Source: frameworks/kit-knowledge-transfer/04-quality.md

04 — QUALITY: Knowledge Transfer

Format: two gates. Gate 1 is mechanical and binary (engine/qc.py — the build is gated on it; you cannot ship a fail). Gate 2 is a judgment checklist run by Claude + the advisor. When to run: after every Create or Update, before the capture is shared.

Why two gates, not 100-point scoring: the mechanical checks (gaps flagged, nothing fabricated, structure intact) are disqualifying — a binary gate is correct. The judgment checks (is the captured tacit knowledge actually the knowledge?) can't be mechanized — a checklist is correct.

Gate 1 — Mechanical (binary, blocking · `engine/qc.py`)

Run: py engine/qc.py ".docx". Must exit 0. It enforces:

#	Check	Blocks ship
1	Every template section is present in the output	Yes
2	No marker leaks — no `[...]`, no `FILL::`, no `FLAG::` in the rendered text	Yes
3	Every gap and every risk is highlighted (highlight count == FILL + FLAG count)	Yes
4	Aptos font (matches the source-doc standard)	Yes
5	Filename matches `- Capture .docx`	Yes

Gate 1 is the anti-fabrication floor: a blank can't go un-flagged, and a flag can't go un-highlighted.

Gate 2 — Judgment (checklist · Claude + advisor)

Mechanical pass ≠ good capture. Verify by hand:

Substantiation

Pick 5 lines at random — each traces to a specific moment in the source. Zero lines you can't source.
No invented values dressed as fact. Anything uncertain is a FILL:: flag, not a confident guess.

The knowledge is actually captured

Watch-outs carry judgment, not restated steps. At least the non-obvious calls a successor would otherwise learn the hard way are present.
The single points of failure are flagged (FLAG:: RISK —), not buried.
A competent successor who never met the expert could run one cycle from this doc + ask only the flagged questions.

Voice + boundary

Plain operator language — no internal jargon, no AI-team names, no builder-speak (see 02 forbidden terms).
If vault-visible: fully anonymized. No real client, person, or account number. Fictional names only.
Alphabetized Rules + Standing-notes lists.

Ship Criteria

Gate 1 exited 0, and every Gate 2 item is true, and the advisor has reviewed. The advisor is the final gate — not the QC pass.

Common Failure Modes

Failure	What Happens	How to Fix
Steps captured, judgment missed	The doc reads like an SOP; the successor still can't make the calls	Re-mine the source for the watch-outs — the "this always…" / "don't do X before Y" moments. If the source doesn't have them, that's a facilitation gap (`06`), not a writing gap.
Blank filled with a guess	A plausible value that isn't in the source slips in	Replace with `FILL::` . The gate can't catch a plausible fabrication — Gate 2 substantiation must.
Real client in a vault example	Anonymization boundary violated	Swap to fictional names/data. Never lift a real subject into a vault-visible capture.
Gate gamed by removing flags	Someone deletes a `FILL::` to make the page "clean"	The gap didn't close — the flag did. Restore it; an unflagged unknown is the exact failure this kit exists to prevent.
One capture, many subjects	A single doc tries to hold several people/processes	One capture per subject/process. Run the kit again.
Process content under a config heading	Ordered steps (1, 2, 3…) render under "Rules — alphabetized"	Pick the template shape that matches: `template-process.json` ("Sequence — run in order") for a method, `template.json` ("Rules — alphabetized") for standing config. See `05`, Two capture shapes.