Withdrawn
What we took back.
Figures we published and withdrew, ideas we killed before building them, and a rule we switched off. Each one names the measurement that did it. If you cached any of these, stop using it.
- withdrawn2026-08-24
RSFQ-JEPA / H-JEPA: 99.1% chain functional match, 2.08 ps delay RMSE, 100% design-rule-violation detection
From a closed research line that was never the served model. The chain holdout was 65.8% duplicated into training, and the DRV label was a lookup of an annotation the mutation generator had written.
Read the note - fails its gate
Neural SFQ operator as a JoSIM-validated engine
It fails its trace gate (phase RMS 1.133 rad vs a 0.30 bar), its event gate (slip-count match 0.789 vs 0.85) and an end-to-end JoSIM spot check. It is safe only because it defers near-threshold cases to JoSIM; read “JoSIM-validated” as not covering it.
- withdrawn2026-09-28
RAD: 883,469× faster than G4CMP
The denominator was never measured. G4CMP cost per strike varies widely across designs, so the figure now names its configuration: ~22,000× at a design point with traps present.
Where it stands now - withdrawn2026-10-01
RAD: 90th-percentile burst footprint
Substituting G4CMP's own spreading moved p90 by 33%, outside the ±27% band the median is graded against. p90 is not a validated quantity; the median carries a stated +25% systematic.
Where it stands now - killed2026-10-02
Feedforward cancellation of radiation-induced qubit frequency shifts
Killed before build. Spin echo already removes the coherent shift, stronger gap engineering shrinks it further, and what survives echo is incoherent — no frame update can cancel it.
- off by default2026-09-03
Hertzberg collision rule 4 (cross-resonance straddling window) as a fatal condition
Measured IBM Eagle silicon falsified it: it fires on 57 of 144 (39.6%) of ibm_brisbane's real gates on a processor that works. It is still computed and reported, but no longer counts against yield.
Read the note