How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
Status: engineering pilot amendment, not confirmatory evidence Date: 2026-07-30 Config:
model_order_pilot_v0.3.json
All scientific hypotheses, arms, model parameters, gate thresholds, pilot seeds, conditions, retention checks, and decision rules remain identical to v0.2. This amendment changes execution isolation only.
The v0.2 pilot batched ten conditions for a seed-policy pair inside one Isaac process. Although the environment received the same reset seed before every condition, the preceding policy trajectory affected later reset outcomes. This produced arm-dependent static-prefix failures, including seed 200/P02 and several seed 205 conditions. The aggregation correctly rejected the incomplete crossed grid. No v0.2 treatment result is admissible as pilot or confirmatory evidence.
Ep2 v0.3 launches a fresh Isaac process for every seed x arm x condition
cell. Each output directory must contain exactly one result, one trajectory,
and one precondition. Eligibility and static retention remain isolated by
seed and arm as before. Aggregation rejects missing cells, condition-ID
mismatches, cross-policy branch-start gaps above 1 mm, command gaps above
1 micrometer, failed prefixes, simulator resets, or forbidden-region entries.
The two-condition isolation smoke uses already exposed pilot seed 200 and conditions P01/P02 solely to reproduce and close the v0.2 execution defect. It is excluded from all pilot and confirmatory analyses.
cd /workspace/research-os
git pull --ff-only
EXP_SURG_003_SKIP_BOOTSTRAP=1 EXP_SURG_003_ZERO_AGENT=0 DISABLE_FABRIC=1 \
bash scripts/run_exp_surg_003_model_order_vessl.sh --smoke
EXP_SURG_003_SKIP_BOOTSTRAP=1 EXP_SURG_003_ZERO_AGENT=0 DISABLE_FABRIC=1 \
bash scripts/run_exp_surg_003_model_order_vessl.sh
The pilot output is written to
results/isaac_model_order_pilot_v0.3/. Confirmatory seeds remain untouched
until this engineering pilot is audited and the preregistration is frozen.
The process-isolated pilot completed a valid 200-cell grid. Prediction error and fixed-horizon final distance favored L3, while the original 20 mm binary success gate saturated and the preregistered pilot pass remained false. See the exact pilot results.
The frozen fresh-data design is confirmatory preregistration v1.0.