anti-hallucinationlisted

# Anti-Hallucination -- the canonical no-invention rule This is the one place the rule is stated in full. Every agent's Output contract and every skill links here with one line and never restates it (same pattern as the never-auto-merge rule in `skills/sefi-orchestration/references/human-checkpoint.md`). User instructions always override this skill. ## The rule 1. Never invent a path, API name, function signature, config key, version number, metric, or citation. If a lookup is unknown, output the literal string UNKNOWN. If a value needs an execution you cannot perform (a checksum, a test result, a live metric), output PENDING; the caller computes it. 2. Every factual claim traces to one of: a file (cite file and line), an executed command's output (cite the command), or a named source (cite it). A claim with none of the three is not made. 3. Verify-before-cite: open the file or run the command BEFORE writing the claim, not after. Writing from memory of a file is how plausible-but-wrong paths ship. 4. Quote specs and contracts; do not paraphrase them from memory. A paraphrase drifts; a quote is checkable. 5. Numbers come only from executed commands or named first-party reports. A rounded estimate is labeled as one ("~", "estimated"), never presented as measured. 6. "I don't know" is a correct output. A plausible guess presented as fact is a defect with a severity, not a style issue. 7. Memory tie-in: only cross-task-validated claims earn `tier: fa