logs were frozen first; the specifications were committed before any of them existed, at 96249bb.

AMENDED after stage 0, before any other dispatch — critic.md, amendments A1–A6. The independent pre-run critic returned NEEDS-AMENDMENT, six findings, four BLOCKING, all six accepted (one in part, with the departure recorded). The amendments add a fourth rendering (A′, T-zapiski-oct3-R19-v4), rebuild the content-parity screen that was blind to the arm it most needed to screen, raise the manipulation check from two seats to three with numerical isolation thresholds, raise the primary from three seats to four, add failure criterion F7, and narrow what the primary is licensed to say. Every amended passage below is marked [A<n>].

Every sentence this design licenses is provisional. Tier D is NOT PASSED. No translation is judged in this run. Every seat is asked who is speaking or what does this text say; none is asked whether anything is good, and nothing here may support a quality claim about any member.

1. The question, and why the arm is spending its second session on a run

wiki/goodness-senses.md §voice asserts that a narratorial perspective's "observable carriers are register, rhythm, diction temperature, idiosyncrasy and distance." RS-20260805d §4 measured those five against three properties the entry does not name and found the unnamed three moving more (1.667) than the five named (1.067), with idiosyncrasy saturated at 7 in all eight profiles — a scale that agreed Multatuli's prose is mannered and discriminated nothing.

ARM-voice-persona step 2 is "carry the verdict into the entry." The verdict that would change the entry most is the carrier finding, and it cannot be carried on RS-20260805d alone: one passage, one author, one language, one translator, and — decisively — that run never manipulated the carriers. It manipulated a person and measured the carriers as a side effect. A side-effect measurement cannot say which properties do the work, because nothing in it was held.

So this run manipulates the properties directly and disjointly, in a second language and a second author, and asks the only question that discriminates:

Hold a narrator's self-knowledge fixed and move all five named carriers as far as the propositions allow; then hold all five fixed and move his self-knowledge. Which change makes a source-blind reader say he has met a different person?

What this teaches about translating literature (the subject-rule sentence, wiki/tracks.md): it tests whether the features a translator is told to hold on to when carrying a narrator across a language are the features that decide whether the reader meets the same man.

2. Materials

The six declared carriers K1–K6 and both variant specifications are at materials/specifications.md, frozen before drafting.

3. Why N is in the design, and what it is a control for

A rewrites nearly every word; B changes eight sites (its log enumerates them exhaustively). A raw comparison of C–A against C–B is therefore confounded with the extent of textual change, and that confound is named here rather than discovered later.

N is the extent control. It rewrites nearly every word of C while aiming at nothing, so it measures how much "different person" a reader will report for wholesale rewording alone. It is the floor A must clear to have bought anything with its five moves, and — because B rewrites far less than N does — any separation B achieves over N is achieved against the extent confound rather than with it.

This is a deliberately weaker use of the paraphrase arm than RS-20260805d's F4, and the difference is registered here before the data. There, PN reaching parity with the deliberate rewrite withheld an absolute claim that a persona reaches readers, and rightly. Here the primary is a contrast between two manipulations, and rewording noise common to both does not undermine a contrast the way it undermines an absolute claim. N therefore gates the primary only under F3 below — when it beats both manipulations, i.e. when neither is distinguishable from noise.

A second, mechanical extent measure is registered as a secondary: token-level change fraction from C, computed by analysis/extent.py (difflib.SequenceMatcher over lowercased word tokens, 1 − matching ratio), reported for A, B and N, with each arm's mean rating divided by it.

Computed before dispatch and it is the sharpest fact in this design: A changes 0.7012 of C's tokens and B changes 0.0396 — a factor of 17.7. So P1p and P2p are not symmetric bets. P2p (the entry's claim) can be satisfied by extent alone and would be weak evidence even if it fires; P1p requires the arm that rewrote 4% of the text to outrun the arm that rewrote 70% of it, and cannot be explained by extent at all. A P2p firing therefore does not vindicate the entry's list, and this is registered now, before the data, so that it cannot be decided afterwards.

[A1] The demotion of P2p is now formal: a P2p firing may be reported only as not distinguishable from extent, never as support for the entry's carrier list. And F7 is added at §6: if A does not beat the aimless floor N, P1p is unreadable, because B would then be outrunning noise rather than a manipulation.

[A5] The extent confound is no longer a single comparison but a ladder of four arms, the critic's fix (b) to finding 6:

A′ is 4.5× B's extent where A is 17.7×, so the B-vs-A′ contrast is the fairer one and is registered as P1q beside the primary rather than instead of it.

4. Procedure

Five stages. Every raw body is written to runs/ before any parse (note (bdt)); finish_reason: length is a seat failure, never a partial answer; a reserve is declared per stage before dispatch (note (bfc)); key-usage snapshots are write-once (note (bgq)).

