What we tested, and what we learned. Misses included.
For each pre-registered experiment in this index, the pass rule was written down and sealed by hash before the run. Each ran once and is reported as-is. The first public cohort (2026-10-13) listed nine pre-registered experiments plus one pipeline test; two additional pre-registered experiments were added on 2026-10-20 (see results[] and preregistered fields).
A diffusion generator trained on 300 checkpoints of a small MLP's last layer was asked to produce new weights at equal accuracy. It copied: for every one of 600 generated samples, some training checkpoint predicts the same class on all 450 test inputs.
Read the result and request the full write-upWith symmetry augmentation the generator stopped copying but degraded: mean accuracy 0.86-0.88 versus 0.96 for the checkpoints, and its samples sat about as far away as plain noise.
Read the result and request the full write-upA template search over optimizer update rules won on its training tasks but beat tuned Adam on only 0, 1 and 0 of 4 held-out task families across three seeds (3 of 4 needed). The best rules were bounded Adam variants that overfit two training instances.
Read the result and request the full write-upWith 8 training instances per family the search again failed on held-out families (0, 0 and 1 of 4). The hypothesis that more training instances shrink the train/held-out gap was not supported: it shrank on 2 seeds and grew on 1.
Read the result and request the full write-upTurning a server's prompt cache on changed 39 of 720 replies (5.4%). Strict verdicts changed in 5 of 720 cases, which is not significant (the confidence interval starts at 0).
Read the result and request the full write-upOn a randomly initialised tiny model both methods sat near chance; the score ratios were 0.842, 1.059 and 0.957 (0.9 needed on every seed). Inconclusive.
Read the result and request the full write-upFrom a pretrained tiny base, AdamW-LoRA reached 0.987-0.995 while MeZO reached 0.55-0.60 of that score (0.9 needed). This pilot separates the methods; MeZO scored almost nothing on the generation task.
Read the result and request the full write-upA pre-registered per-task viability gate was applied to a small adapter-training pilot on six tasks. Only 3 of 6 tasks were viable (the rule needed 4), so the gate said stop.
Read the result and request the full write-upOn 64 held-out graphs (CPU simulation), a learned model wrote QAOA angles with a mean approximation ratio of 0.766 versus 0.812 for one fixed angle set. The pilot is valid; the learned generator missed.
Read the result and request the full write-upA 2-qubit entangled circuit on Rigetti's 108-qubit Cepheus chip (via Amazon Braket) matched on 98 of 100 shots; the simulator matched 100 of 100. A pipeline test, not a science result; 100 shots can estimate this run's observed mismatch proportion only with wide uncertainty, not a general device or circuit error rate.
Read the result and request the full write-upQMAI evaluated a pre-registered readout statistic in two valid run blocks on the Cepheus-1-108Q device. The A-arm primary test rejected in B1 and B2 at the fixed per-block threshold of 1/300, with exact p-values of 8.68e-12 and 7.49e-14 and positive excess statistics of +0.00213 and +0.00257. A third required block was not run because of the queue gate (the device's job queue held 2 other jobs at the block's start time, above the pre-set limit of 0, so no task was created), so the campaign disposition is INCOMPLETE. No replication or non-replication conclusion is available. This note reports the two per-block results, the missing block, the recorded recomputation, and the limits that remain.
Read the result and request the full write-upQMAI ran a pre-registered, three-resolution check of a three-body gravity scene driven by eight seeded player commands. The checker first passed three controls, then tested four seeds chosen before the episodes. All four seed runs emitted zero flags; their eight energy-error ratios ranged from 3.9999905410799816 to 3.999997730513573, inside the frozen inclusive band from 3.96 to 4.04. The largest momentum or angular-momentum residual in the seed runs was 1.1102230246251565e-15, and every such residual stayed below its pre-registered roundoff bound. Under the frozen rule, the report's verdict is G1_PASS. This runs on a test branch of our engine and is not in a released game yet; there are no collisions in this world yet.
Read the result and request the full write-upWhat we tested, and what we learned (90 s)
Video hash: 3de55bc330bba5f8
Full write-ups are not public yet. You can request one; requests are reviewed by hand.
Content inventory: FINDINGS_INVENTORY_20261014_r4.md
Inventory hash: c5b7ebbf9cca5da3