File DA-008 View as Markdown

Verdict

{
  "sessionId": "s_9f2c…",
  "class": "agent",
  "probability": { "human": 0.04, "bot": 0.03, "agent": 0.93 },
  "confidence": 0.81,
  "agent": { "family": "openai", "verified": false, "method": "marker" },
  "reasons": [{ "code": "rhythm.think_then_act", "weight": 1.1 }],
  "recommendation": "tag",
  "profile": "ecommerce", "action": "checkout",
  "stage": "provisional", "model": "doubleagent-2026.09", "ts": 1758614400000
}

Probability says which class is most likely. Confidence says how much evidence it rests on.
recommendation is one of allow, tag, challenge, step_up, rate_limit or deny. We never enforce it.
Families: claude openai perplexity google browser_use browserbase skyvern manus amazon unknown.

Prefix	Evidence
`auto.` `global.` `marker.`	Automation and agent-browser artefacts
`env.`	Environment consistency
`drive.`	How input is driven
`rhythm.`	LLM think-time
`bio.` `human.`	Pointer and typing biometrics
`edge.` `ua.` `net.`	Header evidence
`verified.`	Server-checked identity only. See Declared cover.
`judge.`	The Handler

Properties and events

Property	Values
`da_class`	`human`, `bot`, `agent`
`da_agent`	Family, or `none`
`da_score`	0–100, non-human probability
`da_bucket`	`0-20`, `20-50`, `50-80`, `80-100`
`da_rec`	Recommendation
`da_verified`	`true` only after a server check returned a verified agent

Events: da_classified (once), then da_final (only if the verdict changed).

Default policy: tag every tool; hold back ad conversions when the class is bot with p ≥ 0.8; tag agents and never block them; skip destinations without consent.