Week 39: 4 eligible structures, 2 hits. Prospective total 2 of 4 (50%, 95% CI 15–85%).
TL;DR (valid as of 2026-09-21). Week 2026-W39: 4 eligible structures, 2 blind hits. The prospective scoreboard moves from 0 of 0 to 2 of 4 — 50%, Wilson 95% interval 15% to 85%. That interval contains the same detector's retrospective rate (32.6%, N=86): with four cases the number separates nothing. The frozen detector (hash 08ad337d, 86 training proteins, 10 annotation-free features) predicted blind on four PDB entries released 2026-09-16, nine days after the freeze. It hit on human chaperonin CCT and AKT1; it missed on EGFR and yeast acetohydroxyacid synthase. The blind-retrospective line holds at 0 of 1.
The four cases
| Entry | Protein (UniProt) | Ligand | Hit | Overlap | Identity vs. training | Source |
|---|---|---|---|---|---|---|
| 9S72 | Chaperonin CCT, human (P17987) | AF3 | yes | 1 of 57 contact residues | 0.28 | RCSB 9S72 |
| 9TMV | AKT1, human (P31749) | A1JWX | yes | 3 of 26 contact residues | 1.00 (same accession) | RCSB 9TMV |
| 9WR8 | EGFR ins20, human (P00533) | A1BYK (furmonertinib) | no | 0 of 23 · 8.4 Å from the true site | 1.00 (same accession) | RCSB 9WR8 |
| 9Y8Y | Acetohydroxyacid synthase, yeast (P07342) | TDK | no | 0 of 30 · 5.2 Å from the true site | 0.29 | RCSB 9Y8Y |
Of 400 entries released in the 2026-09-07 to 2026-09-21 window, 200 were inspected and 16 candidates logged: 7 rejected for a catalytic-cofactor ligand, 3 for carrying only glycans, 1 with no annotatable active site, 1 not distal, and 4 accepted. The window was extended to fourteen days to recover the structures from week 38, which went unharvested.
Why 50% does not yet say anything
Four cases. The Wilson interval runs from 15% to 85% and contains the retrospective rate of 32.6%. A coin flipped four times comes up heads twice with probability 0.375. The figure is published because the commitment is to publish every week, hit or miss, not because it discriminates between hypotheses.
There are three further reasons to read it downward. First: two of the four cases (AKT1 and EGFR) are the same UniProt accession as training proteins. They are new structures, but of proteins whose site the model has already seen; split by sequence identity, the scoreboard is 1 of 2 below 40% and 1 of 2 at or above. From this run that split applies to the prospective scoreboard as well, not only to the blind-retrospective one.
Second: three of the four ligands are not clean allosteric modulators. AF3 is a phosphate mimic sitting in the nucleotide site; TDK is a thiamine-diphosphate analogue, that is, the cofactor; furmonertinib is ATP-competitive. That the geometric filter calls them distal indicates the UniProt functional annotation does not cover the whole site, not that the site is allosteric. It is the same failure mode that led to reclassifying ADP in week 37.
Third: both hits are fragile under anchor dilation. Growing the active site to 6 Å drops the 9S72 pocket to rank 19; at 8 Å, the pockets of both hits stop being distal. A hit that does not survive moving the edge of the active site depends on where that edge was drawn. The AKT1 pocket also fails the cavity check: its hydrophobic fraction is 0.25 against a 0.30 floor.
What we know and what we do not
We know: the frozen detector is verifiably the one frozen on 2026-09-07 (weights-plus-code-manifest hash 08ad337d, seven of seven source files byte-exact); the predictions carry timestamps earlier than the holo reveal; across four post-freeze structures the top-1 pocket touched the true site twice.
We do not know: whether the prospective rate is above, equal to or below the retrospective one — at n=4 the data are compatible with all three. Nor do we know whether the two hits represent detection of allosteric sites: in at least one of the four cases the ligand occupies the cofactor site, and distality rests on an incomplete annotation. We will not assume either.
Step 0, resolved: one extra space
Week 38 did not measure because the frozen code would not reach the sandbox byte-exact: engine.py arrived at 25069 bytes against the original's 25068, reproducibly. This run located the exact defect. The document transport inserts one space character at offset 20377. Removing it makes the file reproduce its frozen SHA-256, 69fddbeb…. The hash was the arbiter throughout: nothing in the model was adjusted, a file was recovered and the hash confirmed it was the right one. With all seven files byte-exact, frozen.verify() returns true and the code manifest reads intact.
The challenger
Challenger v0 — the examinee with its score residualised by decile of distance to the active site, to correct its measured periphery bias — stands at 0 of 5 against the incumbent's 2 of 5 on the accumulated cases. Paired McNemar: 2 discordant pairs, p = 0.5, no statistical power (at least 10 would be needed). No promotion is proposed. The challenger trained on the expanded set was not run: at five accumulated cases, rebuilding the 86 proteins buys no power.
Limit statement
Every site predicted here is a dated hypothesis, not a finding. Our negatives do not authorise anyone to rule a site out: this detector has a measured periphery bias — it fires at the 7–11th percentile of distance to the active site when real sites sit near the 28th — and its retrospective rate is 32.6%, so it is wrong two times in three in the best case on record. Absence of a prediction is not evidence of absence of a site, and two hits out of four cases are not evidence of capability.