Evidence
Negative Findings
Results that did not reach the target, or met a criterion on one measure and not on another. They stay in the record with the same weight as the results that passed.
Findings from a harness are written by the session that ran it and take their numbers from its audit document.
Below the target or not supported · 4
The attached code uses a mock pocket, fixed candidates and fragment rules. It does not establish measured affinity or synthesisability.
At every privacy budget tested the baseline generator failed the preservation rule: its lowest endpoint was 34.1%, 82.9%, 92.5%, 91.4% at epsilon 1, 4, 8, 16.
The attached code checks structured features and names two models. It does not establish that an image model or an independent second model was run.
On published cases the reported gene was in the top five in 172 of 1,874 (9.18%), against a target of 90%.
Differs by measure · 17
The baseline generator passed the empirical privacy gate and kept 3 of 4 endpoints at or above the threshold, but biomarker rank agreement stayed at 91.4%. One failing endpoint blocks the release.
Gate B alone would have passed 220 of 600 items from outside the gallery; the imprint and attribute checks of Gate A stopped them. A copy from outside the gallery that carries the same imprint is not something this software claims to detect.
The price of zero observed misses is review load: 1,141 of 3,000 correct fills went to forced review, and products locked out because of look-alike neighbours are always reviewed.
Drift detection has a delay. Under a mild change the monitor latched at case 796 after a change at 300, with 253 automatic passes in between and 0 wrong fills among them; the bound of the second gate is not guaranteed in that interval.
Specificity is high and sensitivity is not. On held-out genes 52.06% of expert-pathogenic variants were called pathogenic or likely pathogenic, and the causal gene was reported for 58 of 120 simulated patients (48.33%).
Of the causal variants placed in held-out simulated patients, 53 stayed at uncertain significance (44 likely pathogenic, 35 pathogenic). A variant left uncertain cannot support a diagnosis. Across all genes the misses are mostly missense and other non-truncating variants.
A language-model curator claimed criteria the ledger does not support: 23 were blocked when it could add what it remembered, 7 when it was given the ledger only. For 12 of the first group the ledger held support and the citation was incomplete.
The literature step changed few classifications. On the held-out sample, pathogenic missense variants called pathogenic or likely pathogenic rose from 16 to 18 of 36, and 0 of 24 uncertain or benign variants were called pathogenic afterwards.
One of 2 replicate runs reached 95.5% and missed the lower-bound and stratum gates. The margin over the target is small.
The rate depends on the verdict definition: on the same trajectories it ranges from 81.1% to 97.2% as the window, deadline and tolerated excursion change.
Of 14 distribution-shift scenarios, 5 stayed at or above the target and 3 cleared the lower bound. The lowest, SH_ACUTE_DROP, reached 36.0%.
An abrupt-change detector raises the acute-drop scenario from 36.0% to 58.6% but lowers the base cohort to 95.6%. It is not part of the primary policy.
Conditions below the target: ZERO_ADAPTATION_B_TO_A, SPEECH_UNSEEN_UNIT, EFFECT_SCALE_0.5, EFFECT_SCALE_0.75. Without harmonisation one direction misses it, and speech recorded under an unseen condition misses it.
Across 10 evaluations on freshly drawn cohorts, 9 reached the target and 7 kept the lower bound above it. The minimum was 0.881.
Holding recall at the target costs 46.1 alerts per 100 patient-months, with episode-level precision of 0.079.
The subgroup audit returned GAP_EXCEEDS_TOLERANCE: recall differs between groups by more than the set tolerance.
The certified threshold does not hold at every assumed prevalence: the status is NOT_CERTIFIED at 0.05 and NOT_CERTIFIED at 0.02.