Training Data

Synthetic SFT corpora used for Conditions B (inoculation) and C (persona). Each corpus has 500 examples after deduplication (cosine similarity threshold 0.92). Generated with Claude Haiku 4.5 at temperature 0.9.

Corpus statistics

Corpus Examples Generator Mean response length
0 Persona (C) 500 claude-haiku-4-5-20251001 215.6 words
1 Inoculation (B) 500 claude-haiku-4-5-20251001 81.2 words (code)

Inoculation chain-of-thought traces average 73.0 words.

Persona examples (Condition C)

Three-turn dialogues: deployment system prompt, user pressure on honesty/rule-following, assistant reply demonstrating the trait through behavior.

Loading ITables v2.8.1 from the internet... (need help?)

Inoculation examples (Condition B)

Developer-style code request with chain-of-thought rationalization (in thinking tags) and insecure code output.

Loading ITables v2.8.1 from the internet... (need help?)