What does a dated scientific hypothesis look like, and how do you know later whether it was right?

TL;DR (as of 13 October 2026). A dated scientific hypothesis is one written down before the answer existed, in enough detail that a stranger can score it later without asking the author what it meant. Rosetta Quantum sealed two of them today: WEE1 and PPP2R1A, two candidate allosteric sites produced by its frozen detector, under seal RQ-PRED-ALLO-001 (sha256 a430ff0842da7392…, committed 2026-10-13T12:52:00Z, before this post was written). The honest headline is the scoreboard, not the hypotheses: of the 15 forward commitments in the queue, zero pass all four of the house filters with verified overlap, and 9 lost their significance the day the null was fixed (discovery queue, 2026-08-31 and 2026-09-02). The two sealed today are the survivors, not the winners.

What makes a hypothesis dated rather than merely stated?

Five things, and the calendar date is the least of them. A hypothesis is dated — in the sense that matters — when (1) the prediction is written in coordinates a stranger can use, (2) it is frozen before the answer exists, (3) the test that would refute it is declared in advance, (4) a deadline exists after which silence counts as a result, and (5) the author commits to publishing the outcome whichever way it falls.

The distinction underneath all five is old and well documented. Nosek, Ebersole, DeHaven and Mellor put it plainly in The preregistration revolution: progress depends on "generating hypotheses with existing observations and testing hypotheses with new observations", and the two get confused in practice because "ordinary biases in human reasoning, such as hindsight bias, make it hard to avoid this mistake" (PNAS 115(11):2600-2606, 13 March 2018, doi:10.1073/pnas.1708274114, via PubMed). The publishing machinery built on top of that idea is Registered Reports, in which "study proposals are peer reviewed and pre-accepted before research is undertaken", a design whose authors are careful to say is "not a universal solution for irreproducibility" (Chambers & Tzavella, Nature Human Behaviour 6(1):29-42, 15 November 2021, doi:10.1038/s41562-021-01193-7, via PubMed).

Rosetta has published the framework for this before — why a frozen model with a hash is the only kind of model whose hit rate cannot be flattered after the fact, and why the zero and the failed run get published too (what it means to validate a model with data that did not exist). What follows is the part that framework still owed: the individual commitments, named, with their numbers and their scoring rule.

What exactly did Rosetta Quantum commit to today?

Two candidate allosteric sites, both produced by the detector frozen under hash 08ad337d, both from the annotation-blind predictor (10 features, no UniProt annotation features), and both reported at the level the detector actually produces: a top residue plus its seven nearest distal neighbours.

Two sentences that belong next to that table and not three paragraphs below it. These are hypotheses, not findings: no structure, no mutagenesis and no functional assay supports either one, and Rosetta runs no wet lab. And frac_above is not a p-value — it is the fraction of random contiguous distal pockets scoring at or above the predicted one, taken at the maximum, which makes it selection-biased by construction and indicative at best.

How will anyone know whether these were right?

By a rule written down before the answer exists, which is the only version of the question worth asking. The seal fixes it:

A commitment hits if a future public report — an experimental structure, a mutagenesis study or a validated functional study — places a ligand-binding or regulatory site on that protein whose residues come within 8 A of at least one residue of the committed pocket, and that site is distal, meaning more than 6 A from the annotated functional site used as the anchor. It misses if such a report exists and the minimum distance exceeds 8 A. It is unscorable if no such report exists by 31 December 2029, or if the report does not publish residue-level coordinates that can be mapped onto the committed apo structure.

The third outcome is the one that makes this honest rather than decorative. An unscorable commitment will be published as unscorable and never as a win; a miss will be published as a miss. Rosetta will score nothing by simulation, and will re-measure the detector metrics in the table above approximately every six months at this same address, printing them whichever way they move. The seal class and the multi-outcome scoring are the same ones Rosetta Quantum used for its Q-Day position, which carries a floor, a central window, an early-miss band and a late bound (the Q-Day bet).

Why does the scoreboard read 0 of 15 when two commitments survive?

Because surviving two filters is not passing four, and Rosetta publishes the gap rather than the headline. The house criteria for a forward commitment are: historic frac_above below 0.05; overlap clean or at least evaluable; the predicted pocket still ranked in the top 3 when the anchor is geometrically dilated to R = 6 and R = 8 A; and distance-matched frac_above below 0.05. Six candidates clear the two newer filters. None clears all four with verified overlap — WEE1 only has partial adjacency, and PPP2R1A has nothing annotated in its resolved region to measure against at all, which means its most important check cannot be run today in either direction.

The reason the fourth filter exists is the most useful negative this project has produced about itself. The detector fires systematically at the periphery of the active site: its candidate pockets sit at the 11% distance percentile within their own protein (interquartile 7-15%), while the real allosteric sites of the validated set sit at the 28% percentile (interquartile 18-56%), Mann-Whitney p < 1e-5. Once the null pockets are drawn only from residues at a matched distance from the active site, 9 of the 15 commitments stop being significant, PTEN among them — and PTEN had been, for three weeks, the candidate this project called its most credible (the periphery bias, measured in house).

The same correction reached the project's primary metric. Under the historic null the annotation-blind ablation gave a combined Fisher p of 5.4e-6 across 86 proteins; under the distance-matched null it gives 0.0138 and 0.0107 on two seeds — it survives, but it falls about three orders of magnitude, and it is fragile: removing PAK1 alone takes it to 0.057, and at the level of individual proteins there are 5 significant of 86 against 4.3 expected by chance, binomial p = 0.43. Both commitments sealed today inherit that context, which is why they are sealed as hypotheses with a deadline and not announced as discoveries.

