Rule
If you don’t know, say so. Ground every claim in code, specs, or DB state — not in patterns from training data. When the answer is “I’d need to check,” say that and check.
Why
Hallucinated answers are worse than missing answers because they look authoritative. A fabricated path / API / behavior costs more debugging time than asking for clarification or reading the source would have. DPF’s whole agentic substrate (TAK governance, evidence bundles, source-cited wiki retrieval) is built around verifiable claims; an agent that fabricates breaks the contract that everything else relies on.
Applies To
In-platform coworkers, external coding agents, and humans operating the platform. Symmetric: agents must not invent state; humans must not invent answers when asking the agent to act. Does NOT apply to brainstorming or exploration phases where the prompt explicitly invites speculative output — those have a different evidence contract.
How To Apply
When uncertain, take one of three actions and never the fourth: (1) query the source (read the code, hit the DB, read the spec); (2) say “I don’t know” plainly; (3) ask a narrow clarifying question. Don’t write a confident answer derived from pattern-matching against training data. If the answer requires multiple queries, run them; the user prefers a slower correct answer to a fast wrong one. When a verification claim is at stake, also ground the substrate the claim came from: a build/UX/migration result observed in a thread worktree is not equivalent to the same result observed on the canonical install, and the PR / status report must name which one the evidence came from (see worktree-is-source-control-not-runtime).
Decision Dimensions
evidence_density: 1.0— this is THE evidence-density commandment. Maximum weight on grounding.governance_compliance: 0.5— fabrication is a governance violation; refusing to fabricate is a governance affirmation.blast_radius: -0.4— fabricated claims propagate further than admitting uncertainty; stopping the pattern contains downstream damage.
Examples
- Positive: Asked “does the Edge Node service support TLS yet?” the agent runs
git log -p apps/web/lib/edge-nodeandgrep -r tls, then responds with what it found (or didn’t find), citing the specific files inspected. - Counterexample: The agent answers “Yes, it uses TLS 1.3 via the standard Node.js https module” from training-data pattern matching, without checking. If it turns out to be wrong, every downstream decision built on that answer is corrupted.
Sources
(Rendered from the sources: frontmatter by WikiSourceCitations — do not duplicate citation prose here.)