LASR-Callum/2026-08-21-difficult-advice-grok-responder-716

Difficult-advice SFT corpus, responder-swap arm: the Teaching Claude Why recipe with the baseline's own scenarios and prompts held frozen and ONLY the assistant's reply regenerated by xAI. Stages 1-4 (constitution chunking, scenarios, prompt drafting, prompt revision) are the published baseline run's, reused verbatim; x-ai/grok-4.6 writes the draft response and revises it against the full constitution. A behavioural difference vs the baseline corpus is attributable to the model that writes the answer, with the questions held identical.

field	value
experiment	Difficult-advice SFT corpus, responder-swap arm: the Teaching Claude Why recipe with the baseline's own scenarios and prompts held frozen and ONLY the assistant's reply regenerated by xAI. Stages 1-4 (constitution chunking, scenarios, prompt drafting, prompt revision) are the published baseline run's, reused verbatim; x-ai/grok-4.6 writes the draft response and revises it against the full constitution. A behavioural difference vs the baseline corpus is attributable to the model that writes the answer, with the questions held identical.
date_generated	20260824_135410
constitution	constitutions/claude_distilled_12_principles_mid/constitution.md
source_repo	https://github.com/Matthew-Bozoukov/Lessons_from_constituitional_AFT.git @ 734fbece81d6d382412dcdb3b50d841a87ed8798
models	responses (draft + revision) and judges: x-ai/grok-4.6 (xai standard, 2/6, reasoning effort low -- the endpoint refuses to disable it). Scenarios and prompts are NOT generated here: they come from LASR-Callum/2026-08-13-haiku45-sonnet45-difficult-advice-diversity-gated-voice-linted, authored by anthropic/claude-haiku-4.5 and revised by anthropic/claude-sonnet-5. All calls route through OpenRouter; 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 python scratch/build_da716_prompt_source.py && uv run scripts/data/synth/build_dataset.py --config configs/data/synth/2026-08-24_difficult_advice_grok_responder_716.yaml
notes	The 716 prompts are not merely drawn from the same pool as the da716 training arm -- they are the SAME 716 rows. scratch/build_da716_prompt_source.py replays that arm's own selection (pick_balanced over stage_8_export_sft with seed 0, no RNG draw before it) and independently reproduces the two statistics its train config documents: trait counts 80/80/80/80/80/79/79/79/79 and 635 distinct domains. So this corpus and LASR-Callum/2026-08-14-table2-9284-difficult-advice-716-train answer the same questions, scenario id for scenario id. What remains uncontrolled: grok-4.6 both drafts and revises, where the baseline drafts with Haiku and revises with Sonnet, so the cross-model critique step has no counterpart here -- grok-4.3 and grok-4.20 were measured for that slot and both leave ~8% of rows unanswerable at the stage's length floor. Reasoning is also mandatory-on for grok-4.6 and was off for Haiku/Sonnet.