AD and lateral-movement pilot

The first source pilot is deliberately limited to five hashed books. The list gives Luna a stable scope for extraction and distillation; it does not grant permission to publish their text or to treat a filename as a reviewed concept.

Source	Why it is in the pilot
Windows Internals Part 1	process, memory, and system foundations
Windows Internals Part 2	security architecture and operating-system boundaries
Windows Security Internals with PowerShell	observable identity and policy behavior
Windows Security Internals	Windows security mechanisms and evidence
Hands-On Penetration Testing on Windows	authorized assessment workflow and reporting context

The exact relative paths and SHA-256 revisions are in examples/pilot-books.json. Re-run inventory before processing; a changed hash is a new source revision and must be reviewed. A renamed file with the same hash is the same content for extraction purposes.

Process the pilot

Start with a dry run to measure the queue:

uv run python scripts/distill_batch.py \
  --input .deadwire/extractions \
  --source-manifest examples/pilot-books.json \
  --dry-run --limit 100000

Inspect progress at any time without contacting LM Studio:

uv run python scripts/distill_status.py --source-manifest examples/pilot-books.json

Then run bounded batches through the host Codex CLI after the model preflight succeeds. LM Studio is reserved for later embedding batches:

uv run python scripts/distill_batch.py \
  --input .deadwire/extractions \
  --source-manifest examples/pilot-books.json \
  --limit 10

Every output is a proposed chunk draft. Human review must validate claims, commands, provenance, and safety before a concept can enter the canonical notes or student path.