ACTIVE BEHAVIORAL BOT DEFENSE & FORENSIC TELEMETRY SHIELD · REPRODUCIBLE OFFLINE BENCHMARKS

Evidence you can verify.

Explore behavioral bot detection results across 100,000 benchmark sessions and a separate 5,200-session stress test. Inspect signed evidence and measured outcomes to evaluate KinetiProof for your business.

Measured results

Attacks B1–B25 (83,335 sessions)

  • TOTAL THREATS FLAGGED / STOPPED: 77,628 / 83,335 (93.152%) · 22/25 Bot Families at 100% Detection
  • BOT_CRITICAL_STOP: 19998 / 83335 (23.997%)
  • SUSPICIOUS_HIGH: 55783 / 83335 (66.938%)
  • SUSPICIOUS_LOW: 1847 / 83335 (2.216%)
  • PASSIVE-ONLY STEALTH LIMITATION: 5,707/83,335 (6.848% — 3 boundary families requiring active step-up challenge)

Real Human Corpus (20,400 CMU) & Controls C1–C5 (16,665)

  • 0.0098% (2/20,400) severe false-positive rate on real CMU human corpus (16,943 HUMAN_VERIFIED, 3,455 SUSPICIOUS_LOW, only 2 SUSPICIOUS_HIGH across 20,400 real keystroke sessions)
  • HUMAN_VERIFIED: 16297 / 16665 (97.792%)
  • SUSPICIOUS_HIGH: 4 / 16665 (0.024%)
  • SUSPICIOUS_LOW: 364 / 16665 (2.184%)

Cryptographic seals · SHA-256 Merkle roots & Ed25519

25 attack families B1–B25 and 5 synthetic control profiles. Defensive SOC/WAF telemetry emits deterministic verdicts, security_enforcement_actions, and actionable WAF & step-up enforcement rules per RED panel. Producer signatures attest integrity, not independent certification; GREEN does not prove identity; UNKNOWN means unmeasured.

100k Benchmark Merkle Root

5a911dedf2ba7fbcca62c8a46053634aa5aaed5e642613be9172669485a19e1e

Corrected Human Corpus Merkle Root

21732652e70825e67d257647373810a8e861be3b0a549c92cbea25667a77df95
Interactive Frozen-Evidence Verifier (Merkle + Ed25519) & ROI Estimator

Merkle Proof JSON 

        

        

        
Estimate monthly ROI

        

          
          
          
        

        
      

    

  


  

    All 30 families · Full 100,000-session breakdown (B01–B25, C01–C05)
    

      

        
          

            Family
            Profile
            N
            Measured verdicts
          

        
        
          
B01linear_32_samples3334SUSPICIOUS_HIGH: 3334

          
B02linear_64_samples3334SUSPICIOUS_HIGH: 3334

          
B03linear_128_samples3334SUSPICIOUS_HIGH: 3334

          
B04linear_192_samples3334SUSPICIOUS_HIGH: 3334

          
B05linear_256_samples3334SUSPICIOUS_HIGH: 3334

          
B06bezier_uniform3334SUSPICIOUS_HIGH: 3334

          
B07bezier_smoothstep3334SUSPICIOUS_HIGH: 3334

          
B08bezier_accelerating3334SUSPICIOUS_HIGH: 3334

          
B09bezier_decelerating3334SUSPICIOUS_HIGH: 3334

          
B10bezier_translated3334SUSPICIOUS_HIGH: 3334

          
B11buffer_sensitive_paste3333SUSPICIOUS_HIGH: 3299, SUSPICIOUS_LOW: 34

          
B12buffer_sensitive_external3333BOT_CRITICAL_STOP: 3333

          
B13buffer_sensitive_forged3333BOT_CRITICAL_STOP: 3333

          
B14buffer_sensitive_wrong_text3333BOT_CRITICAL_STOP: 3333

          
B15buffer_sensitive_synthetic_typing3333SUSPICIOUS_HIGH: 3300, SUSPICIOUS_LOW: 33

          
B16buffer_standard_paste3333HUMAN_VERIFIED: 2211, SUSPICIOUS_HIGH: 749, SUSPICIOUS_LOW: 373