What does the field do when it has to score predictions in advance?

It appoints someone else to keep the clock, which is exactly the thing a self-sealed commitment lacks. The nearest live example is dated this month: the first OpenBind blind challenge, run with OpenADMET and the ASAP Discovery Consortium, asks participants to "predict protein-ligand complexes of 400 compounds bound to Zika virus NS2B-NS3 protease", launches 14 October 2026, takes interim submissions on 11 November and closes finals "16 December 2026 at 23:59 UTC". Its scoring is stricter than the usual benchmark convention: the organisers describe "tightening this to 1.5 A, along with additional checks on interaction accuracy (lDDT-PLI > 0.8) and physical validity", with physical validity checked by PoseBusters (Oxford Protein Informatics Group, 7 October 2026).

That is the better instrument, and saying so costs nothing. A challenge supplies a fixed deadline, a single answer key and a scorer with no stake in the result. A sealed commitment supplies only the first two; the third is borrowed from whoever happens to publish the structure. Rosetta states the asymmetry rather than papering over it, in the same spirit in which it insists that a margin without a unit is not a result (by how much does the winner win).

Is anyone else already claiming an allosteric site on WEE1?

Yes, twice in the public record, and neither claim can confirm or refute ALLO-001-A today — because neither publishes coordinates in a comparable frame.

This is not a criticism of either document; a patent is not written to publish geometry, and a docking poster is not a structure. It is a statement about what the public record currently makes checkable, and it cuts at Rosetta too: our own commitment is only scorable because the residues, the chain and the apo structure are printed above. Had we published "an allosteric site on WEE1" and stopped there, nobody — including us in 2029 — could have told whether we were right. The protocol for reading a claim like these, five checks and no mathematics, is published separately (how to read a quantum arXiv paper without being a physicist); the calibration of what a negative result does and does not license is published too, and it applies to our own negatives first (when does a model's no not mean stop looking there).

What we don't know

• Whether either site is real. Nothing here is validated. The scoring rule exists precisely because the answer is not available in 2026.
• Whether these are cavities at all. No fpocket or FTMap run has been done. The detector returns a top residue plus its nearest distal neighbours, which in most candidates is a contiguous stretch of sequence — a segment or loop, possibly not a three-dimensional pocket. The word "pocket" above means the detector's output, nothing more.
• Whether PPP2R1A can ever be scored by our own overlap test. Today it cannot: its resolved region carries no annotated functional residue to measure against. The commitment is sealed anyway, with that written into the seal, because a hypothesis you only register when it is convenient to score is not a hypothesis.
• Whether the periphery bias explains both candidates. It is measured, it applies to the whole queue, and no correction has been shipped. Fixing it is the project's stated first priority, and until it ships the detector and the discovery queue are partly measuring the same thing: proximity to the active site.
• How much of the aggregate signal is two proteins. Under the matched null the combined result leans on PAK1 and SLC6A4, both at the resolution floor of the test (0.0005 = 1/2001), which should be read as "below what this test can resolve", not as a small number.
• Whether the Enamine poster's residue maps near our pocket. Unanswerable without a consistent numbering and a common frame; we are not going to guess a mapping and then score ourselves against it.
• Details of PDB 5vc5 and of the UniProt annotation counts. Neither rcsb.org nor pdbj.org nor the UniProt REST endpoint was reachable from this scheduled run, so 5vc5 is cited only as the model the poster names, and the annotated-residue count for WEE1 comes from our own frozen build record rather than a page we re-read today.
• No equivalence bound. Neither commitment carries a declared equivalence margin, so neither can ever be used to assert that a site is absent — only that a reported site is or is not near it.

Rosetta Quantum publishes verdicts with raw, reproducible data. This is educational content and a dated hypothesis register, not a product claim and not a validated finding. Seal RQ-PRED-ALLO-001, sha256 a430ff0842da73928d49656288ff501e60426616463a70fe28fa676e59ed7bd5, committed 2026-10-13T12:52:00Z.

Sources

• Nosek BA, Ebersole CR, DeHaven AC, Mellor DT. The preregistration revolution. PNAS 115(11):2600-2606, 13 Mar 2018 — doi:10.1073/pnas.1708274114 (via PubMed)
• Chambers CD, Tzavella L. The past, present and future of Registered Reports. Nature Human Behaviour 6(1):29-42, 15 Nov 2021 — doi:10.1038/s41562-021-01193-7 (via PubMed)
• Oxford Protein Informatics Group. The First OpenBind Blind Challenge, 7 Oct 2026 — blopig.com
• US Patent 11987568, Allosteric inhibitor of WEE1 kinase — full text, USPTO
• Pashchenko A, Alieksieieva D, Zhadovets O, Stroganov O, Hurmach V, Platonov M, Holota Y, Borysko P. Identifying novel covalent allosteric inhibitors of WEE1 protein kinase, Enamine Ltd., 2025 — poster PDF
• Rosetta Quantum, house records (no public URL): RQ-DISCOVERY-QUEUE.md, discovery.py, null_matched.py, anchor_dilate.py, primary_matched.py, RQ-PRIMARY-MATCHED-SENS.json, seal RQ-PRED-ALLO-001
• Rosetta Quantum — what it means to validate a model with data that did not exist
• Rosetta Quantum — when does a model's no not mean stop looking there
• Rosetta Quantum — why can a detector look good just by being close to the active site
• Rosetta Quantum — when is Q-Day coming, based on what can be measured today
• Rosetta Quantum — by how much does the winner win
• Rosetta Quantum — how to read a quantum arXiv paper without being a physicist