Skip to content

Library processing status

The inventory and extraction snapshot was produced on 2026-09-10 from the iCloud Hacking directory; the distillation queue counts below are the last committed checkpoint on 2026-09-11. The queue may advance while a run is active; use scripts/distill_status.py for a live count. This is a processing report, not a claim that the course has absorbed every paragraph.

Measure Result
Supported source paths 503
PDF paths 303
EPUB paths 2
Markdown paths 76
Plain-text paths 122
Ready paths 502
Empty iCloud placeholder paths 1
Distinct ready content hashes 487
Extracted content-addressed bundles 485
Extraction failures requiring review 2
Eligible 4,000-character distillation chunks (dry run) 44,965
Eligible five-book AD pilot chunks (dry run) 2,386
Current chunk drafts staged (proposed, not approved) 1,190
AD pilot chunks staged 1,190
AD pilot chunks without output 1,196
AD pilot chunks not yet attempted 841
AD pilot chunks quarantined for retry 355
Distillation chunks generated and approved 0

The two failures are retained as explicit failures in the local run output and are not silently counted as empty books. The source is not copied into GitHub. Re-run inventory and extraction after iCloud sync completes; unchanged content reuses its local artifact cache. The previous curriculum, audit, and SQLite-ledger experiments are excluded from this report’s source set.

The next distillation stage converts extracted sections into original course concepts with source references, then requires human review before publishing canonical notes. The dry-run chunk count is intentionally recorded so an agent can measure queue completion; it does not mean the course has absorbed the library. 1,190 current chunk drafts exist in ignored staging output, all from the AD pilot, with zero approved distillation chunks; 355 pilot responses remain quarantined for quality, host permission, or transient Codex failures and 841 pilot chunks have not yet been attempted. Older legacy staging files are excluded. A high extraction count is not the same thing as a high-quality course.

Run uv run python scripts/check_student_readiness.py --status-file pilot-status.json --evaluation-file evaluation.json --approved-notes N --practice-days N --graph-rebuild-verified --approved-path before calling the course student-ready. The command exits non-zero until the lesson contract, completed pilot, human canonical review, empirical quality metrics, graph rebuild, approved path, and seven days of practice evidence all pass. It reports failed gates explicitly and never treats generated staging drafts as approved notes.

Use uv run python scripts/distill_batch.py --limit 10 to process small batches through the host Codex CLI with gpt-5.6-luna. Keep batches small enough to inspect failures and prompt drift before widening the run. LM Studio is not used for generation; start it only for batched embeddings. If Codex cannot resolve the requested model, record the failure and stop rather than substituting a chat model.