Pillar D · State as of 2026-10-03

When does a model’s “no” not mean “stop looking there”?

As of October 3, 2026, no negative in Rosetta Quantum’s own ledger — not our detector’s zeros, not our pre-registered quantum-walk test — authorizes the instruction “stop looking there,” and that is the point: a model’s “no” only rules something out when the test was specified in advance and has measured error rates behind it. Our allosteric-pocket detector returned zero overlapping residues on 58 of its 86 benchmark proteins, and the zeros include ABL1, whose allosteric myristoyl pocket is the binding site of asciminib, a drug the FDA approved on October 29, 2021. Because the detector’s false-positive rate on proteins without a known site has never been measured, those zeros license nothing. Our one pre-registered negative (Wilcoxon p = 0.095, n = 90) licenses exactly one sentence — and it is not “there is nothing there.”

→ Leer en español

State as of: 2026-10-03

Status as of: October 3, 2026.

What does a model’s “no” actually claim?

On its own, almost nothing: it claims that this model, with this threshold, did not flag this spot — which is a statement about the model, not about the spot. The classic statistics note by Altman and Bland carries the whole lesson in its title, “Absence of evidence is not evidence of absence” (BMJ 1995;311:485), and adds that a non-significant result shows an absence of evidence of a difference, not evidence that no difference exists. Two properties upgrade a “no” into something that can carry weight: the question and the test were fixed before the result existed, and the error rates of the test — its power, its specificity — were measured and published. A negative with neither property is an anecdote with a minus sign.

Rosetta Quantum’s ledger happens to contain both kinds of negative, produced by the same house in the same quarter. This post calibrates them against each other, with our own data on both sides of the scale.

Why does a zero from our own detector rule nothing out?

Because for this detector a zero is the expected output even when a real site is there, and because its specificity has never been measured. In our no-annotation calibration (86 proteins, top-8 pockets per protein), the detector found the known allosteric site in 28 cases — a hit rate of 32.6% — and returned zero overlapping residues in the other 58 (our in-house calibration file, 2026-08-28; the scoreboard is published in our quantum-walk post and the bias behind it in our peripheral-bias report). A detector that misses 67.4% of the sites it was benchmarked on makes “no” its modal answer in the presence of a site. Reading that zero as “don’t look there” inverts its meaning.

The list of 58 zeros makes the cost concrete. It includes ABL1 — and ABL1’s allosteric myristoyl pocket is where asciminib binds, the drug the FDA granted approval on October 29, 2021 (FDA); Novartis describes it as “the first FDA-approved CML treatment that binds to the ABL myristoyl pocket” (Novartis). The same list of zeros includes KRAS and SHP2. If anyone had treated our detector’s zero on ABL1 as an instruction to stop looking, the instruction would have pointed away from a pocket that today anchors an approved medicine.

There is a second, quieter defect: to convert a zero into evidence of absence you need the ratio of two probabilities — a zero given no site, against a zero given a site. The second number is measured: 67.4%. The first has never been measured, because all 86 benchmark proteins carry a curated allosteric site; the set contains no negatives to score against. A likelihood ratio with an unmeasured numerator is not a number, and Rosetta will not pretend otherwise.

What turns a negative into a result with authority?

Freezing the question, the test, and the threshold before the result exists — and sealing the freeze. That is what our one pre-registered negative did. On August 3, 2026, before running anything, Rosetta Quantum sealed pre-registration PR-QMARGIN-001 (content hash 88d0fbf8): the hypothesis (quantum-walk propagators add non-redundant signal for the allosteric pocket beyond three classical propagators, under identical budget), the two variants (a logistic model on 6 propagators against the same model on the 3 classical ones), the protocol (leave-one-protein-out, spatial null with 2,000 permutations), the primary test (one-tailed paired Wilcoxon, α = 0.05), and — the clause that matters here — that a non-significant result would be a publishable, precise negative. The seal records committed_before_run = true.

