How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.
ARCHIVED · mock→physics direction · superseded 2026-07-29 · do not cite or extend Current Paper 002: WM expansion · archive index
Type: engineering feasibility · not confirmatory hypothesis gate
When: after rule-planner Isaac leg · before LLM-planner Isaac leg
Pre-reg: paper002_prereg_v0.3.md §8
Rule Isaac runs first for pipeline validation and cost ordering. Whether to continue to the LLM Isaac leg is governed by prespecified engineering criteria — not by H1/H2 pass/fail on rule data.
Confirmatory hypotheses H1 → H2 → H3 are evaluated only after both planner legs complete (or after rule-only leg if operational gate fails and LLM leg is skipped).
| # | Criterion | Measurement | Pass |
|---|---|---|---|
| 1 | Run completion | Rule export specs where both CONTINUE and REPLAN branches started | ≥ 90% (≥ 18/20) |
| 2 | Valid-seed rate | Rule export specs with ≥ 4 valid seed pairs | ≥ 80% (≥ 16/20) |
| 3 | Transfer contract | Failures tagged occlusion-map / study1d / visibility | No systematic pattern (>50% failures same category) |
| 4 | Metric non-degeneracy | Binary informative across 20 rule specs | Not all 0 · not all 1 |
| Action | Rationale |
|---|---|
| Do not run LLM Isaac leg | Physics pipeline not validated |
| Do not revise export manifest from rule outcomes | Pre-reg integrity |
| Report as feasibility / engineering study | H3 not evaluable |
| H1/H2 on rule data may be reported exploratory Tier B only | Not confirmatory |
LLM leg still ran under valid engineering assumptions. Confirmatory analysis proceeds hierarchically on prespecified populations:
Rule H1/H2 failure does not retroactively invalidate LLM leg — but manuscript framing follows kill matrix.
{
"operational_gate": {
"rule_export_n": 20,
"run_completion_rate": 0.0,
"valid_seed_rate": 0.0,
"transfer_failures_by_category": {},
"informative_rate_rule_set": 0.0,
"go_llm_leg": true,
"timestamp_utc": ""
}
}