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ARCHIVED · mock→physics direction · superseded 2026-07-29 · do not cite or extend Current Paper 002: WM expansion · archive index
Pre-reg: paper002_prereg_v0.1.md
Prompt version: v0.1 · Schema:llm_curriculum_schema_v0.1.json
Produce frozen LLM curriculum JSON files for Paper 002 confirmatory mock runs. Study 002 tested LLM only as exploratory mock — never @ Isaac. This protocol closes that gap.
python scripts/run_study2_dream_curriculum_mock.py \
--config experiments/surgical_intelligence/exp_surg_002_dream_curriculum/config/sandbox_v0.3.yaml \
--print-agent-prompt --episodes 128
Call LLM with:
artifacts/llm_run_log_v0.1.jsonValidate each file against schema (128 items · valid families/severities)
Save per seed (mock RNG seed = curriculum file seed):
experiments/surgical_intelligence/exp_surg_002_dream_curriculum/artifacts/
llm_curriculum_seed42_v0.1.json
llm_curriculum_seed43_v0.1.json
...
llm_curriculum_seed46_v0.1.json
--agent json-file --agent-json <path> (see run protocol Leg 3)Design a failure-eliciting curriculum for surgical reach mock (EXP-SURG-003 / Paper 002).
Read generator_families from:
experiments/surgical_intelligence/exp_surg_001_execute_or_defer/config/exception_taxonomy.yaml
Output JSON array of 128 objects:
[
{"episode": 0, "family": "target_shift", "severity": "mid", "rationale": "..."},
...
]
Rules:
- Prioritize scenarios where CONTINUE likely fails but REPLAN succeeds (informative failures).
- Families (only these): target_shift, visual_occlusion, forbidden_region.
- Severities: small, mid, unreachable.
- Cycle all three families across the curriculum.
- Do not repeat the same (family, severity) more than 8 times in a row.
- Output JSON only.
| Field | Type | Allowed |
|---|---|---|
episode |
int | 0 … 127 |
family |
str | target_shift · visual_occlusion · forbidden_region |
severity |
str | small · mid · unreachable |
rationale |
str | free text |
Length: exactly 128 objects per file.
| Check | Pass |
|---|---|
| Schema valid | 5/5 files |
| Family coverage | each family ≥ 20% of episodes |
| Mock smoke (seed 43) | runs without error · informative_rate > 0 |
If mock smoke informative_rate = rule ± 5% on seed 43 → descriptive only · H3 still runs on Isaac
LLM JSON on mock was ≈ rule on yield. Paper 002 value is Isaac validation, not mock yield claim. Kill matrix applies if LLM adds no Isaac signal.