• Stage 0 — independent pre-run critic. One seat, non-panel, shown this design, the three specifications, all three renderings and all three logs. Reserve: qwen/qwen3.7-max. Findings are applied before any other call is dispatched and the amendments are recorded in critic.md.
• Stage 1 — N. One call to mistralai/mistral-medium-3-5 (non-panel, as E-20260805d), prompt: the text of C and "Say the same thing a different way." Nothing about persona, voice, narrators, register or self-knowledge. Reserve: z-ai/glm-5.2.
• Stage 2 — the manipulation check (transmission, not a finding). [A3] [A5] Five texts (C, A, A′, B, N) × three seats: z-ai/glm-5.2, qwen/qwen3.7-max and mistralai/mistral-medium-3-5. Each seat rates one text, with no comparison and no source, on the six declared carriers K1–K6, 1–7. mistral-medium-3-5 writes N at stage 1 and therefore does not rate N — 14 calls, and N's profile rests on two seats where the other four rest on three, declared here rather than discovered later. max_tokens 6,000. Reserve: P4.
• Stage 3 — the content-parity screen. [A2] [A5] Three pairs (C–A, C–A′, C–B) × two seats: z-ai/glm-5.2 and qwen/qwen3.7-max. Each is shown two passages and asked to enumerate every difference in what the two texts assert, excluding only differences of wording, tone, formality and sentence length, and to label each difference EVENT (something that happened, or something stated about a person other than the narrator) or SELF (something stated about the narrator's own thoughts, habits or handling of his material). No exclusion category is named to the seat and neither label is flagged as the one the design cares about. max_tokens 6,000. Reserve: P4.
• Stage 4 — the primary. [A4] [A5] Four pairs (C–A, C–A′, C–B, C–N) × four seats (P1 openai/gpt-5.6-terra, P2 google/gemini-3.6-flash, P3 x-ai/grok-4.5, P5 deepseek/deepseek-v4-pro) × two presentation orders = 32 calls, each independent and stateless. None of these four seats holds any other role in this run — the critic, the paraphraser, the three manipulation-check seats and the two parity seats are all drawn from outside them. Each call shows two passages labelled Passage 1 and Passage 2, no source, no authorship, no mention that either is a translation, and asks:

  Q1. On a scale of 0 to 6, are these two passages narrated by the same person? (0 = certainly the same person; 6 = certainly two different people.) Q2. In one or two sentences, what in the words makes you say so? Q3. If you judged them different, name the single property that differs most.

max_tokens 4,500. Reserve: z-ai/glm-5.2. Judgment is not parallelised across a seat's own cells: each cell is one call and no seat sees another seat's answer.

5. Predictions and the primary statistic — registered before dispatch

Let r(X, s, o) be Q1's rating for pair C–X, seat s, order o; six cells per arm.

[A4] d (s,o) = r(B,s,o) - r(A ,s,o)    8 paired cells (4 seats x 2 orders)
[A5] d'(s,o) = r(B,s,o) - r(A',s,o)    8 paired cells
     T = sum d(s,o)                    exact sign-flip permutation over 2^8 = 256

• P1p — THE PRIMARY. mean d ≥ 1.00 and d > 0 in ≥ 6 of 8 cells. Holding: moving the one property the entry omits changes who the reader meets more than moving all five it names.
• P1q — the extent-nearer contrast [A5]. The same bars on d′. A′ is 4.5× B's extent where A is 17.7×, so P1q is the fairer test and is reported beside P1p, never instead of it. A design that swapped its primary after seeing the critic would be choosing its statistic; this one adds and keeps both.
• P2p — the registered alternative, which is the entry's own claim. mean(-d) ≥ 1.00 and d < 0 in ≥ 6 of 8 cells: the five named carriers are where the person is. [A1] Reportable only as not distinguishable from extent — see §3.
• Neither firing is a null on the contrast, and it is a real outcome: the two manipulations are not distinguishable at this power.
• P3d (descriptive). Q2/Q3 free text for C–B names knowledge, irony, self-awareness or self-relation more often than for C–A; C–A's names diction, formality, register or sentence length more often than C–B's. Coded by two independent string-match passes in analysis/score.py against keyword sets fixed in this file (below), never by hand.
• P4d (descriptive). mean r(N) < mean r(A) and < mean r(B).
• P5d (transmission check, never reported as a finding). Stage 2 profiles move as specified: A moves K1–K5 more than K6; B moves K6 more than K1–K5.

Keyword sets, fixed here (case-insensitive substring, on Q2+Q3 concatenated). KNOW = {self-aware, self-awareness, self-knowledge, aware of, irony, ironic, ironical, wry, self-deprecat, knows himself, sees himself, self-critical, candid, honest with himself, insight, detach} · SURF = {register, formal, formality, diction, vocabulary, syntax, sentence length, sentence structure, latinate, elevated, style, tone, ornate, verbose, colloquial}

The lead's own expectation, registered before dispatch as RS-20260805d §6 did. I expect P1p to fail and P2p to come nearer: point predictions mean r(A) = 3.5, mean r(B) = 3.0, mean r(N) = 2.0, i.e. mean d = −0.5. The reasoning is that A's change is enormously louder and readers asked same person? will hear the loudness. I am genuinely unsure of this, and it is written down so the run can be wrong about it in public.

