Dayan public methods
Fifteen compact methods explain the decision rules shared across the public Skills.
For applied routes, start with the public playbooks.
For the broader candidate map, read Core knowledge.
- Complexity ladder — Prefer a deterministic rule, then a pipeline, then an agent. Escalate only when the simpler form cannot handle uncertainty or tool choice.
- Multi-pass decomposition — Separate framing, generation, evidence collection, verification, and release so one fluent output cannot silently approve itself.
- Adversarial validation — Use a checker with a different objective and, where possible, fewer permissions than the builder.
- Confidence and evidence chain — Tie every strong conclusion to observable evidence, scope, date, and a signal that would change it.
- Reversible action — Move quickly on bounded reversible work. Require explicit authority before publication, spending, credentials, deletion, signing, or production impact.
- Falsifiable problem framing — State a best-current problem hypothesis, plausible alternatives, and the real-world signal that would disprove it.
- Human authority and AI boundaries — AI may accelerate analysis and production; consequential authority, accountability, takeover, and release remain explicit human responsibilities.
- Six levels of completion — Distinguish artifact ready, internally verified, human accepted, reality released, external outcome verified, and economically validated.
- Project harness — Represent work as a dependency topology with waves, approvals, evidence receipts, stop conditions, and an archive that can be resumed.
- Delivery asset pipeline — A reusable delivery includes the artifact, deterministic checks, evidence, handoff, version, and a path to reuse or retire it.
- Simple altruistic communication — Lead with the decision, explain why it matters to this reader, show evidence, and end with the one useful action.
- Audience-clean artifacts — Reader-visible output contains only reader-relevant content; production notes, placeholders, internal roles, and hidden instructions live elsewhere.
- Tool-before-agent — If a failure can be detected deterministically, make a tool detect it before asking an agent to remember it.
- False-positive control — Treat polished output, fluent language, and structure checks as signals rather than proof of usefulness, safety, or truth.
- Human authority ledger — Write down which consequential actions remain human-controlled before the workflow starts.