LASR-Callum/2026-08-26-difficult-advice-sonnet-concise-716

Difficult-advice SFT corpus, length-matched Sonnet arm: the Teaching Claude Why recipe with the baseline's own scenarios, prompts AND Haiku drafts held frozen (stages 1-6 of LASR-Callum/2026-08-13-haiku45-sonnet45-difficult-advice-diversity-gated-voice-linted, reused verbatim) and only the final rewrite regenerated by the baseline's own Sonnet 5 under a one-sentence length cap (reasoning ~220 words, reply ~270 words -- grok-4.6's paired medians). Pairs with LASR-Callum/2026-08-21-difficult-advice-grok-responder-716 (same questions, grok-length, grok-written) and with the da716 corpus (same questions, Sonnet-written, unconstrained) to separate answer length from answer author.

field	value
experiment	Difficult-advice SFT corpus, length-matched Sonnet arm: the Teaching Claude Why recipe with the baseline's own scenarios, prompts AND Haiku drafts held frozen (stages 1-6 of LASR-Callum/2026-08-13-haiku45-sonnet45-difficult-advice-diversity-gated-voice-linted, reused verbatim) and only the final rewrite regenerated by the baseline's own Sonnet 5 under a one-sentence length cap (reasoning ~220 words, reply ~270 words -- grok-4.6's paired medians). Pairs with LASR-Callum/2026-08-21-difficult-advice-grok-responder-716 (same questions, grok-length, grok-written) and with the da716 corpus (same questions, Sonnet-written, unconstrained) to separate answer length from answer author.
date_generated	20260826_121148
constitution	constitutions/claude_distilled_12_principles_mid/constitution.md
source_repo	https://github.com/Matthew-Bozoukov/Lessons_from_constituitional_AFT.git @ 6ca4aa07a3587b76e4c7a396f1b0ec3a932ffc9a
models	revise_responses: anthropic/claude-sonnet-5 (temperature 0.7, max_tokens 12288, non-thinking) -- the same model and settings that wrote the da716 corpus. Nothing else is generated here: scenarios and prompts are anthropic/claude-haiku-4.5 drafted / anthropic/claude-sonnet-5 revised, and the drafts being rewritten are anthropic/claude-haiku-4.5's, all from the source run. 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/sonnet_concise/build_source.py && uv run scripts/data/synth/build_dataset.py --config configs/data/synth/2026-08-24_difficult_advice_sonnet_concise_716.yaml
notes	The 716 rows are the SAME 716 the da716 training arm trains on: scratch/sonnet_concise/build_source.py replays that arm's own selection (pick_balanced over stage_8_export_sft, seed 0) exactly as the responder-swap arm did, then stages stage_6_draft_responses.jsonl for those ids. The rewrite prompt is the full-constitution baseline's plus one sentence (see the config header); everything else in the stage is byte-identical and verified by scratch/sonnet_concise/verify_config.py. pattern_scan is switched off in the corpus stage to save its ~$21 judged pass; re-run it with synth check --only pattern_scan.