QOMN LAB — Extreme Conditions Test Harness HOSTILE

NASA / oil & gas style test harness. Survives, degrades gracefully, fails safe. Every suite runs against the live production API — no separate environment.
← Back to public benchmark  ·  desarrollador.xyz/benchmark2.html (LAB)
What this does. Fires hostile workloads at the live /api/plan/execute endpoint: adversarial inputs, boundary sweeps, parallel execution, fault injection. Scoreboard graded A+ through F. No client state is stored; reload resets everything.

■ MISSION READINESS SCOREBOARD

—
Run suites to compute mission readiness.
Each of the 6 suites contributes to the overall grade. All 6 must pass to reach A+.
TOTAL RUNS
0
PANICS
0
NaN OUTPUTS
0
DETERMINISM
—
AVG THROUGHPUT
—
WORST p99 (ms)
—
Executes all 6 suites sequentially. Takes ~45-90 seconds. Safe to abort by reloading.
—

● SUITE 1 — BLACK SWAMP

80% valid + 15% edge-case + 5% corrupt inputs. Measures panic rate, safe-rejection rate, NaN contamination.
valid_runs —
rejected_safely —
panics —
nan_outputs —
p99 latency — 

    


    
    

      
—

      
❄ SUITE 2 — DEEP FREEZE

      

        Subnormals, values near zero, catastrophic cancellation.
        Tests numerical stability at IEEE-754 boundaries.
      

      

        subnormal_stable —

        cancellation_ok  —

        overflow_safe    —

        underflow_safe   —

        worst_rel_err    —
      

      
      

      

    


    
    

      
—

      
🔥 SUITE 3 — REDLINE

      

        32 concurrent requests, same input.
        Measures throughput and bit-exact determinism under parallelism.
      

      

        concurrent    —

        all_success   —

        throughput    —

        bit_identical —

        p99 ms        —
      

      

      
      

    


    
    

      
—

      
👁 SUITE 4 — PHANTOM SENSOR

      

        Physically impossible inputs (neg flow, eff>1, zero diameter).
        Must reject with explicit reason, never silently compute.
      

      

        tested           —

        rejected_explicit —

        silent_accept    —

        physics_guard    —

        worst_violation  —
      

      
      

      

    


    
    

      
—

      
🧠 SUITE 5 — SPLIT BRAIN

      

        Same input executed across independent endpoints
        (/api/simulation/repeatability, /api/plan/execute).
        Must return bit-identical results.
      

      

        paths_tested   —

        cross_hash     —

        max_delta      —

        determinism    —

        runs_compared  —
      

      
      

      

    


    
    

      
—

      
⛔ SUITE 6 — MISSION ABORT

      

        Aborted requests (100ms timeout), recovery probe,
        rapid-fire bursts. Measures clean recovery, no orphan state.
      

      

        aborts_clean   —

        recovery_ok    —

        burst_handled  —

        post_health    —

        state_leak     —
      

      
      

      

    


  


  
  

    
■ BOUNDARY SWEEP — parametric stability scan

    

      Sweeps Q_gpm from 1 to 5000 in 50 steps against plan_pump_sizing.
      Looks for NaN, discontinuities, monotonicity violations (HP should strictly increase with flow).
    

    
    

    

    

  


  
  

    
■ GRADE RUBRIC

    

      A+ — all 6 suites pass, 0 panics, 0 NaN, bit-identical determinism

      A  — all 6 suites pass with minor degradation (<1% timeout)

      B  — 5/6 suites pass, no safety regressions

      C  — 4/6 suites pass or minor invariant failures

      D  — 3/6 or fewer, determinism degraded

      F  — any panic, NaN contamination, or silent physics violation