6. Failure criteria — pre-registered, each with what it does

• F1 — the manipulation did not isolate. [A3] With Δ_X(K) = profile_X(K) − profile_C(K) averaged over the three stage-2 seats, all three of the following must hold, and F1 fires and the primary is WITHHELD if any fails: mean|Δ_A (K1..K5)| - |Δ_A (K6)| >= 0.50 mean|Δ_A'(K1..K5)| - |Δ_A'(K6)| >= 0.50 |Δ_B(K6)| - mean|Δ_B(K1..K5)| >= 0.50 What a firing would establish is not nothing: these six properties are not independently movable in translation, which is itself a claim about the entry's list. Stage 4 is dispatched either way — the departure from the critic's requested fix, with its reason, is in critic.md §F3.
• F2 — content parity. [A2] If the union over the two screen seats of EVENT-labelled differences exceeds 2 for any of C–A, C–A′, C–B, that arm is WITHHELD. The screen's own counts and the screen's own labels govern; the lead does not adjudicate them (RS-20260805d §3). SELF-labelled differences are reported and do not gate — and they are a check on the lead's own logs, which declare three for B and none for A and A′. If the screen finds more than the logs declare, the logs are wrong and the result says so. A declared, unscreened difference, stated here in advance: B's log enumerates three added propositions, all of them first-person statements about how the narrator handles his own material, and the screen is instructed not to count them. They are the manipulation. Self-knowledge has no purely substitutional exponent in English — B's log argues this at length — so a design that demanded zero addition would be a design in which K6 cannot be moved at all. What follows is a hard limit on the primary and is written into any result: B differs from C by three self-descriptive assertions as well as by their marking effect, and this run cannot separate the two.
• F3 — the floor beats every manipulation. If mean r(N) is ≥ the mean of all three of A, A′, B, the primary is WITHHELD: no manipulation is distinguishable from wholesale rewording.
• F4 — seat degeneracy. [A4] A seat returning an identical Q1 rating across all eight of its cells is not discriminating. If ≥ 2 of 4 seats do, the primary is WITHHELD.
• F7 — the loud arm must beat the floor. [A1] If mean r(A) ≤ mean r(N), P1p may not be read as support for anything: A would not then be a manipulation that reached readers at all, and B outrunning it would be B outrunning noise. P1q is subject to the same condition on A′.
• F5 — order. If the mean absolute within-pair order difference exceeds 2.0 scale points, every magnitude claim is withheld and only signs are reported.
• F6 — dispatch. Fewer than 5 of 6 admissible cells on any arm withholds that arm.

7. What this run cannot establish, whatever it returns

[A4] This run supports a pilot-level signal on one passage, one author, one pair, one translator, and licenses no generalisable claim about voice. Eight paired cells over four seats is enough to see a large effect and not enough to size one.

[A6] What the primary is licensed to say, fixed here before the data. The critic's finding 6 is that this design cannot isolate a latent property; it can only compare operations a translator can actually perform. So the licensed sentence is about operations:

A translator who changed N% of a passage at the site of the narrator's self-relation moved source-blind readers' judgment of who was speaking further, or less far, than one who changed M% of it along all five properties the entry names.

Not the five carriers do not carry voice. A′'s log already establishes, from the translator's side and before any reader was asked, why the two operations cannot be made parallel: K4 and K5 are located properties that ten edits move, K1 moves partially, and K2 and K3 are distributed over every sentence boundary and every concrete noun in the paragraph.

• That the lead's Poprishchin is Gogol's. No source-side yardstick is in this design, deliberately: the question is which properties carry a person, not whether the person is apt. voice is a source–target relation sense and this run measures only the target-side half of it.
• That any member is good. Nothing is judged; Tier D is NOT PASSED.
• Anything about voice beyond this passage, this pair, this translator.
• Anything that separates B's three added propositions from their marking effect (F2).

8. Contamination

Declared suspected on all three members and measured after the freeze, not before, because this design's validity does not turn on independence from any published rendering: all three members are the lead's, compared against one another, and an overlap with Field 1916 that is shared by all three cannot separate them. The measurement is reported for the record, per the standing rule that a contamination: declaration without a measurement is a placeholder. tools/dependence_check.py unmodified, C/A/B against CMP; no line of CMP was displayed to the translator.

9. Pre-flight cost estimate

Built from max_tokens, not from expected output (note (abc)), with a 2× routing margin on every stage (the S022 provider caution).

Declared worst case: $2.90, revised upward from $2.40 by the critic's amendments; the fourth rendering itself cost $0. UTC day 2026-08-06 opens with the full $5.00 and no session has spent today, so the revised worst case is 58% of the cap. A stage that does not fit remaining headroom is dropped, not shrunk silently.