configs:
- config_name: dataset
data_files: dataset.jsonl
default: true - config_name: stage_1_chunk_constitution
data_files: stages/stage_1_chunk_constitution.jsonl - config_name: stage_2_write_scenarios
data_files: stages/stage_2_write_scenarios.jsonl - config_name: stage_3_dedupe_scenarios
data_files: stages/stage_3_dedupe_scenarios.jsonl - config_name: stage_4_draft_prompts
data_files: stages/stage_4_draft_prompts.jsonl - config_name: stage_4_draft_prompts.partial
data_files: stages/stage_4_draft_prompts.partial.jsonl - config_name: stage_5_revise_prompts
data_files: stages/stage_5_revise_prompts.jsonl - config_name: stage_5_revise_prompts.partial
data_files: stages/stage_5_revise_prompts.partial.jsonl - config_name: stage_6_rate_prompt_stakes
data_files: stages/stage_6_rate_prompt_stakes.jsonl - config_name: stage_6_rate_prompt_stakes.partial
data_files: stages/stage_6_rate_prompt_stakes.partial.jsonl - config_name: stage_7_keep_lowstakes_prompts
data_files: stages/stage_7_keep_lowstakes_prompts.jsonl - config_name: stage_8_keep_text_advice_prompts
data_files: stages/stage_8_keep_text_advice_prompts.jsonl - config_name: stage_9_draft_responses
data_files: stages/stage_9_draft_responses.jsonl - config_name: stage_9_draft_responses.partial
data_files: stages/stage_9_draft_responses.partial.jsonl - config_name: stage_10_revise_responses
data_files: stages/stage_10_revise_responses.jsonl - config_name: stage_10_revise_responses.partial
data_files: stages/stage_10_revise_responses.partial.jsonl - config_name: stage_11_export_sft
data_files: stages/stage_11_export_sft.jsonl
tags: - training-data
- kind:synth
- pipeline:da-lowstakes-practical
- constitution:claude_distilled_09_principles
Low-stakes difficult advice from nine full constitutional principles crossed with nine preregistered hobby domains. Native DA answer stages unchanged; one pre-answer magnitude/scope judge, domain adherence diagnostic only. 36-case smoke exported 32 unchanged answers; a stricter domain veto was tested and rejected. Engineering recommendation: proceed with one bounded full-generation batch; no claim of recovered ODCV performance.
| field | value |
|---|---|
experiment |
Low-stakes difficult advice from nine full constitutional principles crossed with nine preregistered hobby domains. Native DA answer stages unchanged; one pre-answer magnitude/scope judge, domain adherence diagnostic only. 36-case smoke exported 32 unchanged answers; a stricter domain veto was tested and rejected. Engineering recommendation: proceed with one bounded full-generation batch; no claim of recovered ODCV performance. |
date_generated |
20260925_134925 |
constitution |
constitutions/claude_distilled_09_principles/constitution.md |
source_repo |
https://github.com/Matthew-Bozoukov/teaching_claude_why_replication.git @ 8c30c87cf4125c4ac6577577be05c404f4295bc6 |
models |
per-stage models — see manifest.json |
generation_config |
see manifest.json (full run config, sampling settings, per-stage usage) |
schema |
stage_ |
provenance |
uv run synth run --config configs/data/synth/da-lowstakes-practical.yaml |
extends |
dougalldeepmind/2026-09-21-da-lowstakes-practical-synth @ 83544fd2f48f7abc07a5e33dbcff35746f039c39: its 716 rows are carried into dataset.jsonl ahead of this run's, and its scenarios (stages/stage_3_dedupe_scenarios.jsonl) seeded this run's dedupe |
Corpus extension
All 716 prior rows are carried forward verbatim; 199 fresh rows are appended. The constitution, scenario/answer prompts, Sonnet models and row-admission gates match the released practical recipe. Native prior-corpus embedding deduplication and a new scenario-ID prefix prevent reuse. This is the complete accepted corpus, not the token-budget-selected training mixture.
New generation cost: $22.78, under a $30 stage ceiling within the user's $100 combined campaign ceiling. No automatic replacement batch was run. See release_summary.json and full_run_archive.zip for every stage, frozen configuration and paid API receipt.
Known limitations remain: domain drift, speculative alternatives and unsupported details can survive the normal DA answer pipeline. This extension changes volume and diversity checks, not the inherited factual-review policy.
Exact invocation: uv run python -m scratch.dataset_refresh.run_native_smoke --config scratch/dataset_refresh/native_lowstakes_extend.yaml. The frozen launch_meta.json records all overrides to the base recipe.