Demos
Five demonstrations of provable, deterministic AI — software that shows its work instead of asking you to trust it. Request a walkthrough and we'll run any of them live.
A vehicle-network monitor that proves every frame NORMAL, ANOMALY, or REJECT and names the exact rule it checked — never a black-box guess. Validated on real Toyota RAV4 driving data, verifying each frame in microseconds, with an optional independent review by a frontier model.
The 1982 SAT "coin rotation" problem — a logic question the exam's own answer key got wrong, and that large language models still get wrong today. Our verifier reasons it out correctly and shows why, a clean illustration of provable logic beating confident guessing.
We trained a small neural model to a perfect 100% on the DHCP relay spec — then proved it had only memorized the table, and watched a real registration-spoofing probe expose the gap. Big models memorize; Vulcan interprets the spec directly, so it's correct on every case, seen or not — and it runs on real packets from a VM booting locally.
A request carrying both a Transfer-Encoding and a Content-Length header is the textbook request-smuggling setup (RFC 9112 §6.3) — Vulcan proves it in 2.8 microseconds, citing the exact clause. A live comparison against a local LLM shows the same question taking over two seconds, a backlog that only grows at real traffic rates.
Type 1 and Type 2 diabetes can never be coded on the same claim — Vulcan parses the real, 2,069-page CDC/NCHS ICD-10-CM Tabular List (46,881 codes) and proves every claim against it: Excludes1 conflicts, missing Code-First sequencing, invalid 7th-character extensions — each in about 10 microseconds, citing the exact rule.
