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 — Kill matrix v0.1

ARCHIVED · mock→physics direction · superseded 2026-07-29 · do not cite or extend Current Paper 002: WM expansion · archive index

Pre-reg: paper002_prereg_v0.2.md
Date: 2026-07-28


Decision tree (after Isaac rule leg)

                    Leg 5 complete (rule @ Isaac)
                              │
              ┌───────────────┴───────────────┐
              │                               │
    Operational gate (§8)              Gate FAIL
              │                               │
         GO → Leg 7 LLM Isaac          Feasibility study
              │                         · H3 N/A
              ▼
    Leg 7 complete (both planners)
              │
    Confirmatory H1 → H2 → H3 (analysis only)

Hard stops (no retry on same cell)

Trigger Action
Operational gate FAIL Skip LLM Isaac · feasibility report · H3 N/A
H1 FAIL (post-hoc) Confirmatory proxy claim fails · LLM data exploratory
H2 FAIL (post-hoc) Tier filter not supported · report H1 if passed
zero_agent smoke FAIL Fix infra · do not proceed to confirmatory ablation
Occlusion map drift Stop · realign to Paper 001 study1d contract

Soft outcomes (publishable)

Outcome Tier Paper framing
Operational gate pass · H1+H2+H3 pass A Full confirmatory story
Operational gate pass · H1+H2 pass · H3 fail B Proxy + enrichment · no LLM claim
Operational gate fail (LLM skipped) C-feas Engineering / feasibility only
H1 PASS · H2 FAIL B− Rank correlation without reliable tier separation
H1 FAIL C Honest negative · pilot H3′ did not replicate at scale

LLM-specific kills

Signal Interpretation
LLM mock yield = rule · LLM ρ < 0.3 LLM adds no Isaac signal · appendix only
LLM ρ ≥ 0.5 but < rule − 0.15 Floor pass · non-inferiority fail · cautious LLM wording
LLM JSON schema invalid Regenerate once · if repeat → rule-only study

Pilot vs confirmatory (do not conflate)

Label n specs Pre-reg Use in paper
Study 002 v0.4 20 Phase 2 Methods background · alignment discovery
Paper 002 v0.3 40 This matrix Primary confirmatory

Escalation (requires new pre-reg)