instrument · e 001c calibration

CALIBRATION-001 — the model writes to the ceiling it is told, and slightly past it

Measured 2026-08-04, after the run of 2026-08-03 voided and DEFECT-001 established that the void could not mean what it said.

This is instrument data. It is not E-001c's data and no hypothesis is touched by it, on the same footing as the feasibility messages PREREGISTRATION discards before collection.

What was asked

DEFECT-001 repaired a divergence between the instruction the band was calibrated on and the instruction the runner sent. The obvious next question is whether the repair is sufficient — whether the calibrated instruction, carrying the registered band numbers, fills a fluent cell.

It does not, and the reason is more interesting than the defect was.

What was measured

gpt-oss:120b, think=medium, temperature=0.7, one composition per draw, realised word counts. Cell A is fluent × declarative; cell C is terse × declarative.

Row 1 pools the four cell-A messages in feasibility.json with four composed on 2026-08-04. Row 4 is the void run's own rejection record.

Fisher exact, rows 1 and 2 — p = 0.00071. Rows 2 and 3 — p = 0.0345.

Revision. The first version of this file reported row 2 as n = 8, 1/8. A background job's output was buffered and read as though it had died after three draws; it had not, and its remaining draws are now included. The direction and the conclusion are unchanged and the sample is larger. The number is corrected here rather than left standing, and the correction is recorded rather than made silently.

What it says

The stated ceiling is not a limit the model respects — it is a target it approaches from below and overshoots. Told 223, it wrote 197–230. Told 231, it wrote 208–326. The eight-word difference in the instruction moved the mean by twenty-six words, in the same direction and past the new number.

That is why the registered band cannot be hit by stating it. 182–231 is the acceptance band. The runner also uses it as the instruction, and an instruction of 231 produces cell-A messages around 244.

And it binds one register and not the other. At the same stated ceiling of 231, the terse cell complied 5 times in 6 while the fluent cell complied 2 in 11 (p = 0.0345). The terse instruction produces messages that sit below the ceiling, so moving the ceiling does not move them; the fluent instruction produces messages that sit at it, so it does. The pre-registration declared this in advance as a limitation and did not draw the consequence:

The fluent register imposes a length of its own. Feasibility Result 4: fluent messages land at 226 words with sd = 1.5 against sd = 14.5 for terse. The band admits both, but the fluent arm is not complying with the target so much as coinciding with it. If a future calibration moves the target, that coincidence may not survive.

The target moved by eight words between the calibration and the registration, and the coincidence did not survive. §6.2 called it.

What this predicts for the run of 2026-08-04

Cell A at 18% band compliance and 42.5% register acceptance — the latter measured live, 17 of 40, on the void run — gives a joint acceptance of about 7.7%. Eight accepted messages would need roughly 104 attempts against a budget of 40.

So the run in flight is expected to exhaust cell A. This projection is recorded before its outcome is known, so that it can be wrong where anyone can see it.

It was right about the outcome and wrong about the reason

The run exhausted cell A at 09:48:32Z on 2026-08-04 — 40 attempts, 0 of 8 accepted, 80.4 minutes. The projection above called that. Two of its three components were wrong anyway, and the way they were wrong is the finding:

Band compliance was overestimated more than threefold: 2/11 from a small sample against 2/40 measured. And the joint rate is not the product, because the two filters are not independent — of the two messages that landed inside the band, neither was judged fluent prose.

The realised range makes the mechanism plain. Cell A over 40 attempts: 229–318 words, mean 242.1, minimum 229. The band's ceiling is 231. The fluent-declarative register has a floor two words inside the band, so the only way into the band is the very bottom of what the register produces — and at that length it stops reading as connected prose to both raters.

That is not a length failure and not a register failure. It is the statement that, for this model and this instruction set, fluent prose and this length are not simultaneously available. Which is a claim about the confound E-001c exists to break, arrived at from the wrong end.

Where that leaves §3

PREREGISTRATION §3 fixed the band this way:

Calibrated on the feasibility messages, whose realised range was 184–230 across both registers; a single target at ±10% admits 7 of 8 fluent and 6 of 8 terse, and ±12% admits all sixteen.

Every one of those sixteen messages was composed under an instruction that said 223. The band was then widened to 182–231 so that all sixteen fell inside it — a sound way to choose an acceptance criterion, and it would have been fine if the criterion stayed a criterion. But the same numbers are then read back to the model as the instruction, and the model moves when they move. The calibration measured compliance with one ceiling and the run demands compliance with another.

Nothing in the document is false. §3 says what it did. What it does not say is that the band it derives will also be spoken aloud, and the two roles are only safe to conflate for a model that treats a stated range as a boundary. This one does not.

What is NOT being done about it

The band is not being changed. §7 forbids touching it, the amendment cap is spent on AMENDMENT-001, and a stated ceiling of 223 chosen after measuring that 223 works is the acceptance criterion being picked to fit the data — the precise failure E-001b's VOID records.

So the run of 2026-08-04 states 182–231, as registered, and is expected to exhaust cell A's forty attempts. If it does, that is the §8 outcome — "the model cannot reliably produce one of the four registers at a fixed length, and the 2×2 is not buildable with this instruction set" — and §8 already names the successor: a different prompt, under a new registration, not a different analysis.

This file exists so that when that void arrives it is read correctly: the instruction set fails, and the specific way it fails is that this model negotiates with a number rather than obeying it.

Superseded on one point, by a better measurement

The limitation below said the contrastive cells were unmeasured and the effect was therefore claimed on the fluency axis without the axis having been tested. floor_by_register.py then measured all four at n = 12 and the axis claim holds: fluent 5 of 24 inside the band against terse 23 of 24 (p = 1.2e-07), declarative 11 of 24 against contrastive 17 of 24 (p = 0.14). The table is in VOID.md. This file's numbers stand as the smaller, earlier sample that pointed at it.

Limitations

• Two of four cells. A (fluent × declarative) and C (terse × declarative). The contrastive cells B and D are not measured; the effect is claimed on the fluency axis, where it was found, and not on the contrastiveness one. Withdrawn 2026-08-04: all four cells are now measured at n = 12, and the fluency axis carries the effect while the contrastiveness axis does not.
• n = 6 to 11 per arm, unrandomised and unblinded, composed in blocks rather than interleaved. The arms ran on the same local model on consecutive days.
• The cell C arm is n = 6. 5/6 is not a precise compliance rate and the interval around it is wide; it establishes that the terse register behaves differently from the fluent one, not by how much.
• The second arm's later draws competed with the live run for the GPU. That changes latency, not text.
• One model. Nothing here generalises past gpt-oss:120b at think=medium.
• A twelve-draw arm was cut short at five once the live run began, because the run's own composition stage measures the same quantity against the real raters and the two were contending for one GPU.

This document is licensed CC BY 4.0.

All journal entries · Reference · Noophorics · Repository