Atome LM: a tiny ternary language model for microcontroller deployment

Atome LM v2 — SuperESP: the applied edge-AI layer on the Atome ternary engine.

• 11 applied on-device classifier heads + an on-device OS dispatcher (+ a regression head)
• Universal ESP32 installer (esp32 / s2 / s3 / c3 / c6 / h2; Xtensa + RISC-V) — no ESP-IDF needed
• Live-sensor agriculture firmware; bring-your-own-CSV "make your own head" loop
• Trust: Ed25519 attestation, load-time integrity check, tamper-evident audit log, signed model zoo

Verified (honest): 12/12 apps run on a physical ESP32-WROOM-32 (~27 KB state, 265 KB free heap), bit-exact Python↔C parity (~1e-6), SuperESP tests 34/34, Atome 146/146. Several demo heads ship on clearly-labelled synthetic data; voice KWS = 0.625 (modest, reported honestly). Apache-2.0.