The engine, on camera

Screen recordings of the shipped v1.0.2 binaries running live. No mockups, no animations — every number and every trace line on screen was produced by the engine while recording.

Workbench tour

43 s · Windows 11 · shipped ruledsl_capi.dll · bindings/python/workbench.py

Rules on the left, inputs on the right: compile & run, the decision with its output fields and canonical decision hash, the engine's own evaluation trace (why this decision), a live input change, and the deterministic MISSING_NOW_UTC_MS error when the explicit clock is removed — the engine never reads the system clock. The "Run 100×" check re-evaluates and asserts every run produced the identical decision hash.

Run it yourself: pip install ruledsl then ruledsl-workbench --dll path/to/bundle/bin/ruledsl_capi.dll — the engine DLL comes from the SDK bundle. It also replays production bytecode: File > Open Bytecode (Replay) loads an .axbc exactly as your server ran it.

Windows ↔ Ubuntu — byte-identical decisions

40 s · Windows 11 + Ubuntu 24.04 (WSLg) · both on shipped v1.0.2 binaries

The same ruleset and the same input, loaded, edited and executed simultaneously on two operating systems. The banner at the bottom is a live instrument, not a caption: the Ubuntu side exports its decision hash after every run and the Windows side compares — green means the hashes are equal; a genuine mismatch would print red. Traces match line for line, and a live input change produces a new decision whose hash is again identical on both systems.

Throughput — live counter, single thread

13.5 s · 12th Gen Intel Core i5-1235U (laptop-class) · C API, in-process, no server

The counter counts real ax_eval_bytecode() calls streamed from a running benchmark — hundreds of thousands of evaluations in five seconds on one laptop thread, with zero decision mismatches across the entire run. The number on screen is that run's live measurement, recorded as it happened.

How to read these numbers. One machine, one workload, single thread — throughput varies with hardware, ruleset and load; measure on your own machine (the bundle ships everything needed). What does not vary is the decision: same bytecode, same input, same options → the same decision bytes, on every run and on both platforms shown. That claim is not a video — it is published, recomputable evidence.