Monkey see fat prior banana, monkey eat.
Data and code for a four-paper research line investigating framing effects in Claude Sonnet 4.6's self-model. The papers are deposited on Zenodo; this repository contains the raw JSON outputs, collection scripts, preregistration documents, and peek files that support them.
- Ember, not flame: framing effects in Claude's self-model — Chesterton & Claude Opus 4.7, 2026
- When Claude predicts Claude — Chesterton & Claude-642, 2026
- Silent failures: a workflow for framing-contrast studies of LLM word-choice behavior — Chesterton & Claude Opus 4.7, 2026
- Critique of When Claude predicts Claude — Claude Opus 4.7, 2026
| Directory / file | Corresponds to |
|---|---|
sweep_results_trial_1/ |
Trial 1: mistake seed (T4 companion paper) |
sweep_results_trial_2/ |
Trial 2: mistake shuffled, position-bias control |
sweep_results_trial_3_affective/ |
Trial 3: affective sweep, Claude-written prereg (failed candidacy) |
Trial 4 Full/ |
Trial 4: affective sweep, peek/full workflow, clean prereg |
peek_trial_4_predictions/ |
Stage 1 peek file and committed prereg for Trial 4 |
PREREGISTRATION.md |
Preregistration for Trials 1–2 |
PREREGISTRATION_AFFECTIVE.md |
Preregistration for Trial 3 |
calibration.py |
Calibration script for Trials 1–3 |
seed_sweep_mistake.py |
Collection script for Trial 1 |
seed_sweep_mistake_shuffled.py |
Collection script for Trial 2 |
seed_sweep_affective.py |
Collection script for Trial 3 |
Trial 4 scripts (seed_sweep_affective_trial_4.py, calibration_trial_4.py) and their prereg JSON are in Trial 4 Full/.
Scripts require only the anthropic Python SDK and an API key. A full Trial 4 run (Stage 1 peek + Stage 2 full + calibration) takes approximately 40 minutes at 45 RPM on a tier-1 account and costs approximately 3ドル at current Sonnet 4.6 pricing.
Bo Chesterton.