Multi-step gate flow paced at machine-regular intervals (near-zero inter-step CV — a timer, not a human)
bh.session_flow_robotic · convicts
What it catches
Session-intent coherence — the machine-REGULARITY tell, the blue counter to a PACED bot (the arms-race move after bh.session_flow_superhuman, which a bot evades by adding think-time between gates). Over the per-ks_sid flow log, when the inter-step gaps' coefficient of variation (stddev/mean) across >= _FLOW_REGULAR_MIN_STEPS (5) completions is below _FLOW_REGULAR_CV (0.03) AND the mean gap is above the physiological floor (a PACED flow, not a superhuman one — the floor tell owns those), the detector injects behavioral.session_flow_robotic. No human's reaction+think-time is that regular: even a metronomic human clicking at a steady pace carries reaction-time variance (CV ~0.05-0.1), so 0.03 sits below the human floor — FP-SAFE BY CONSTRUCTION. Catches the NAIVE fixed-sleep humanization tier (a sleep(N) loop between steps); a JITTERED humanizer (variable think-time, CV well above 0.03) evades it — that residual is a RED WIN recorded on the cadence sub-axis (the deeper counter, the think-time DISTRIBUTION shape, is external-calibration-bound). Convicting (w0.8). GROUNDED LIVE 2026-07-04: a fixed-pace arena-solver (KS_PACE_MS, constant inter-step sleep) trips session_flow_robotic while NOT tripping the floor tell (paced, not superhuman); the jittered humanizer (KS_HUMANIZE) and a synthesized diverse-human population (including a metronomic human at CV ~0.06) are ALL silent. See docs/research-radar.md (session-intent axis).
Signals it reads
behavioral.session_flow_robotic
How it fires
present
Bypassed by 10
Frontier evaders that reach the detector uncaught (scored only suspicious, defeating every convicting tell) — this check is not one that stops them. The red-team frontier this detection still has to convict.