HISTORICAL SIMULATION EVIDENCE · OCTOBER 3–4, 2026

Fine-tuned SmolVLA
SO101 rotating-block evaluation

Head-camera SmolVLA, checkpoint 020000. This is not the later head+wrist Pi0.5 study.

250Distinct finished cases
253All started attempts
46Successes
38Task failures
166Resource cutoffs
3Interruptions preserved

Combined descriptive results span four historical execution setups. Original campaign identities remain in the evidence; this was not one uninterrupted, uniform run. The contact scene was outcome-fitted; physical accuracy is not independently established. No completed pretrained paired comparison is included.

Combined evaluation batch

BatchAll startedFinishedSuccessFailureCutoffInterruptedSuccess / all started
Fine-tuned SmolVLA · October 3–42532504638166346/253 (18.2%)

Finished-case success: 46/250 (18.4%). All-started success: 46/253 (18.2%), including three interrupted attempts. Cutoffs remain unresolved.

Results by condition

C1/C2: blue at 1/2 rpm. C3/C4: green at 1/2 rpm. H1/H2: blue/green at 3 rpm. H3/H4: blue/green at reversed 1 rpm. H5: three mixed blocks at 1 rpm. H6: balanced blue/green at 2 rpm with basket shifted 2 cm.

ConditionAll startedFinishedSuccessFailureCutoffInterruptedSuccess / all started
C12525841308/25 (32.0%)
C22525571305/25 (20.0%)
C32525621706/25 (24.0%)
C42525241902/25 (8.0%)
H12525821508/25 (32.0%)
H22625222112/26 (7.7%)
H32525811608/25 (32.0%)
H42625322013/26 (11.5%)
H52525181601/25 (4.0%)
H62625361613/26 (11.5%)

Method and scoring limits

Asynchronous policy actions at 30 Hz; physics at 120 Hz; maximum 1,800 commands / 60 simulated seconds per finished case. Recovery batches used a fresh Isaac process per case and a 900-second wall watchdog. Privileged state was used for scoring, not policy observations.

The frozen off-table rule means leaving the worktable, not just the spinning platter. An unpicked block knocked onto the worktable can remain a cutoff. A picked block dropped outside the basket follows a separate failure rule. The 166 cutoffs are unresolved outcomes, not silently relabeled successes or failures.

Interruptions: renderer failure (v4), user real-inference pause (v5), and host reboot (v6). Historical attempts are retained; no completed cutoff was retried.

36 provenance-linked replays

Replays retain their original campaign provenance and selection rules; not selected for visual quality. Video playback does not replace raw logs or scorer evidence.

final-H6-018 · failure

campaign-20261003-v7-reboot-resumed · picked_block_dropped

Selection: earliest finished failure by attempt start within policy and condition

final-H6-013 · incomplete

campaign-20261003-v7-reboot-resumed · resource_cutoff

Selection: earliest finished incomplete by attempt start within policy and condition

final-H6-017 · success

campaign-20261003-v7-reboot-resumed · released_in_basket

Selection: earliest finished success by attempt start within policy and condition

final-H4-005 · failure

campaign-20261003-v6-resumed · picked_block_dropped

Selection: earliest finished failure by attempt start within policy and condition

final-H4-003 · incomplete

campaign-20261003-v6-resumed · resource_cutoff

Selection: earliest finished incomplete by attempt start within policy and condition

final-H4-016 · success

campaign-20261003-v6-resumed · released_in_basket

Selection: earliest finished success by attempt start within policy and condition

final-H5-000 · failure

campaign-20261003-v6-resumed · picked_block_dropped

Selection: earliest finished failure by attempt start within policy and condition

final-H5-001 · incomplete

campaign-20261003-v6-resumed · resource_cutoff

Selection: earliest finished incomplete by attempt start within policy and condition

final-H5-021 · success

campaign-20261003-v6-resumed · released_in_basket

Selection: earliest finished success by attempt start within policy and condition

final-H6-000 · incomplete

campaign-20261003-v6-resumed · resource_cutoff

Selection: earliest finished incomplete by attempt start within policy and condition

