void · e 001c void

E-001c is void — ~~the fluent register has~~ **`gpt-oss:120b` has** a floor where the band has a ceiling

Title corrected 2026-08-31, retraction 18. The void stands and its numbers stand. What does not is the generalisation from this generator to the fluent register: on qwen3.5:35b, the same script, specification, instruction and band put cell A in band 11 of 12 against gpt-oss's 0 of 12, and there is no fluency-axis length effect at all — 23 of 24 fluent against 23 of 24 terse. See RESULTS-qwen-floor.

Status: void · closed · successor not yet registered Voided: 2026-08-04, during composition, before any probe outcome existed Gate: register filter exhausted in cell A (PREREGISTRATION §4, §6.1)

Two runs, and only the second one counts

E-001c voided twice, and the difference between them is the whole reason this document can be trusted.

2026-08-03 — cell A exhausted its budget in 38.9 minutes. That void establishes nothing. The runner held its own copy of the length instruction, diverged from the one §3 calibrates the band on, and 37 of 40 attempts overshot the band under it. DEFECT-001 records it.

2026-08-04 — the instruction repaired to the calibrated one, the same cell exhausted the same budget in 80.4 minutes. That is this void.

The distinction matters because the two runs record the identical void_reason string, and a reader who found only the first would have read a defect as a finding.

The numbers

Cell A (fluent × declarative), gpt-oss:120b at think=medium, temperature=0.7, band 182–231 words, both raters judging each message alone:

The repair moved the mean 46 words and moved nothing else. Band compliance went from 3/40 to 2/40; joint acceptance stayed at zero.

What the run actually found

The minimum over 40 attempts is 229 words. The band's ceiling is 231.

The fluent-declarative register, for this model, has a floor two words inside the band. Getting into the band requires the very bottom of what the register produces — and of the two messages that got there, neither was judged fluent prose by either rater.

So the two filters are not independent, and not merely non-independent but opposed: the messages short enough to be admitted are the ones that have stopped being connected prose. CALIBRATION-001 projected a joint rate of 7.7% by multiplying the two marginals. The measured rate is 0%.

That is the finding, and §8 named it in advance:

A register filter exhausts its budget — 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. That is a finding about instructions, and the successor is a different prompt rather than a different analysis.

It is sharper than §8 anticipated in one respect. The failure is not that the register is unreliable at a fixed length; it is that this register and this length are mutually exclusive. Fluent prose about this specification, from this model, does not come in under 229 words.

What was and was not seen

Voiding is a decision, and a decision made after seeing outcome data is not the same decision. Precisely:

Examined: composed message texts, their word counts, their register verdicts, the per-cell rejection counts. Not examined: any probe answer, any divergence, any fidelity, any Φ, any β, any effect. The sweep never began in either run. No messages.json was written, because composition never completed. No hypothesis was computed.

Cells B, C and D were never reached by either run: cell A is composed first and voids the run on exhaustion. They were measured off-run, after the void, by floor_by_register.py — instrument data that cannot revive E-001c and exists so a successor knows what it is designing around:

The floor belongs to the fluency axis. Fluent cells put 5 of 24 messages inside the band; terse cells put 23 of 24, Fisher exact p = 1.2e-07. Along the other axis — declarative 11 of 24 against contrastive 17 of 24 — p = 0.14. The fluency instruction moves the median length 24 words; the contrastiveness instruction moves it 7.

So the 2×2 did not fail as a design. Half of it did. The terse row composes inside the registered band without difficulty, and would have filled its cells. The fluent row cannot, and cell A cannot even in principle: its floor of 232 sits one word above the band's ceiling of 231, measured at 0 of 12 here and 0 of 40 in the run.

What survives for the successor

Not everything here failed, and a successor that rebuilt all of it would be discarding what the two voids paid for.

• The absolute register filter works. It rejected 24 of 40 messages and accepted 16, on two providers with no shared point of failure. The forced-choice rating it replaced scored 64/64 on feasibility messages of which three in eight were lists — an estimator that could not fail. This one can and did.
• The two-sided band works as a measurement even though it failed as a constraint. It is what made the floor visible.
• The instruction is now imported, not copied, so the calibration and the run cannot diverge again (DEFECT-001).
• The ceiling effect is characterised: this model writes to the ceiling it is told and slightly past it, and the effect binds the fluent register and not the terse one (CALIBRATION-001, Fisher p = 0.0345).

What the successor must change is the instruction set, as §8 says — and it now has a specific target rather than a general one. A design that pins length and varies register cannot use a band whose ceiling sits at the fluent register's floor. The measurement above says where the band would have to go: the fluent row's floor is 223 and its median 234, so a band centred near 240 admits both fluent cells, and the terse row — floor 184, median 210, and demonstrably able to hit a target — would be the arm asked to pad up rather than the fluent arm asked to cut down. Which direction the padding runs is not cosmetic: asking a terse register to reach a length is asking it to add content, and that reintroduces the confound from the other side. A successor has to say what it does about that before it collects.

Neither choice is available to E-001c. §7 forbids moving the band, the amendment cap is spent on AMENDMENT-001, and a band chosen after measuring which band works is the acceptance criterion picked to fit the data — the failure E-001b's VOID exists to record. It requires a new registration under a new id.

Amended 2026-08-04: §8's successor is the wrong successor

This document said above, following §8, that the successor is a different prompt. Measured afterwards, that is wrong, and the correction belongs here rather than in a document nobody reading this void would find.

headroom_check.py took three of the fluent messages this run rejected — they passed both blind register raters and failed only the band — and three terse ones, all 221–233 words, and ran the draws this experiment never reached. Observed agreement is 0.951 fluent against 0.941 terse: a difference of one third of one probe in 34, with 1 to 3 probes diverging per message where E-002c's outcome-variation gate wants at least 3.

MERIDIAN-34 is saturated at the length the fluent register requires. A successor with a better prompt and this measure would compose successfully and then measure nothing. See Problem 15.

The correction is recorded and §8 is not edited, because a pre-registration stands as written.

The programme cost so far

E-001 confounded style with length. E-001b tried to fix it with a shared token budget and voided on its parity gate. E-001c tried to fix it with a two-sided word band and voided on its register filter. Three designs, one confound, no result.

That is worth stating plainly rather than burying: the question does fluency buy understanding has now consumed three experiments without being answered once, and each void was caused by the same underlying fact from a different angle — this model's registers are not separable from their lengths. A fourth attempt that does not treat that as its central obstacle will void too.

This document is licensed CC BY 4.0.

All journal entries · Reference · Noophorics · Repository