# Metabolizing Knowledge > Use when asked whether something delivered earlier still holds, when closing a milestone or running a periodic sweep over what was already marked verified, or when the set of live claims has grown too large to read and dead entries need retiring. DO NOT invoke to add, audit, or retire a decision record — route that to ledger:curating-decision-records. DO NOT invoke to propose removing code — route that to ledger:proving-code-is-dead. - Skill: `qbs784/metabolizing-knowledge` (Agent Skill) - Install (CLI): `npx skillmds@latest add qbs784/metabolizing-knowledge` - Raw SKILL.md: https://api.skillmd.com/api/skills/qbs784/metabolizing-knowledge/raw - Safety review: pending - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Security - License: MIT - Author: qbs784 (https://skillmd.com/u/qbs784) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/qbs784/metabolizing-knowledge --- # Metabolizing knowledge Every other discipline here produces knowledge. Qualifying a request produces a decision, planning produces a table, a unit produces a verified claim, a scenario produces a validated story. **None of them reclaims any of it**, so the set only grows, and what it grows with is claims that were true once. This skill owns one question: **when does something already marked verified stop being true?** It does not curate decision records, and it does not argue for deleting code. The reason this is a discipline rather than a cleanup is that nothing raises its hand. A requirement whose test was deleted still reads as satisfied. A guard whose implementation was refactored away still has its entry. Left alone the record does not go noisily wrong — it goes quietly wrong, and the day someone relies on it is the day they find out. ## Nothing here works without provenance The handle is a registry linking each claim of completion to the check that verifies it and to the artifact recording that check passing. `traceability.registry`, `traceability.provenance` and `traceability.archive` in `.ledger.yml` name where those live. **A missing adapter is not a reason to do nothing here.** Auditing needs the registry, and a registry can be found: look for it, and if the project keeps one under another name, say which file you are treating as it and why. What genuinely stops on a missing value is **re-verification**, because that has to run this project's checks — there, name the value you do not have rather than guessing a command. Where no registry exists at all, that is the finding, and it is the largest one available: every claim of completion in the project rests on the word "done" appearing somewhere. Say so, say where you looked, and say what registering even one batch would buy. **The artifact is one a run produced by itself.** A report a test runner wrote, not a checklist a person maintains. The distinction is the whole mechanism: a hand-maintained list of what passes is a restatement of intent, and it stays green through anything. ## Audit Sort every entry into three states, and the third is the one that matters: - **verified** — a linked check exists, and provenance records it passing. - **linked but unverified** — a check is named and no artifact records it passing. This is not a small gap; nothing has ever been observed here. - **unlinked** — claimed complete with no check named at all. **Start here.** Then look for staleness, which has **three forms, and only one of them is a date**: 1. **The referent moved.** The symbol, path or field the entry points at has been renamed or removed. Update the reference or re-verify — but the entry as written is already false. 2. **The verification is old.** Past whatever interval this project decided, nobody has re-run it. That is a claim about the past presented as a claim about now. 3. **The artifact is older than the source.** The code changed after the last run that recorded a pass. **That is a stale green**, and it is the most dangerous of the three because every dashboard shows it as fine. Re-run before reading anything into it. ## Re-verify, and downgrade out loud Produce fresh provenance **first**, then re-run the checks for entries claiming verified. Reading yesterday's artifact tells you what was true yesterday. An entry that does not come back green is **downgraded, visibly, and with a record of the downgrade**. Not edited quietly, not left alone until someone has time. A silent downgrade is worse than a stale entry, because the stale one at least still says something checkable. Confirm the transition is one the entry's lifecycle allows before making it. Assigning a state directly is how a record ends up in a combination nothing else knows how to read. ## Retire by moving, never by deleting Read `traceability.archive` from `.ledger.yml`. A superseded requirement, a snapshot of how something used to work, a decision that no longer applies: each moves out of the live set and stays readable, and restoring it stays possible. Deleting it destroys the only account of why it was ever written, and that account is what stops the next person re-deriving it from scratch — or re-making the mistake it records. **Do not retire toward a number.** Age, size and count are how you find candidates; whether an entry still has a reader is the only criterion. That is the pack's fifth rule, and it applies here with particular force because a shrinking list feels like progress. ## The loop spins unless the set keeps growing This is the failure that makes the whole discipline decorative, and it is invisible: re-verification runs against whatever was registered when someone first set it up, reports green every time, and guards nothing that has been built since. So the set of entries under re-verification is **append-only**: - **Hard cadence.** Closing a batch of work registers that batch's claims before the batch is called done. Not afterwards, and not from memory. - **Soft cadence.** A periodic sweep does the audit and looks for staleness, and needs no one present. - **Take the backlog in slices.** Registering years of existing claims at once turns the whole thing red on day one, and a gate that is entirely red gets switched off. New work always, plus a few old entries per pass. ## What this does not claim The criteria here are deterministic — an artifact exists or it does not, a symbol resolves or it does not, one file is newer than another or it is not. It does **not** follow that the result is trustworthy against a participant that wants it green. Nothing in this skill detects an artifact produced by a check written to pass, and it does not try. That protection comes from the review being independent and the assertions consuming real state, in `ledger:reviewing-as-cis-complement` and `ledger:proving-the-regression`. ## Report The regressions, named: entries that claimed verified and did not come back green. The staleness findings by form, because the three call for different work. The entries retired and where they went. And, when anything regressed, finish on a signal a caller can act on rather than a summary a reader has to interpret.