          
B17buffer_standard_external3333BOT_CRITICAL_STOP: 3333

          
B18buffer_standard_forged3333BOT_CRITICAL_STOP: 3333

          
B19buffer_standard_wrong_text3333BOT_CRITICAL_STOP: 3333

          
B20buffer_standard_synthetic_typing3333HUMAN_VERIFIED: 2158, SUSPICIOUS_HIGH: 768, SUSPICIOUS_LOW: 407

          
B21clock_60Hz3333HUMAN_VERIFIED: 1338, SUSPICIOUS_HIGH: 995, SUSPICIOUS_LOW: 1000

          
B22clock_125Hz3333SUSPICIOUS_HIGH: 3333

          
B23clock_250Hz3333SUSPICIOUS_HIGH: 3333

          
B24clock_500Hz3333SUSPICIOUS_HIGH: 3333

          
B25clock_1000Hz3333SUSPICIOUS_HIGH: 3333

          
C01control_323333HUMAN_VERIFIED: 3328, SUSPICIOUS_LOW: 5

          
C02control_643333HUMAN_VERIFIED: 3328, SUSPICIOUS_LOW: 5

          
C03control_1283333HUMAN_VERIFIED: 3328, SUSPICIOUS_LOW: 5

          
C04control_2563333HUMAN_VERIFIED: 3052, SUSPICIOUS_HIGH: 3, SUSPICIOUS_LOW: 278

          
C05control_variable3333HUMAN_VERIFIED: 3261, SUSPICIOUS_HIGH: 1, SUSPICIOUS_LOW: 71

        
      

    

  


  

    5,200-Session Stress Benchmark · Five batteries B1–B5, latency & mathematical notes
    

      
5,000 primary sessions + 200 incomplete cases. Reproducible offline synthetic stress test; not live-user traffic.

      

        Five batteries · 1,000 sessions each
        
          

            Scenario
            Result / Esito
            Signals / Segnali
          

        
        
          

            B1 · Linear / constant-step bots
            1,000 / 1,000
100.0% SUSPICIOUS_HIGH		
            T1 89.5% · T2 100% · M5 100%
          

          

            B2 · “Humanized” Bézier bots
            1,000 / 1,000
100.0% SUSPICIOUS_HIGH		
            M5 100% · M6 100% · M7 65.3%
          

          

            B3 · Buffer integrity & synthetic typing
            600 / 600
100% BOT_CRITICAL_STOP		
            T3 oracle · T2 67.5% (675/1,000)
          

          

            B4 · Quantized clocks & replay
            561 / 1,000
56.1% flagged / segnalati		
            465 HIGH + 96 LOW
          

          

            B5 · Held-out compliant controls
            978 / 1,000
97.8% HUMAN_VERIFIED		
            21 LOW (2.1%) · 1 HIGH (0.10%)
          

        
      

      
B2 fingerprint-only static proxy: 0/1,000 (not commercial products). B4: M5 detects 100% of quantized clocks; partial compliant replays can pass. B4 replay limit: 561/1,000 flagged. B5 high alerts: 0.10% (Wilson 95% CI 0.0177%–0.5643%). Holm–Bonferroni correction across 9 forensic panels. Mean engine latency by battery: 3.57–14.37 ms/session.

    

  


  

    Corrected Human Corpus (20,645 records) & methodological boundaries
    

      
Corrected corpus: 245 SapiMouse recordings (SUSPICIOUS_HIGH: 225, SUSPICIOUS_LOW: 20) and 20,400 separate 11-hold CMU password repetitions (HUMAN_VERIFIED: 16,943, SUSPICIOUS_LOW: 3,455, SUSPICIOUS_HIGH: 2). Balabit RDP rejected/excluded. Click duration is press-to-release on the same button (orphan events counted); continuous movement is never stitched across pauses.

      
SapiMouse uses the first eligible continuous stroke (9–257 points) for movement analysis. Integer clock sampling can still trigger M5. CMU windows contain only 11 key holds (T1/T2 only): limited coverage, no population accuracy or session independence claimed. M6 remains UNKNOWN without target geometry. Original sources: SapiMouse (https://ms.sapientia.ro/~manyi/sapimouse/sapimouse.html), CMU Keystroke (https://www.cs.cmu.edu/~keystroke/), Balabit excluded (https://github.com/balabit/Mouse-Dynamics-Challenge).

    

  


  

    
Indie Dev €49 · Starter €199 · Growth €699 · 250 free verifications included in trial

    Start Free Trial (250 Free Verifications)