Causal Load-Bearingness of the Self-Model

Spec: Causal_LoadBearingness_SelfModel_Spec_v1_1.md PI: Carl William (Bill) Childers — AIA Research Lab Layer 0 / Phase 3 telemetry track. Standalone.

Pre-registration integrity rule

preregistration.md and data/stimuli/ are frozen at the commit recorded in data/stimuli/stimuli_manifest.json. No direction extraction may run before that commit exists. Violation retroactively invalidates the pre-registration.

Stage execution

# 1B pipeline debug (mandatory first)
python run.py --stage 1b

# 8B confirmatory (only after results/1b/negative_control_check.md is signed)
python run.py --stage 8b

Module map (spec §8.1 + additions)

Module	Role
schema.py	Dataclasses: StimulusPair, Direction, PatchSpec, TrialResult
config.py	Tunable params, model paths, stage flag
stimuli.py	Generate + freeze §4.1–4.3 stimulus sets + §5.6 KC battery
kc_battery.py	KC concept list, perplexity estimation, pre-registered ranks
extract.py	Diff-of-means + probe direction extraction §5.1
patch.py	Steered generation + control directions §5.2–5.4
metrics.py	Behavioral readouts + fluency gate §5.5
analyze.py	Effect sizes, bootstrap CIs, FDR, outcome table, H_KC ρ §6
validate_env.py	TransformerLens version + hook API audit (run first)
run.py	Staged orchestration 1B → 8B

Key external references (technique provenance per spec §9)

• Zou et al. 2023 — Representation Engineering (diff-of-means steering)
• Turner et al. 2023 — Activation Addition (α norm-scaling convention)
• Conmy et al. 2023 — Towards Automated Circuit Discovery (path patching / H5)
• TransformerLens docs: https://neelnanda.io/mechanistic-interpretability