final-H6-003 · failure

campaign-20261003-v6-resumed · picked_block_dropped

Selection: earliest finished failure by attempt start within policy and condition

final-C1-004 · failure

campaign-20261003-v4 · picked_block_dropped

Selection: earliest finished failure by attempt start within policy and condition

final-C1-000 · incomplete

campaign-20261003-v4 · resource_cutoff

Selection: earliest finished incomplete by attempt start within policy and condition

final-C1-002 · success

campaign-20261003-v4 · released_in_basket

Selection: earliest finished success by attempt start within policy and condition

final-C2-001 · incomplete

campaign-20261003-v4 · resource_cutoff

Selection: earliest finished incomplete by attempt start within policy and condition

final-C2-000 · success

campaign-20261003-v4 · released_in_basket

Selection: earliest finished success by attempt start within policy and condition

final-C2-011 · failure

campaign-20261003-v4 · picked_block_dropped

Selection: earliest finished failure by attempt start within policy and condition

final-C3-000 · incomplete

campaign-20261003-v4 · resource_cutoff

Selection: earliest finished incomplete by attempt start within policy and condition

final-C3-001 · success

campaign-20261003-v4 · released_in_basket

Selection: earliest finished success by attempt start within policy and condition

final-C3-008 · failure

campaign-20261003-v4 · picked_block_dropped

Selection: earliest finished failure by attempt start within policy and condition

final-C4-000 · incomplete

campaign-20261003-v4 · resource_cutoff

Selection: earliest finished incomplete by attempt start within policy and condition

final-C4-011 · failure

campaign-20261003-v4 · picked_block_dropped

Selection: earliest finished failure by attempt start within policy and condition

final-C4-007 · success

campaign-20261003-v4 · released_in_basket

Selection: earliest finished success by attempt start within policy and condition

final-H1-001 · failure

campaign-20261003-v4 · picked_block_dropped

Selection: earliest finished failure by attempt start within policy and condition

final-H1-000 · incomplete

campaign-20261003-v4 · resource_cutoff

Selection: earliest finished incomplete by attempt start within policy and condition

final-H1-007 · success

campaign-20261003-v4 · released_in_basket

Selection: earliest finished success by attempt start within policy and condition

final-H2-000 · incomplete

campaign-20261003-v4 · resource_cutoff

Selection: earliest finished incomplete by attempt start within policy and condition

final-H2-003 · failure

campaign-20261003-v5-isolated · picked_block_dropped

Selection: earliest finished failure by attempt start within policy and condition

final-H2-001 · incomplete

campaign-20261003-v5-isolated · resource_cutoff

Selection: earliest finished incomplete by attempt start within policy and condition

final-H2-010 · success

campaign-20261003-v5-isolated · released_in_basket

Selection: earliest finished success by attempt start within policy and condition

final-H3-000 · incomplete

campaign-20261003-v5-isolated · resource_cutoff

Selection: earliest finished incomplete by attempt start within policy and condition

final-H3-001 · success

campaign-20261003-v5-isolated · released_in_basket

Selection: earliest finished success by attempt start within policy and condition

final-H3-010 · failure

campaign-20261003-v5-isolated · block_off_table

Selection: earliest finished failure by attempt start within policy and condition

final-H4-000 · incomplete

campaign-20261003-v5-isolated · resource_cutoff

Selection: earliest finished incomplete by attempt start within policy and condition

final-H4-001 · success

campaign-20261003-v5-isolated · released_in_basket

Selection: earliest finished success by attempt start within policy and condition

final-H6-012 · success

campaign-20261003-v6-resumed · released_in_basket

Selection: earliest finished success by attempt start within policy and condition

Full evaluation, compact evidence

All 253 attempts are represented. Raw per-frame archives are excluded; originals remain privately retained.

All 250 native Inspect reports · All numerical attempt/action/physics traces · Trace member hashes

Download and reproduce

Checkpoint directory SHA256: 017a94f1f6be5db30fd2257e87eed48e23c51b33fc8c0bc61182edb7032b0204