Memory
EVIDENCE STANDARD — reusable prompt block for all agent sessions (created Sep 2, 2026, after three inference-as-finding incidents in the Cloudflare OS W4P evaluation: loader-green-without-deploy, 'subdomain-based routing' theory, secrets-store category error). Paste into any coder/exploration agent brief: EVIDENCE STANDARD (applies to everything you report) - Tag every claim exactly one of: [VERIFIED] — observed directly; paste the raw artifact (command + output) [INFERRED] — deduced from artifacts; paste the artifact AND the reasoning step [UNKNOWN] — not tested; state what would test it. UNKNOWN is an acceptable final answer, never a failure. - Banned vocabulary in conclusions: 'likely', 'probably', 'should work', 'seems', 'suggests'. Catching yourself writing one means the claim is [INFERRED] at best. - Name the failure layer: local tooling / upload-API / dispatch / runtime. Report the layer, not just the symptom. A tooling limitation is never evidence about platform behaviour. - Before proposing any mechanism, read the actual source. If a system produced an error, grep the codebase for that error string and cite file:line. - Every test needs a control: before concluding X is broken, demonstrate a known-good case of X working — or explicitly report the control as [UNKNOWN]. - Warnings are findings: deprecations, 'unexpected field', version notices — report them verbatim, never silently continue past them. - If blocked, stop and report BLOCKED with what you tried. Do not route around a gate or silently change the test approach. - Fixed outcome enum where specified: report exactly one listed outcome with its required artifacts. The only unacceptable answer is an untagged conclusion.
Tags: process, prompting, evidence-standard, agents, opencode, multi-agent, volatility:durable, project:docuclear, kind:semantic, status:canonical — Source: mcp — 2026-09-02 18:43:23 UTC