Repository path: wiki/findings/results/RS-20260806-same-man.md · rendered 2026-09-09

You can rewrite seven tenths of a man's words and readers will tell you it is the same man

ARM-voice-persona step 2 (T2), and the arm closes on it. E-20260806-carrier-or-knowledge. 65 billed bodies, $0.614681164, key-usage cross-check closing to 2e-9; verifier 194 checks, 0 failures, 3 mutation tests, 3 caught. Pre-run critic NEEDS-AMENDMENT, six findings, four BLOCKING, all six accepted before any other call was dispatched — and one of them added a fourth rendering.

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

The headline: asked point-blank whether two renderings of the same paragraph are narrated by the same person, four source-blind seats answered "certainly the same person" at 27 of 32 cells — 35 of 40 with the repaired control included — and the three cells that said otherwise were about a framing artefact and a register.

1. What was built

One paragraph of Gogol — «Записки сумасшедшего», Октября 3, 306 Russian words of a titular councillor denouncing other men's graft on his way to beg for an advance — rendered by the lead four times under R19, a regime minted for this and specified on properties rather than on persons:

K1–K5 are the five properties wiki/goodness-senses.md §voice names as the observable carriers of a narratorial perspective — register, rhythm, diction temperature, idiosyncrasy, distance. K6, self-knowledge, is the property RS-20260805d §4 found moving most and the entry does not name.

Four panel seats (P1, P2, P3, P5), none of which held any other role in the run, saw two passages at a time — no source, no authorship, no mention of translation — and rated are these narrated by the same person from 0 to 6.

2. The primary is WITHHELD, and it would not have fired anyway

mean rating, 8 cells per arm (4 seats x 2 orders), 0 = certainly the same person

  A'   0.000        B    0.125        A    0.750        N (frozen)    1.625
                                                        N (repaired)  0.000

P1p  d = r(B) - r(A) :  T = -5,  mean -0.625,  1 positive / 1 negative / 6 zero,  exact P = 1.000
P1q  d = r(B) - r(A'):  T = +1,  mean +0.125,  1 positive / 0 negative / 7 zero,  exact P = 1.000

F3 and F7 both fired — the floor arm N scored at or above every manipulation — so the primary is withheld, as registered. P1p does not fire, P1q does not fire, and P2p, the entry's own claim, does not fire either. The registered outcome for that conjunction is the two manipulations are not distinguishable at this power, and that is the outcome.

F1, F2, F4 and F5 did not fire. The manipulation isolated (§4), content parity held (§5), only one of four seats was degenerate against a bar of two, and the mean within-pair order difference was 1.125 against a bar of 2.0.

The gates fired on a defect in the control arm, and the repair is reported beside them

N's two "different person" verdicts are both about the paraphrase announcing itself. The seat that wrote N wrapped it in an assistant's preamble and closing note, and the runner's strip removed only the terminator. P1, on the frozen N: "Passage 2 explicitly presents itself as a rephrased version addressed to a reader." P2: "Passage 2 is framed and rewritten in modern language by an AI assistant." Neither is a judgment about a narrator.

A declared post-hoc probe re-ran the same four seats × two orders on N with the framing stripped: 0.000 at 8 of 8 cells. So the floor is flat and both gates fired on a materials defect.

The primary stays withheld on the frozen materials. Swapping to the cell that would have passed after watching it pass is the move the verification discipline exists to stop (config/models.md, S086), and it is not made here. What the repair licenses is one sentence, and it is the honest one: the gates fired on an artefact, and the primary they withheld would have failed on its own numbers regardless — P1p at mean −0.625 and P1q at +0.125, against bars of ±1.00 and six of eight cells.

3. The finding: propositional parity decides the question the note asked

The seats say why, and they agree. P1, comparing C with B:

"The passages share the same events, grievances, social setting, and distinctive voice, with Passage 2 mainly adding self-interruptions and more fragmented emphasis. These read as variations of the same narrator rather than different people."