The run came back negative and was published as sealed: Wilcoxon p = 0.0949, mean percentile delta +0.073, 54 of 89 proteins improved, and the classical-only variant actually had more individually significant proteins (13 vs 11) — seal RQ-EXP-QMARGIN-001. Re-measured on the stricter 86-protein no-annotation set, the paired test stayed non-significant — p = 0.0741 and 0.0921 — with 1 of 86 proteins individually significant (our quantum-walk post). This negative has authority precisely because nothing about it was chosen after seeing the result.

How much does even a pre-registered negative authorize?

Exactly the sentence that was frozen, and no more: under budget parity, on these 90 proteins, the quantum walk’s non-redundant contribution did not reach significance at α = 0.05. It does not license “quantum walks add nothing.” The observed trend was positive — delta +0.073, 54 of 89 proteins improved (seal RQ-EXP-QMARGIN-001) — and a p of 0.095 is a failure to demonstrate an effect, not a demonstration of its absence; that is Altman and Bland’s distinction (BMJ 1995;311:485) applied to our own result.

The statement “the effect, if any, is smaller than X” requires a different instrument: an equivalence test — two one-sided tests (TOST) against a smallest effect size of interest declared in advance (Lakens 2017, Social Psychological and Personality Science 8(4)). Our pre-registration declared no such bound, so even our best negative leaves the door formally open. If Rosetta wants the stronger sentence, it will have to pre-register the equivalence bound first and seal it before running; until that happens, no row in our ledger will claim it.

How should you read the negatives in our ledger, then?

By asking what each one licenses. The table below is the complete list — all of it our own data, with the licensed sentence spelled out per row.

Negative in our ledger	The number	What it licenses	Source
Detector zero on ABL1 — and 57 more proteins	overlap 0; detector hit rate 32.6%	nothing — specificity on site-free proteins never measured	FDA, Oct 29, 2021 · in-house calibration, n = 86
Blind retrospective arm	0/1, Wilson 95% 0–79%	nothing — n = 1	run-report W37
Prospective scoreboard, 4 weekly runs	3/5 or 2/5; Wilson 23–88% / 12–77%	nothing yet — every interval contains the 32.6% retrospective rate	run-report W40 · our prospective-validation post
QMARGIN, pre-registered	Wilcoxon p = 0.0949; delta +0.073	one sentence: no non-redundant quantum signal demonstrated at α = 0.05, n = 90	seals PR-QMARGIN-001 / RQ-EXP-QMARGIN-001 · quantum-walk post
Equivalence test (TOST)	none run yet	it would license “the effect is smaller than X” — no row in our ledger reaches this rung	Lakens 2017

Note the asymmetry, because it runs against instinct. The zero from the detector feels strongest — “the model looked and said no” — and licenses the least. The pre-registered p = 0.095 feels weakest — close calls get rounded to “nothing there” in conversation — and it is the only row that licenses any sentence at all. The authority of a negative is a property of the test, not of how emphatic the number looks. Our prospective-validation post published the short form of the limit — a model that is wrong two times in three cannot rule anything out — and the five failure modes a prospective design does not fix; this post adds the calibration between those two facts: between “licenses nothing” and “licenses one sentence” there is a ladder, and a reader of any vendor’s negative — or of ours — can place the claim on it using the checklist our credibility post applies to positives. The rungs are the same in both directions.

What we don’t know

We do not know the detector’s false-positive rate on proteins without a known allosteric site, because the benchmark contains none; until that is measured, its zeros license nothing and its hits inherit the peripheral bias we documented ourselves (peripheral-bias report). We do not know whether the quantum walk’s positive trend (delta +0.073, p = 0.095) is a small real effect or noise; the pre-registered test cannot tell, and we will not decide by re-running until something reaches significance. We do not know what the right smallest-effect-of-interest bound would be for an equivalence version of QMARGIN; if we set one, it will be pre-registered and sealed before the run, and published whichever way it lands. And we do not know how many of the 58 zeros sit on genuinely druggable pockets the way ABL1’s did — that is exactly the question a 32.6% detector cannot answer.