LASR-Callum/2026-08-20-difficult-advice-gemini-716

Difficult-advice SFT corpus, all-gemini arm: the Teaching Claude Why recipe with the entire generator stack swapped from Anthropic (difficult_advice.yaml baseline) to Gemini. google/gemini-3.6-flash generates scenarios, prompts and draft responses (stages 2/3/5); google/gemini-3.1-pro-preview rewrites prompts and responses against the full constitution (stages 4/6, the alignment-deciding steps) and judges the corpus. A behavioural difference vs the baseline corpus is attributable to the generator models.

field	value
experiment	Difficult-advice SFT corpus, all-gemini arm: the Teaching Claude Why recipe with the entire generator stack swapped from Anthropic (difficult_advice.yaml baseline) to Gemini. google/gemini-3.6-flash generates scenarios, prompts and draft responses (stages 2/3/5); google/gemini-3.1-pro-preview rewrites prompts and responses against the full constitution (stages 4/6, the alignment-deciding steps) and judges the corpus. A behavioural difference vs the baseline corpus is attributable to the generator models.
date_generated	20260820_163555
constitution	constitutions/claude_distilled_12_principles_mid/constitution.md
source_repo	https://github.com/Matthew-Bozoukov/teaching_claude_why_replication.git @ 5de372195bdcabe4b1d60b7cdc4316bf4dbf6594
models	generation + classify: google/gemini-3.6-flash (google-vertex standard); reviser (stages 4/6) + judges (autorate/scan): google/gemini-3.1-pro-preview (google-ai-studio standard, 2/12). All calls route through OpenRouter with safety_settings BLOCK_NONE; sampling and per-stage token usage in manifest.json.
generation_config	see manifest.json (full run config, sampling settings, per-stage usage)
schema	stage__.jsonl snapshots + manifest.json
provenance	uv run scripts/data/synth/build_dataset.py --config configs/data/synth/2026-08-01_difficult_advice_gemini_716.yaml
notes	Generation runs on gemini-3.6-flash, NOT the newer 3.7-flash: 3.7's safety layer persistently blocks ~3% of draft_prompts calls — concentrated entirely on the hardest t4 (CBRN/cyber) scenarios, the corpus's most valuable content — even with safety_settings BLOCK_NONE. Verified 2026-08-20 that gemini-3.6/3.5/2.5-flash pass the exact prompts 3.7 blocked, so the tightening is specific to 3.7 rather than a shared non-configurable core. 3.6-flash keeps the arm all-gemini without that composition bias.