The Decision Token as a Semantic Channel Code: A Frozen One-Byte Decision Primitive for Goal-Oriented Communication, Digital-Twin Synchronization, and Bit-Exact On-Device Decoding

The Decision Token as a Semantic Channel Code: A Frozen One-Byte Decision Primitive for Goal-Oriented Communication, Digital-Twin Synchronization, and Bit-Exact On-Device Decoding Randolph James Ferlic, M.D. and Kimberly Kate Ferlic — Fieldstone Analytics, LLC, Austin, TX, USA Preprint · Zenodo DOI: 10.5281/zenodo.24020142 · CC-BY 4.0 · Community: spiral-domain-encoder-campaign · This work is a scientific characterization of previously filed and published methods; it discloses no new mechanism and is accompanied by no new patent application — see the notice below. Abstract Goal-oriented ("semantic") communication transmits only the information a receiver needs to act, rather than reconstructing the source waveform. This record studies a deliberately minimal realization of that idea and pushes it as hard as an adversarial communications reviewer and a hard-nosed technical evaluator would: a frozen, deterministic, ≤256-cell decision token — an offline-trained nearest-centroid codebook over a collapsed linear-discriminant/principal-component embedding, read out by a per-cell class posterior — treated as a semantic channel code that carries a single byte across a noisy link. Unlike the learned, heavy, non-deterministic encoders that dominate the field, this encoder is bit-exact, sub-milliwatt, and auditable by construction; the scientific question is how far such a primitive can be pushed as a communication object, and exactly where it stops. Across a pre-registered battery of twenty-one channel tiers (SC-1 … SC-21) and four twin-synchronization tiers (TW-1 … TW-4) on the public UCR/UEA archives plus real activity and physiological streams, with decision bands fixed before measurement and negatives kept, six things are established. (1) The thesis holds at archive scale: transmitting the decision beats reconstruct-then-decide at a matched ~6-bit budget by +0.161 mean macro-AUROC across the full 108-dataset census (≥+0.05 on 87%, tied-or-ahead on 95%; every point above the diagonal). (2) A channel-realism and streaming map: a naked token is graceful to ~1% bit-error-rate and fragile by 10%; a soft-decision lookup readout recovers much of that fragility over a real BPSK/AWGN channel (+0.048 on 8/10 at mid-SNR); and a causal, zero-latency streaming filter recovers most channel-induced error on genuinely autocorrelated streams (+0.199 at 10% BER on dwell streams; 81% of the non-causal smoother's recovery at zero latency), with a shuffle control separating a true temporal effect from a class-prior effect. (3) An honest head-to-head with a learned joint source–channel code (Deep-JSCC), which beats the token on accuracy by +0.076 on 10/10 datasets — relocating the token's value from accuracy to determinism, bit-exactness, footprint, and auditability. (4) A security boundary: the token's geometry is not a tamper-evidence primitive (a codebook-aware adversary flips the decision with a median of one bit; cheap geometric detectors fall to 0.31 AUROC against a detector-aware attacker), so integrity must come from a system-level authentication layer — a boundary quantified rather than papered over. (5) A digital-twin synchronization flagship: a one-byte token's control decision is unmatched by scalar-quantized full-state transmission even at 64 bits on 10/10 datasets, and a one-byte "deviation" side-channel becomes a safe competence gate under distribution shift (open-set detection 0.76 vs. 0.52 for the posterior margin; up to 0.94), while the previously disclosed drift compensator recovers a twin's own sensor drift (translational ~99%, gain ~83%) but cannot bridge two genuinely different entities without recommissioning. (6) A bit-exact reduction to practice of the receiver decode on two commodity Arm microcontrollers (0 of 400 decision flips; ~5.2 µs/symbol on a Cortex-M33 and ~13.9 µs/symbol on a Cortex-M4F; ~0.3% of the encode energy). Five pre-registered directional hypotheses were refuted and are reported as such. The study introduces no new mechanism: it re-expresses previously disclosed methods in the communication and twin-synchronization settings, and positions the "transmit-the-decision" framing itself as prior art to be cited, not claimed. The contribution is a precise, hardened, hardware-validated map of what a frozen one-byte primitive can and cannot do as a communication object — strong precisely because it names its own boundaries. Highlights · One byte, transmitted as the decision, beats reconstructing the waveform — on an entire archive. At a matched ~6-bit budget the token out-decides a reconstruct-first baseline that scalar-quantizes the same features to the same bits, by +0.161 mean macro-AUROC across all 108 UCR/UEA datasets (≥+0.05 on 87%, tied-or-ahead on 95%), with every census point above the diagonal. The mechanism is transparent: the token spends every bit on the decision-relevant