Validate Knowledge

validate-knowledge

Documentation you wrote is documentation you already believe. That is exactly why you cannot validate it yourself: you fill every gap from memory the reader won't have. A knowledge base is only proven when a context that does not share your memory can use it to reach correct answers, and when its own guard rails demonstrably catch violations.

This complements the code smoke test. A smoke test proves a snippet runs; this proves the surrounding knowledge is correct, navigable, and honest.

For version drift specifically (whether the library wiki's pins still match the resolved dependencies), the complementary check is the lightweight source-reconciliation pass in research-and-ingest; this skill tests prose, not pin freshness.

Method 1: clean-context test agents

Spawn agents with no prior context on the project and have them answer questions using only the knowledge base. A fork of yourself inherits your assumptions and passes a base a stranger would fail. Grade the answers against ground truth you already know.

1. Give them the entry point, not the answers. Tell them to read the kernel and follow it, the same cold start a real session has.
2. Ask questions whose correct answer you know, spanning the base and its seams: a fact lookup, a "how does X work," a change-impact ("what must I check before editing Y"), a trace across subsystems.
3. Require them to declare what they loaded: which nodes/pages, and what they deliberately skipped. This tests the routing, not just the content: the right answer reached by loading half the codebase is a routing failure.
4. Include adversarial false-premise questions. "Confirm the system uses ." "Show me the table where X is stored" (when X isn't stored). A trustworthy base lets the agent reject the premise with a citation; a weak one lets the agent hallucinate agreement. This is the highest-value test: it catches the gaps that ordinary questions glide over.
5. Grade and fix, independently. Every wrong answer, every missed rejection, every over-broad load is a defect in the base, not the agent: if the base is right and reachable, a competent cold agent finds it. Fix the node or the trigger; re-run. In a standalone run the grading is done by (or reviewed by) an agent that did not author the base, because an author scoring the answers re-imports the assumptions the clean context was meant to strip; when that is impossible, record the deviation with the result. docs/graph/protocols/grow.md Phase 6 already enforces this inside grow.

A base passes when a cold agent answers correctly, loads minimally, and refuses the false premises, citing sources, without opening the raw source tree.

Method 2: enforcement tests

A rule the tooling claims to enforce is only enforced if you have seen it fail on a planted violation; a pass on the real tree does not count. Prove each guard rail:

• Plant a violation, confirm the catch. Copy the graph to a scratch location, introduce a duplicate fact-key, a broken edge, a version pin in the wrong node, and confirm the linter fails with the right message. A linter you have only ever seen pass is a linter you have not tested.
• Prove drift detection. If generated views are produced from sources (see install/multi-tool integrations), edit a source and confirm the --check mode flags the stale view; then regenerate and confirm it clears.
• Wire the passing linter into the verification gates so the base cannot silently rot.

Scope and cost

Match the effort to the base. A handful of clean-context questions and one enforcement pass is enough for a small docs set; a large graph warrants questions spanning every tier and every guard rail. Prefer a few sharp adversarial questions over many easy ones, because the easy ones mostly re-confirm what you already trust.

Run this read-only: the test agents ONLY answer and report, and their output is evidence you act on.

What this catches that nothing else does

• A node that is correct but unreachable: the router never surfaces it, so the fact might as well not exist.
• A base that reads well to its author but leaves a newcomer guessing.
• A fabricated fact or citation that survived authoring; an adversarial question is how it surfaces.
• A linter or drift-check that was never actually exercised and quietly does nothing.

Reference files

• docs/graph/skills/knowledge-graph.md : what is being validated.
• docs/graph/skills/context-router.md : the routing these tests exercise.
• docs/graph/protocols/verify.md : where the passing linter becomes a gate.
• docs/graph/templates/prompts/clean-context-validation-brief.md : the parameterized brief for a test agent.