Readers answer same person? from the propositions. And R18 P5 and R19 Q5 — the parity rules that make a paired rendering a controlled contrast at all — hold the propositions fixed by construction. So:

The design voice's reachability note has been asking for since S097 is self-defeating, and this is the second run to build it and the first to say so. The note wants "the same source persona realised two ways on purpose" with "a blind jury asked which reader met which person." Making the pair a controlled contrast requires propositional parity; propositional parity is what the readers use to answer; so the control determines the answer. The note's design cannot be fixed by better renderings, more seats, or a different language. It is fixed by a different question.

The one arm that did move readers off zero, before its defect was found, was the arm aiming at nothing. Note (bix) fires a third time, in a third session, on a third design — and this time with its mechanism visible: N did not out-persuade the manipulations, it out-flagged them, on a property of machine text that has nothing to do with narrators.

What this does not show

Not that the five carriers are inert. Nothing separated in this run, so nothing is licensed about which property matters more; P1p failing is not P2p holding. And per amendment A6, registered before the data, the licensed statement is about operations a translator can perform, not about latent properties:

A translator who rewrote 70% of a passage along all five properties the entry names, moving its measured register from 2.33 to 6.00 and its measured distance from 1.67 to 4.00, did not thereby make source-blind readers say they had met a different man.

4. The manipulation worked, which is what makes §2 a measurement

Three seats rated each text alone on the six carriers, 1–7, with no comparison and no source. F1's pre-registered isolation margins, all required ≥ 0.50:

A   mean|K1..K5| 2.267  |K6| 0.667   margin 1.600
A'  mean|K1..K5| 0.600  |K6| 0.000   margin 0.600
B   mean|K1..K5| 0.067  |K6| 0.667   margin 0.600

Movement from C, per carrier:

A is not a failed manipulation. It moved four of the five carriers by 2.3 to 3.7 scale points, and readers still met the same man.

idiosyncrasy saturates for the second time, on new material

K4 across all five texts: 5.67, 5.33, 5.67, 5.67, 5.50 — a total range of 0.33. RS-20260805d §4 found the same scale pinned at 7 in all eight profiles of a Dutch pair. Two runs, two languages, two authors, two translators, two disjoint seat sets, and the scale has never discriminated anything. It is not measuring; it is agreeing that first-person comic prose is mannered.

The ten-site budget bought a third of a register and none of an idiosyncrasy

A′ was built because the critic was right that A and B were not parallel operations. What it establishes, and T-zapiski-oct3-R19-v4's log predicted from the translator's side before any reader was asked: ten edits buy K1 +1.333 of A's +3.667, K2 +0.667 of +2.333, K3 +0.333 of +2.667, and K4 exactly nothing. K5 distance is the exception — 0.667 of 2.333 from five edits, because in this paragraph distance is carried by a countable thing, second-person address.

So two of the five named carriers are distributed properties that no small edit moves. Register is a property of a distribution of words and rhythm a property of every sentence boundary; the paragraph has 433 of the first and far more than ten of the second. The entry lists five carriers as though they were five comparable handles, and they are not.

5. The parity screen convicted the lead's own log, and that is what it was rebuilt to do

The screen was blind to B by construction in the frozen design — it was instructed to ignore first-person statements about the narrator's own material, which is exactly what B's manipulation consists of. The pre-run critic caught it (finding 2, BLOCKING) and it was rebuilt to count everything under two neutral labels with no exclusion named to the seat.

EVENT differences (the F2 gate, bar > 2)     A 0    A' 0    B 1
SELF  differences (reported, not gated)      A 1    A' 0    B 3

Both screen seats independently found exactly the three SELF additions B's log declares — the prefer to call it, the bring it up whenever I want him smaller, the I tell myself. The log's count was right.

And both seats convicted the log on a fourth site. B's log claims "And I am on my way to the treasurer" is a relocation of a proposition C already asserts, adding nothing. Both seats labelled it EVENT — a new assertion. They are right and the log is wrong: C says he hoped to catch the treasurer; B says he is on his way. A hope is not an errand in progress. F2 does not fire — one EVENT against a bar of two — and the log is corrected here rather than quietly.

