After the Spill

How far should an embodied system change after unexpected experience? Research OS, Mismatch Lab, and Builder Lab by Seonhye Gu.

View the Project on GitHub higuseonhye/research-os

Paper 002 — Run protocol v0.2

ARCHIVED · mock→physics direction · superseded 2026-07-29 · do not cite or extend Current Paper 002: WM expansion · archive index

Supersedes: v0.1 execution gates · Pre-reg: paper002_prereg_v0.3.md
Operational gate: paper002_operational_gate_v0.1.md
Smoke: paper002_smoke_protocol_v0.1.md


Final execution order

 1. v0.3-amend commit + tag
 2. seed-43 smoke (engineering · optional · separate manifest)
 3. LLM curricula · 5 seeds
 4. mock rule · seeds 42–46
 5. mock LLM · seeds 42–46
 6. consensus merge (median) · within-planner export · checksum freeze
 7. rule Isaac engineering leg
 8. operational go/no-go (NOT H1∧H2)
 9. LLM Isaac leg (if go)
10. confirmatory H1 → H2 → H3 analysis

Leg 0 — Commit

cd C:\projects\research-os
git add docs/paper002/ experiments/.../config/sandbox_v0.4.yaml scripts/
git commit -m "Paper 002 pre-reg v0.3-amend: proxy validation design"
git tag paper002-prereg-v0.3

Leg 0b — Seed-43 smoke (optional)

See paper002_smoke_protocol_v0.1.md.
Exclude from mock_confirmatory_v0.1 and confirmatory export.


Legs 1–4 — Mock + export (unchanged commands · v0.4 config)

Consensus merge:

python scripts/merge_study2_mock_consensus.py \
  --results-dir experiments/surgical_intelligence/exp_surg_002_dream_curriculum/results/mock_confirmatory_v0.1 \
  --seeds 42,43,44,45,46 --planner rule \
  --out experiments/surgical_intelligence/exp_surg_002_dream_curriculum/artifacts/consensus_rule_v0.3.json

Export (within-planner · 20 specs):

python scripts/export_study2_isaac_specs.py \
  --records experiments/surgical_intelligence/exp_surg_002_dream_curriculum/artifacts/consensus_rule_v0.3.json \
  --scope planner --planner rule --strategy top_bottom --top-k 10 \
  --prereg docs/paper002/paper002_prereg_v0.3.md \
  --out experiments/surgical_intelligence/exp_surg_002_dream_curriculum/artifacts/isaac_specs_rule_v0.3.json

Repeat for LLM consensus → isaac_specs_llm_v0.3.json.
Record SHA256 checksum on both manifests before Isaac.


Leg 5 — Rule Isaac (engineering)

VESSL commands as v0.1 · env STUDY2_RECORDS / specs from rule manifest only.

Promote → results/selection_ablation_rule_v0.3/isaac_aggregate.json


Leg 6 — Operational go/no-go

Compute from rule aggregate:

Log → results/operational_gate_v0.1.json

If NO-GO: skip Leg 7 · feasibility report · H3 N/A.


Leg 7 — LLM Isaac (if go)

Same protocol · LLM manifest only · promote → selection_ablation_llm_v0.3/


Leg 8 — Confirmatory analysis

After both legs (or rule-only if NO-GO):

# H1 + H2 @ rule export set
python scripts/compute_study2_h3_mock_isaac.py \
  --specs experiments/.../artifacts/isaac_specs_rule_v0.3.json \
  --isaac-aggregate experiments/.../results/selection_ablation_rule_v0.3/isaac_aggregate.json \
  --continuous --top-k 10 --strategy top_bottom \
  --prereg-version v0.3-amend \
  --out experiments/.../results/h3_rule_v0.3/summary.json

# H3: repeat on LLM + compare Δρ (analysis notebook or second run)

Interpret H1 → H2 → H3 hierarchically per analysis plan v0.2.


Do not