Kyle's updated goal: after the 20x sample “if it looks good after that, expand it to do the FULL set. keep monitoring the quality (with some random samples) and the budget hourly” and deliver a secret-gist HTML report + HTML Preview link when finished.
Conditional workstream: wait for issue #3's 2,000-unit generation, independent review, and RAG checkpoint; the assistant must inspect actual quality and failure causes before starting. If acceptable, cover the full eligible plan (10,829 single calls + 4,087 B2B call pairs, no B2C multi-call) using the adopted v9 prompt/quality gates. Reuse completed expanded source units and compatible request caches under verified identical configuration; do not count copied ledger costs twice. Preserve existing raw cohorts and exclusions. Quarantined/bad questions must not re-enter the active dataset merely through cache reuse.
Budget is $5,000 total for the expansion + full work, not per phase. Re-estimate remaining costs after the expanded run; preserve conservative generation, review, and RAG guards. Record unknown provider usage and separate turbopuffer/VM costs rather than presenting estimates as a complete invoice. Stop before spending beyond the authorized ceiling if the remaining work cannot fit without reducing the agreed quality protocol.
Hourly assistant inspections must read live process evidence, progress/costs, and random question/answer/source samples. A scripted monitor is supplemental. Final verification must cover generated artifacts, frozen independently reviewed QA, the RAG outcome set, reproducible rollups, and the rendered secret-gist report with its HTML Preview link. HF publication is not automatically implied by this goal.
Status: waiting for the expanded checkpoint; full generation has not started.
Kyle's updated goal: after the 20x sample “if it looks good after that, expand it to do the FULL set. keep monitoring the quality (with some random samples) and the budget hourly” and deliver a secret-gist HTML report + HTML Preview link when finished.
Conditional workstream: wait for issue #3's 2,000-unit generation, independent review, and RAG checkpoint; the assistant must inspect actual quality and failure causes before starting. If acceptable, cover the full eligible plan (10,829 single calls + 4,087 B2B call pairs, no B2C multi-call) using the adopted v9 prompt/quality gates. Reuse completed expanded source units and compatible request caches under verified identical configuration; do not count copied ledger costs twice. Preserve existing raw cohorts and exclusions. Quarantined/bad questions must not re-enter the active dataset merely through cache reuse.
Budget is $5,000 total for the expansion + full work, not per phase. Re-estimate remaining costs after the expanded run; preserve conservative generation, review, and RAG guards. Record unknown provider usage and separate turbopuffer/VM costs rather than presenting estimates as a complete invoice. Stop before spending beyond the authorized ceiling if the remaining work cannot fit without reducing the agreed quality protocol.
Hourly assistant inspections must read live process evidence, progress/costs, and random question/answer/source samples. A scripted monitor is supplemental. Final verification must cover generated artifacts, frozen independently reviewed QA, the RAG outcome set, reproducible rollups, and the rendered secret-gist report with its HTML Preview link. HF publication is not automatically implied by this goal.
Status: waiting for the expanded checkpoint; full generation has not started.