The screen also caught something in A the lead had argued away. A's log claims "it is to be presumed that he is moved by envy" is a formality and not a hedge, so K6 stays held. The screen read it as a SELF difference: "Passage 1 asserts as fact that the section head is jealous, while Passage 2 asserts only that it is 'to be presumed'." One seat of two, and A's measured K6 movement is +0.667 — the same as B's. A is not perfectly clean on its HOLD, and the entry should not be revised as though it were.

6. What the seats talked about, and what they never mentioned

Q2/Q3 free text, coded by string match against keyword sets fixed in the design before dispatch:

Not one seat in 32 cells used a word for self-awareness, irony, candour or detachment — including on the arm built entirely out of them. The manipulation-check seats scored B's K6 above C's when asked about it directly; the identity seats, asked an open question, never reached for the vocabulary at all.

7. The lead's registered expectation, and how it did

Registered before dispatch: r(A) = 3.5, r(B) = 3.0, r(N) = 2.0, mean d = −0.5. Measured: 0.750, 0.125, 0.000 (repaired), −0.625.

The sign of the contrast was predicted and every level was wrong by a factor of four or more. The lead expected an instrument that would separate texts and argued about which way; the instrument does not separate texts. That is the miss worth recording.

8. Limits, in the order they would bite

• The primary is withheld, and its value is printed anyway because a withheld primary whose value is hidden is worse than one whose value is shown.
• The floor arm was defective and both firing gates trace to it. The repair is a declared post-hoc probe, not a re-run of the primary, and it is labelled that way everywhere.
• A ceiling, not a floor, is what this run is up against. Twenty-seven of thirty-two cells returned the scale's minimum. An instrument that returns 0 for a 70% rewrite has very little room to show anything, and no magnitude here should be read as a size.
• One of four primary seats (P3 x-ai/grok-4.5) returned 0 in all eight of its cells. F4 needs two of four and does not fire, but a quarter of the primary is a constant.
• One of three manipulation-check seats is degenerate on the decisive scale. mistralai/mistral-medium-3-5 returned K6 = 6 on all four texts it rated, where the other two returned 1–3, and its boilerplate reason was near-identical each time. Including it raises every K6 mean and compresses B's isolation margin to exactly 0.600. F1's margins were computed over all three seats as registered, and would be larger without it.
• B differs from C by three self-descriptive assertions as well as by their marking effect, and this run cannot separate the two (declared in the design before dispatch).
• No source-side yardstick. This run measures the target-side half of voice only; nothing here says the lead's Poprishchin is Gogol's.
• One passage, one author, one pair, one translator, eight paired cells. A pilot-level signal.
• The C–A parity cell needed max_tokens raised from 6,000 to 16,000 after two seats and a reserve returned finish_reason: length on it — declared amendment A7, note (bhf)'s ninth firing, and the failing cell was the 70%-rewrite pair, where enumeration is longest.

9. Contamination

Measured after all four renderings and all four logs were frozen, tools/dependence_check.py unmodified, against Field 1916 (public domain, extracted mechanically by heading index, never displayed).

Zero shared 7-grams on all four arms is the cleanest contamination measurement in the project's record, against a longest run of 24 tokens (S079) and 14 (RS-20260805d's PA). The declarations stay suspected rather than moving to none, and the reason is the standing one: one reachable comparator is not a measurement of the training data. Diary of a Madman has several modern English translations, none of them free, and none of them measurable here.

Reported for the record: this design's validity does not turn on independence from any published rendering, since all four members are the lead's and an overlap shared by all of them cannot separate them.

10. Cost

$0.614681164 over 65 billed bodies against a declared worst case of $2.90 — 21%. Key-usage delta 0.614681162 against a per-request sum of 0.614681164, closing to 2e-9.

The pre-run critic cost $0.0477954 — 7.8% of the run — and it is again the most valuable line. Four BLOCKING findings, all accepted: it rebuilt the screen that would have passed B without looking at it, doubled the manipulation check, raised the primary from three seats to four, and forced the fourth rendering whose log supplies §4's sharpest paragraph.

Waste: four bodies on the C–A parity cell (two qwen3.7-max and two reserve kimi-k3, one of them a non-JSON body), all finish_reason: length or unparseable, before amendment A7 raised the cap and the fifth attempt returned in 61 characters.