Everything below executes the real logic client-side in plain JavaScript (SHA-256 and HMAC implemented here and checked against known-answer vectors on load). The embedded numbers are the paper's measured and calculated outputs. What is measured: lossless codec ratios and keying cost on a 2-vCPU host with a 0.5B model. What is not claimed: any engine throughput, any device or cross-host result.
key_i = HMAC_K( "thox.kv.chunk/v1" || key_{i-1} || fingerprint || tokens_i as uint32 big-endian || len(extra) || extra ). Edit the chunks and watch the keys: the first differing chunk changes its key and every key after it (Propositions 1 and 2).
Weights plus KV reserve admit against usable RAM; priced workspace admits against the system budget. First-fit contiguous, anchor first; no partial admission. Floors: Air 2 GB, Mini 4 GB, Clip 6 GB (ThoxOS whitepaper v0.1.0). The default model is the 27.8B Q4 demo geometry; try the 2.49 GB Phi-4-mini file.
T_xfer = L k / (r B) + a; T_rec = L / P. The lossless ratio r defaults to the paper's measured value. Link and prefill rates are inputs, not measurements.
Standard LoRa time-on-air formula with a 16-symbol preamble and the 16-byte Meshtastic header. Presets from the Meshtastic radio settings table. Payload ceiling 239 bytes.