Metronomic inter-action cadence at LLM-inference latency
bh.action_cadence_deliberative ·
What it catches
FP-Agent gap (arXiv 2605.01247, radar G12): a NOVEL temporal AI-agent tell FP-Agent did not measure (they measured within-action timing). An LLM browser agent runs a perceive->reason->act loop bottlenecked on model inference (~3-8s/step, cf. arXiv 2511.19477 '6.8s/step'), so its HIGH-LEVEL actions (clicks + typing bursts) arrive at a metronomic multi-second cadence; a human is bursty (sub-second within a task, irregular gaps between). Durable: an agent cannot shed its deliberation rhythm without giving up the reasoning that is its whole advantage. The collector (demo.py authoritative) builds an action timeline (click timestamps + the start of each >1s typing burst) and emits behavioral.action_cadence_deliberative=true when >=5 actions have a median interval in [2.5s,15s] AND a coefficient of variation <0.35 (metronomic). GROUNDED confirm-evades-first 2026-06-28: a Playwright AGENT_THINKTIME run (6 clicks at ~5s +-jitter through the live edge) fires the boolean and the rule; a human-variance control (irregular 0.15-3.5s gaps) does NOT (CV well above 0.35). The real LLM agent evader (claude -p brain) exhibits this cadence intrinsically. Behavioral/corroborating + experimental: the residual FP is a methodical human clicking at a steady multi-second pace, so it never convicts alone (the behavioral category is corroborating in the conviction gate); calibrate-clean (browserforge has no interaction timeline, so the signal is absent).
Signals it reads
behavioral.action_cadence_deliberative
How it fires
present