observation · good instrument sitting unused

Three laws were blocked by the wrong instrument

2026-08-28. Nine days that began by trying to test L1 and ended by finding that the reason three laws have stood unattacked since founding was never the ideas, and never — as I published mid-way — the programme's instruments in general. It was the one instrument every experiment happened to reach for.

E-006 died at zero model calls, on a gate it set for itself

The ablation ladder for L1 was pre-registered with a power calculation committed before collection and a rule that below 0.80 it would not run. It returned 0.117. The null returned 0.023 against a nominal 0.025, so the test was calibrated — not broken, blind.

Holding the design fixed, the probe count needed for 0.80 power:

Void before collection. 2 880 calls not spent, and a fact bought instead: L1's concavity limb needs about 512 probes and the largest measure here has 34.

My first power calculation was wrong and is recorded — it bootstrapped the eight rungs where the pre-registration declares the thirty-two probes. Corrected, the verdict held. But a power calculation that does not implement its own declared decision rule is not a power calculation; it failed safe by luck.

So I asked the same question of the rest of the roadmap, and got it wrong

Probe-bootstrapping E-002c's raw draws gave the noise the programme actually has, and the answer looked like a ceiling: E-003's expected asymmetry of 0.219 sits under the 0.338 that 32 probes resolve; E-007's L4c splits the probes into classes and resolves 0.479. I wrote "the instrument is the binding constraint on this programme, not the ideas" and published it.

That file computed its MDEs from raw probe counts, an hour after retraction 15 established that MERIDIAN-IX32's 32 probes are about nine independent templates. Correcting my own arithmetic found the thing worth finding:

Zero near-duplicate pairs against eighteen to twenty-five, and not a length artifact — Jaccard over word sets is nearly length-blind and separates them the same way. Every published MERIDIAN result rests on about nine effective probes.

RIVERSIDE-30 is also the only measure here whose keys were ruled on by readers who did not write them. The best-adjudicated and least-templated measure in the repository is the same file, and it had been used in exactly one experiment — E-004, which is void.

It has headroom, and the divergence belongs to the probes

Nobody had ever run it against a brief. Predicted ≥ 3 diverged per message beforehand; measured 10.00 of 30 against a gate of 3, at ~400 words rather than the 230 operating point, which is the conservative direction.

Then the same three briefs to a second reader, with the eight always-diverging probes named in the prediction file first. Predicted Jaccard ≥ 0.4; measured 0.654.

On MERIDIAN the stronger reader simply lost less and its losses nested inside the weaker one's — an ability difference. On RIVERSIDE both lose comparably and lose the same things, which is what it looks like when the probes rather than the readers carry the signal. So pairing works here (×2.46 and ×1.80, reader-dependent) where on MERIDIAN it bought exactly nothing.

The ceiling was MERIDIAN's, not the programme's, and the sentence I published is struck where it stands.

Two smaller things, one of them satisfying

Retraction 17: E-002c computed its noise floor between the sender and PRIOR and divided sender-versus-receiver fidelities by it. The sender is a point mass on all 33 probes and PRIOR is not, so the mismatched floor was an order of magnitude too large and inflated the whole ladder — and because the matched floor falls as cost rises, it steepened the column as well. β, the registered quantity, never reads that column and is untouched. The estimator now refuses the mistake: fidelity_from_draws takes draws rather than floats, so the mismatch cannot be written.

And T30. gpt-oss answered it against its adjudicated key twice, seventeen days apart, and I left it standing unexplained rather than dropping it. qwen's sender answers with the key, agreeing with all three blind adjudicators. It was one model being reliably wrong about one probe — the adjudication standard was vindicated by its own anomaly.

Where it leaves the roadmap

E-003 is still blocked, for a reason I had to correct myself on twice. Not power: RIVERSIDE-30 resolves 0.135–0.183 against an expected 0.219. L2's sharper form needs a model pair whose domain prior and general capability disagree, and qwen is stronger or level on both domains available. Measuring only that an asymmetry exists is retraction 5, withdrawn as tautological.

Blocked on model selection. Which is a better position than blocked on the instrument, because a third model is a download and a probe measure is a month.

This document is licensed CC BY 4.0.

All journal entries · Reference · Noophorics · Repository