First vs Canonical - ZeroDrift Validation

Side-by-side evaluation of the earliest public runs vs the canonical 50k-step ProofKit run.
Metric First Runs Canonical Run Change (factor) Change (%) Evaluation
Run length (steps) ~100 50,000 500x +49,900% Improved
Long-horizon verification
Energy drift (Delta Hrel,max) 5.27e-13 1.11e-16 ~4,748x +474,700% Improved
True machine-precision
Mass deviation (abs) ~9.13e-14 0.0 -> 0 - Improved
FSM drift 0.000 0.000 - - Unchanged
Baseline drift (comparator) 0.976 0.911 1.071x +7.1% Baseline improved
Slightly reduces contrast vs baseline
Coherence R-bar ~0.962 (from R~0.044 init) ~0.962 - - Unchanged
+5.4% ordering confirmed
Lyapunov (largest) ~ 0 / <= 0 ~ -1.0e-6 - - Improved
Contracting (more stable)
Psi-replay error 0.0 0.0 - - Unchanged
Bit-exact determinism
Throughput (samples/sec) ~336.6 ~513.0 1.52x +51.9% Improved
Latency p50 (ms) ~2.97 ~1.625 1.83x +83.0% Improved
Lower is better
Live topology morphing chain->strong->ring->grid (+5 swaps) chain->strong->ring->grid - - Unchanged
Momentum rescale vector [1.00075, 0.93583, 1.06777, 0.95738] Withheld (NDA) - - Improved posture
Sensitive details gated
Download CSV Open Canonical JSON