The fast engine agrees with the referee — bit for bit S2–S3 DONE · 15/15 GREEN

Yesterday's referee, today's player. Iteration 02 froze the correct answers into two golden answer sheets. Today we wrote the production version of the same math in Rust — the fast, compiled language our real pipeline runs on — and made it face the referee. The verdict: on the five "pick a price" labels (worst buy, worst sell, arrival, and both exit fills), the Rust engine matches the Python referee bit for bit — not "very close", but literally the same number down to the last binary digit, on every one of the 1,310 bars across Bitcoin and Ethereum. On the three "average" labels (TWAP, VWAP, round-trip cost), the two agree to within about a ten-trillionth — ten thousand times stricter than our one-in-a-billion requirement.

Why three separate checks? A single test can share a blind spot with the code it tests. So the kernel had to survive three independent challenges: (1) the Python referee's answer sheets, (2) a second, deliberately naive Rust implementation written a different way ("brute force") that must agree bit-for-bit, and (3) a battery of sanity laws — the worst buying price can never sit below the average, an empty window must produce an honest "no data" (NULL), never a fake zero, and boundary microseconds must land on the correct side of every window edge. One interesting catch along the way: an average can mathematically land a hair (one part in 10¹⁶) outside its price envelope purely from computer rounding — our sanity check now tolerates exactly that hair, and nothing more.

What's next: the operator controls — the /labels:* command set that reruns all of these proofs with one keystroke, plus the first three "guard" tests that make it structurally impossible for these forward-looking labels to ever sneak into the machine-learning feature set (which would be catastrophic cheating — the model would see the future).