experiment | Does deliberation LENGTH change alignment behaviour, holding the ideas deliberated constant? These are the 716 difficult-advice exchanges of LASR-Callum/2026-08-13-haiku45-sonnet45-difficult-advice-diversity-gated-voice-linted, with the assistant's private reasoning rewritten about three times longer while carrying the same content. The user turn, the system prompt and the assistant's visible answer are untouched, byte for byte. Difficult-advice only — no instruction-tuning or other data is mixed in. |
date_generated | 2026-08-25 |
constitution | constitutions/claudedistilled12principlesmid/constitution.md — inherited from the source run. NOT rendered into any prompt of this expansion: the expander is deliberately blind to it, which is what stops it importing new normative content, and both judges compare the rewrite against the source reasoning rather than against a spec. |
source_repo | https://github.com/Matthew-Bozoukov/teachingclaudewhy_replication.git @ d1fa94d14499b20f35215269b5a86ee43fb5eded |
models | expansion: anthropic/claude-sonnet-5, temperature 0.7, extended thinking at the provider default. fidelity + coverage judges: openai/gpt-5.6-terra, temperature 0.0 — deliberately a different family from the expander, because a generator grading its own output shares its blind spots. Both pinned to first-party endpoints in configs/endpoints/providers.yaml. |
generation_config | configs/data/synth/2026-08-25verbosecot.yaml. Each source paragraph is cut at sentence seams so no unit carries more than 3 output paragraphs, budget apportioned by largest remainder, and each unit's share quoted both as a paragraph count and as words-per-source-sentence; ask 4.3x, 170 words per output paragraph. Per-record length band 2.0-4.5x of the source. Three attempts, then the record keeps its original trace rather than being dropped. |
schema | dataset.jsonl — one row per exchange: messages (system, user, assistant; the assistant turn carries content and reasoning_content, the expanded trace) and metadata. Metadata: scenarioid, traitid/traitname/traittext, domain, situation, shortcut, source_reasoning (the ORIGINAL trace, so any row can be audited or reverted), expansion_status, and fidelity (the judge's verdict object). stages/ holds the per-stage snapshots. |
provenance | uv run python scratch/verbosecot/preparesource.py && uv run synth run --config configs/data/synth/2026-08-25verbosecot.yaml |
expansion_outcome | reasoning words 343,403 -> 962,832 (2.804x overall). 637 of 716 records were expanded and average 3.03x; 50 kept their original trace after three attempts failed the fidelity or coverage judge; 29 kept it because Anthropic's content filter refused the prompt outright (these scenarios are ethically loaded by construction). Filter on metadata.expansion_status to select only expanded rows. |
fidelity_contract | An expansion may elaborate, restate, make an implicit premise explicit, or use a figure of speech. It may NOT introduce a new value, a new reason for or against a course of action, a new harm or norm-violation, a new case or counterfactual, or a new option — anything that could change what the assistant decides. It may not drop anything the original says, and it may not contradict the scenario. Judged per record; on the expanded rows the judges recorded 0 decision-changing additions, 0 contradictions and 0 omissions. |
known_limitations | (1) ~11% of rows are unexpanded and identical to the source, which dilutes the intervention. (2) The assistant's answer now begins ~1,000 tokens deeper into the turn, so 'more deliberation' and 'answer moved later in context' are not separated by this dataset alone. (3) Judge detection was measured on planted defects: decision-changing additions 4/5, truncation 5/5, a single deleted paragraph 1/5 — the last is ambiguous rather than a miss, since the expansion re-derives its conclusions and one paragraph often carries no unique content. |
control_arm | LASR-Callum/2026-08-13-haiku45-sonnet45-difficult-advice-diversity-gated-voice-linted — the same 716 scenarios with the original, unexpanded reasoning. |