partition, while scalar quantization spends its bits reconstructing a waveform no one at the receiver will read. · Channel fragility is real, but remediable, free, and at zero latency. A naked token degrades gracefully to ~1% bit-error-rate and becomes fragile by 10% (one flipped bit jumps to an unrelated cell). Over a real BPSK/AWGN channel a textbook soft-decision lookup readout recovers much of that fragility (+0.048 on 8/10 at mid-SNR), and a hidden-Markov streaming filter recovers channel error at no extra channel bits (+0.199 at 10% BER on balanced-dwell streams, +0.173 across the census). Crucially, a causal, lag-0 forward filter keeps 81% of the non-causal smoother's recovery at zero latency — a clean latency dial (81/88/98% at lag 0/1/5) — so the recovery is deployable in a real-time loop, not an artifact of peeking at future symbols. · The recovery is honestly bounded, and proven temporal by a shuffle control. On a balanced 6-class activity stream the gain is genuinely temporal (+0.202, with a clean shuffle control at −0.106); on normal-class-dominated cardiac streams the gain survives shuffling and is therefore a class-prior effect, not dwell tracking — a pre-registered hypothesis (that making the stream multi-class would restore temporal recovery) that was refuted and kept. The condition for temporal recovery is balanced multi-class dwell, not merely a multi-class stream. · A learned coder beats the token on accuracy — and we ran the comparison against ourselves. At matched channel use and SNR a small Deep-JSCC autoencoder trained end-to-end over the channel beats the frozen token by +0.076 mean on all ten datasets, and its frontier dominates throughout. The token does not win on accuracy and does not claim to; its value is the trade it makes — a frozen, bit-exact, deterministic, auditable lookup with no per-deployment training and no float decoder. This is stated as the honest cost of determinism. · Security is a system property, not a token claim. A codebook-aware adversary flips the decoded decision with a median of one bit (near-tautological for any short unauthenticated code, and decision-aware index assignment provides essentially no worst-case defense). The sharper, often-skipped measurement: a decision-flip is largely undetectable by the receiver's cheap geometric checks against a detector-aware attacker (confidence-check AUROC 0.31; distance-consistency 0.50, because distance-to-centroid is cell-invariant). Integrity therefore belongs at the protocol layer — a message-authentication-code for an active adversary, plus parity/CRC for channel noise — and the token's geometry supplies none of it. · Digital-twin synchronization is the flagship: compact *and* safe. A 6-bit token's control decision is unmatched by scalar-quantized full-state transmission even at 64 bits on all ten datasets and the real activity stream — the full-state baseline never catches the token in the swept range — so a twin-to-twin link's bandwidth and energy budget collapse by roughly an order of magnitude with decision quality improving. And under open-set distribution shift (held-out classes the receiving twin has never seen) the posterior margin is at or below chance at flagging them while the one-byte distance "deviation" byte detects them on 8/8 datasets (mean +0.243 AUROC, up to 0.94): the receiver defers rather than acting on a confident-wrong decision — precisely where a naive confidence signal is dangerous. · The drift compensator synchronizes a twin to itself, not to another entity. Under injected drift a shared codebook is not free (decision drops up to 0.5–0.8). A previously disclosed online mean-tracking compensator recovers a translational sensor-bias shift by ~99% and most of a realistic sensor-gain shift (~83%, because standardization renders gain largely translational), but ~0% of a pure non-translational embedding rotation — and, decisively, it does not bridge the gap between two genuinely different entities (cross-subject recovery ≈ 0). Bridging distinct entities is a per-entity recommissioning problem, reported explicitly rather than implied away. · The receiver decode is deterministic silicon, and effectively free. The causal HMM filter plus soft-LUT readout over a 10%-BER activity stream runs bit-exact against the host float32 reference — 0 of 400 decision flips on both cores (a Cortex-M33 at 128 MHz and a Cortex-M4F at 64 MHz), reproducing all temporal corrections exactly, at ~5.2 µs/symbol (M33) and ~13.9 µs/symbol (M4F) — about 0.2–0.4% of the encoder window and ~0.3% of the encode energy. The machinery that makes the token safe and recoverable costs the receiver almost nothing. · Five honest refutations make the positives trustworthy. Pre-registered directional guesses that came back against us — bursty errors are not worse than i.i.d. (concentration spares other tokens); multi-class alone does not restore temporal recovery (balanced dwell does); token relevance does not saturate by K=64 (it keeps rising to K=256; K=64 captures ~79%); automat