instrument · e 001c feasibility

Can the four registers survive being pinned to a common length?

Instrument check, before registration. No hypothesis is tested here and none may be read off it. It exists to answer whether E-001c is buildable at all, after E-001b was voided by its cost parity gate.

Answer: yes, with a calibration problem that is measured rather than assumed.

What was replaced

E-001b instructed Stay within 350 tokens — a ceiling. A ceiling bounds above and says nothing about below, so it cannot produce parity even when obeyed, and it was not obeyed: every cell exceeded it, the fluent cells by 76%.

E-001c states a two-sided word band instead, and verifies realised token cost afterwards. Words rather than tokens because a model cannot observe its own tokenizer, and E-001b is direct evidence that it does not honour a token limit. Nothing else in the prompts changed: the same five blocks, the same two varying blocks, only the output block swapped.

Result 1 — cost parity is achieved

Target 220 words (band 198–242), k=3 per cell, gpt-oss:120b, think=medium, temperature 0.7:

The 1.5× gap collapses to 1.10, and the ordering reverses — terse now costs slightly more per message than fluent, which is what one expects once length is the controlled variable and register is free to vary within it.

Parity across cell means: 1.160, inside the 1.30 threshold. Parity across individual messages: 1.320, marginally outside it, driven entirely by within-cell spread (a terse message at 474 tokens against a fluent one at 359). The runner requires both readings to pass, so this configuration would still be voided — see the calibration below, which is the reason it is close rather than comfortable.

Result 2 — the manipulation survives

This is the half that mattered more. Convergence bought by destroying the distinction would be a null experiment with extra steps.

The two registers remain clearly separated at a common length: fluent writes sentences nearly twice as long and uses subordinating connectives three and a half times as often. Neither register produced bullet lists, so "terse" is manifesting as short declarative sentences rather than as note form.

These are proxies, not a validity check. They distinguish prose from notes; they do not establish that the intended style distinction is what survived. Establishing that needs independent blind rating, and it is a precondition of E-001c's registration, not something this script can supply. What the numbers do rule out is the loud failure mode — pinning the length did not collapse the registers into one.

Result 1b — the calibration, confirmed rather than assumed

Asking for 203 words, as the offset predicted:

Both readings now pass, and the style separation survives the shift (fluent 19.75 vs terse 11.52 mean sentence words; 1.77 vs 0.70 connectives per 100).

The overshoot is cell-dependent, which the single-target correction cannot remove: A +14.4%, B +8.3%, C −2.8%, D +3.0%. Fluent registers run long, terse ones land. Word counts therefore do not equalise across cells — but the requirement is cost parity, and cost parity is what passes.

Result 1c — max/min over messages is the wrong statistic, and this is why

1.290 against a threshold of 1.30 is not a pass, it is a warning. The maximum and minimum of a sample both drift outward as the sample grows, so max/min is not scale-free: the same design fails the same gate simply for collecting more messages.

Parametric bootstrap from the observed per-cell cost distributions (normal, 20 000 draws):

At E-001b's own pre-registered k = 8, a design with genuinely parous costs fails the across-messages gate 86% of the time. A gate that gets harder as you collect more data penalises statistical power, and would have voided E-001c for the sin of being adequately sampled. The cell-means reading is flat across every k.

So the ambiguity flagged in the void — the pre-registration said "across cells", the runner compared messages — resolves in favour of the pre-registration, on a statistical argument rather than a convenient one. The argument is that the sample max/min is not scale-invariant, which is true regardless of which side it favours; but the resolution does happen to favour the reading that passes, and that is stated here rather than left for a reader to notice.

The runner is corrected to use cell means, and E-001c will register that reading explicitly. A stricter successor worth considering is an equivalence test (TOST) on cost across each axis: "the difference is inside a pre-specified negligible band" is the claim actually wanted, and a ratio threshold is a crude stand-in for it.

Result 3 — the model overshoots, systematically and in one direction

Six of twelve messages fell outside the ±10% band. Every miss was long. None was short.

target 220 words -> realised mean 238.2 (+8.3%)

A one-sided error is much better news than a symmetric one: it is a calibration offset, not noise. Asking for 203 words should centre the realised distribution on 220. That prediction is being measured at k=4 before it is relied on, because predicting a calibration and assuming it are different things, and this experiment has already spent a day on the difference.

Compliance at 50% also sets the price of rejection sampling on words: roughly two compositions per accepted message. Affordable — and worth contrasting with rejection sampling on cost, which the voided run showed has acceptance zero for every band, because the two families' realised costs were disjoint.

What E-001c needs before it can be registered

1. The calibration confirmed, not predicted. In progress.
2. Independent blind rating that the four cells still instantiate the intended registers at a pinned length. The structural markers are not sufficient and are not offered as sufficient.
3. A decision on the parity reading. E-001b's pre-registration said "across cells" and its runner compared messages. E-001c will state one, require both, and evaluate them at composition — where their inputs already exist — rather than after the sweep.
4. M33 repaired or removed. MERIDIAN-34 is superseded for new experiments (SENSITIVITY-M33); a probe whose key the source does not determine should not be inherited.

Registration follows those. Not before.

This document is licensed CC BY 4.0.

All journal entries · Noophorics · Repository