The runs behind the work: what we built, where it ran, and what came out.
168 of 168 held-out written-rule rows resolved by the compiled specialist in the sealed row-only run. The two general-model baselines resolved 72 and 45 rows with the same row-only prompt, and 167 and 163 when given the full rule.
49,140 distinct specialist weight hashes were placed across a 12-card Akida simulator layout: 4,095 per card, with 13 mission specialists. The run completed 72 of 72 cyber samples, 72 of 72 synthetic sensor rows and 24 of 24 written-command-rule cases.
In 20 of 20 stop trials, a six-step mission halted cleanly at step three.
Specialists were mapped through one-, two-, four- and twelve-card simulator layouts. Held-out routing completed 100 of 100 single-chain cases, and ambiguous inputs used the intended fallback in 16 of 16 cases.
An account-relation task built on 72 real OpenSSH event bursts and 3,559 held-out rows. The compiled four-bit specialist weighs 78,054 bytes.
Across 82,332 sensor windows, the UNO R4 plus Tesla T4 hybrid replay reached F1 0.919036 versus 0.895308 for the always-on GPU path, while making 53,489 rather than 82,332 GPU model calls: 35.0% fewer.
Official test replay with physical UNO R4 and Tesla T4; board decision mean 3.021 ms.
Read the ALF language, install ReasonRDN, or bring a workload for a side-by-side run.
Run yours with us →