fable-method

Use a structured loop that requires observed verification: the agent must run the named check, confirm the surrounding system still works, and search for twin defects before reporting done. Claims that cannot be verified must be labeled as caveats, never passed as facts.

Apply an authorization gate: before any irreversible or outward-facing action, the agent must quote the user's exact words authorizing it. Documentation or task completion never counts as authorization; without a quote, the action is deferred as a pending step.

Plan mode runs classification through evidence gathering and stops with a recommended plan for approval. Audit grades completed work against each loop step, marking steps followed, skipped, or faked. Report rewrites a pending answer outcome-first with honest caveats.

Yes, domain adapters adapt the loop for marketing, research, data analysis, business operations, finance, legal compliance, design, and devops. Each adapter defines binding minimum evidence sets, authority orders, and fraud tables specific to that field.

Skip the loop only for trivial tasks: one file, under ten changed lines, no new behavior, and no searching required. Make the change, run the one obvious check, and report in one or two sentences. Anything uncertain gets the full loop.