eproc · Sequential evaluation lab

Loading C11 WASM (72 KiB)…

Check a stopping rule before you put it in your evaluation loop. Paste bounded scores or run a preset to see when the evidence crosses your threshold. All calculations run in your browser.

Python release wheelInstall eproc 0.1.0
Single-Header C11 / C++17 (0 malloc)#include "eproc_single.h"
WebAssembly / JS / TS (0 Emscripten)initEproc('./eproc.wasm')
MCP 2026-07-28 Server (7 tools)eproc mcp
Evaluate a stream of scores
For the upper-tail test, the null bounds the expected next score given the history. Set the score bounds, threshold and error level before collecting data.
EProcess Verdict (Ville)
—
e-value / anytime p
Confidence sequence
—
Empirical-Bernstein interval
Conformal 90% / 95% Radii
—
5-rate crossing-free PAVA
C11 WASM Execution
—
640 B wire digest
Click "Run C11 WASM Engine" or pick a preset to run EProcess + ConfidenceSequence + MultiRateConformal + 640-byte ConformalMonoid locally in WebAssembly.
Repeated z-test rejection rate
—
Continuous peeking
E-process rejection rate
—
Across 200 simulated streams
Median evaluations used
—
Up to 600 per stream
Mean evaluations saved
—
Relative to 600 calls per stream
Sample Log E-Process Wealth Trajectories (25 live paths, threshold = ln(1/α)) H0: μ ≤ 0.50 · T = 600 ticks
Verified Kernel Speedups vs Pure-Python Baselines [LIVE-CPU]
Intel Xeon 2.30 GHz (AVX2+FMA) · Python 3.12 · Median of 5 paired alternating runs · benchmarks/receipts.json
Engineering Optimization Ladder
1. Native CPython C-API (_native)56× faster
eprocess_scalar: 6,629 ns (ctypes) → 118.5 ns · 0 deps · 0.88 ms cold import
2. Runtime CPUID AVX2 in Portable Wheels2.56× faster
Portable wheel conformal_5rate: 799.6 ns (scalar) → 311.9 ns with 0 -mavx2 flags & 0-ULP identity
3. Exact Gaussian Max-Entropy B&B161× faster
sift_voi_2048ch: 2,005 µs → 12.4 µs · 8,192-ch firehose 0.752 F1 vs 0.100 naive (0 loud FA)
4. Slim Seqlock + Fixed-Share BOA58× faster
SafeFuser C11: 566 ns → 445.7 ns (2.24M ops/s) · Attack MSE 0.0220 (-88.7% vs anchor, -48.8% vs median)
Naive 8-Agent MSE
—
SafeFuser Attack MSE
—
Naive Echo Posterior
—
EchoPool Exact Posterior
—
Rolling Absolute Error (log10 scale): Naive Mean vs Anchor vs eproc SafeFuser (Attack t=150, Regime Shift t=300)