The Predictive Reach of a Decision Token: An Honest Map of Forecasting, Anticipation, and Complementary-Modality Fusion on a Single-Token Substrate

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The Predictive Reach of a Decision Token: An Honest Map of Forecasting, Anticipation, and Complementary-Modality Fusion on a Single-Token Substrate Randolph James Ferlic, M.D. and Kimberly Kate Ferlic (Fieldstone Analytics, LLC, Austin, TX, USA) Preprint · Zenodo DOI: 10.5281/zenodo.22736921 · CC-BY 4.0 · Community: spiral-domain-encoder-campaign Abstract A previously described class-discriminant codebook encoder reduces each window of a multivariate stream to a single 8-bit decision token — a compressed symbol chosen to preserve a downstream decision rather than to reconstruct the signal. Prior work established the token as a classification and detection primitive. Here we ask a different question, under pre-registered protocols with frozen outcome bands, placebos, and controls: how far does a decision token reach into prediction — forecasting, anticipation, and multi-sensor fusion — and where does it honestly fail? Four experiments on real public data map the answer. (1) Forecasting. Inserting a logarithmic-spiral front-end to forecast the continuous embedding is a NO-GO: a log-spiral is the exact trajectory of a linear dynamical system, so a plain vector-autoregression (VAR) already captures it, and on real embeddings the spiral extrapolation is numerically unstable or worse; the constructive spill-over is that a plain linear VAR on the embedding genuinely forecasts three of four streams (skill vs. persistence up to +0.30 at twelve steps), while the discrete token-transition model is a poor continuous forecaster and must not be repurposed as one. (2) Anticipation. Decision-state change can be anticipated above chance, and the cheap order-1 token-transition model is a strong, low-cost near-ceiling; engineered linear dynamics and order-2 history do not beat it, though a stronger nonlinear model combining the token-transition rate with cheap dynamics gives a modest, stream-dependent gain (a k-nearest-neighbor model beats order-1 by +0.04 on one of four streams, confidence interval excluding zero). (3) A precursor pattern, reported honestly. Qualitatively, anticipation is strong on monotone industrial degradation (near-failure AUC 0.972 on turbofan run-to-failure, novelty-vs-remaining-useful-life correlation −0.64), modest on gradual regime crossings, and weak on sudden onset (paroxysmal atrial fibrillation, ~0.65); but a pre-registered common-axis test did not confirm this as a quantitative law — a metric-agnostic precursor-strength measure did not predict anticipation quality (the sudden-onset dataset has the strongest measured precursor yet the weakest anticipation) — so we report the degradation early-warning result on its own and the cross-regime pattern as an observation, not a law. (4) Fusion. Single-token fusion of two genuinely complementary modalities helps (GO) when performed at the posterior level (per-modality token → per-modality posterior → product rule): +0.033 macro-AUC on twelve-lead ECG (limb vs. precordial leads) and +0.035 on robot force vs. torque, with bootstrap confidence intervals excluding zero; fusing both modalities into a single token instead does not help (the byte-wide bottleneck dilutes complementary information); the benefit scales with a measured complementarity that, estimated on training data, predicts the test gain (non-circular) and is robust to the fusion rule. Honestly, posterior fusion does not beat a full-feature model — not even one built on a single modality; its value is that within the extreme single-token regime the encoder targets, fusing two complementary tokens beats one. (5) Trustworthiness. Beyond accuracy, the one-byte decision is on average harder to adversarially flip than a full model (robustness ratio 1.25×, 1.9× on clinical ECG), though its posterior is somewhat less well-calibrated (an honest, recalibratable bound); and the fusion gain is predicted by the complementary label-information the two tokens carry (Spearman +0.68, with an identical-views null at the origin) — an information-theoretic account, not just a curve. The contribution is an honest, pre-registered map of a decision token's predictive reach and its trustworthiness: it forecasts through a linear reader, anticipates change up to a strong, cheap transition-model near-ceiling whose height is not explained by a simple precursor measure, fuses complementary sensors at the posterior level within its compression regime, and yields a one-byte decision that is mostly harder to adversarially subvert than a full model — each result reported with its boundary. Highlights · An honest, pre-registered map, not a single result — five experiments ask how far one 8-bit class-discriminant decision token reaches into prediction (forecasting, anticipation, fusion) and where it honestly fails; every hypothesis carried a frozen Go/No-Go band, placebos, and controls, and a second pre-registered hardening pass deliberately attacked the paper before writing. · Forecasting — a NO-GO with a mechanism — a logarithmic-spiral front-end is redundant with linear dynamics (a log-spiral is the exact trajectory of a linear system, so a vector-autoregression already captures it); the constructive spill-over is that a plain linear reader on the continuous embedding genuinely forecasts three of four streams (skill vs. persistence up to +0.30 at twelve steps), while the discrete token-transition model is a poor continuous forecaster and must not be repurposed as one. · Anticipation — a strong, cheap near-ceiling — decision-state change is anticipated above chance, and the encoder's own order-1 token-transition model is a strong low-cost baseline that engineered dynamics and order-2 history do not beat; a heavier nonlinear challenger beats it only modestly and stream-dependently (a k-nearest-neighbor model by +0.04 on one of four streams, CI excluding zero). · A precursor pattern, reported honestly — anticipation is strong on monotone industrial degradation (near-failure AUC 0.972 on turbofan run-to-failure, novelty-vs-remaining-useful-life correlation −0.64), modest on gradual crossings, and weak on sudden onset; but a pre-registered common-axis test did not confirm a quantitative precursor law (the sudden-onset dataset has the strongest measured precursor yet the weakest anticipation), so the cross-regime picture is reported as an observation and the degradation early-warning result stands on its own. · Complementary-modality fusion — a bounded GO — single-token fusion of two genuinely complementary modalities helps at the posterior level (per-modality token → per-modality posterior → product rule): +0.033 macro-AUC on twelve-lead ECG (limb vs. precordial) and +0.035 on robot force vs. torque, CIs excluding zero; cramming both modalities into one token does not help; the gain scales with a train-estimated complementarity (non-circular) and is robust to the fusion rule. Honestly, it does not beat a full-feature model even on one modality — the value is confined to the extreme single-token regime. · Trustworthiness of a one-byte decision — beyond accuracy: the token is on average harder to adversarially flip than a full model (robustness ratio 1.25×; 1.9× on clinical ECG), because crossing into a different-class Voronoi cell is farther than crossing a linear boundary — and the minimal flip distance is a closed-form, per-decision robustness certificate; its posterior is somewhat less well-calibrated (mean ECE 0.086 vs. 0.069), a gap that temperature scaling reduces but does not eliminate; and the fusion gain is predicted by the complementary label-information the two tokens carry (Spearman +0.68, with an identical-views null at the origin) — an information-theoretic account, not just a curve. · Corrections made under scrutiny, reported verbatim — the pre-registered hardening pass confirmed two claims (the linear forecaster; the fusion non-circularity and rule-robustness) and forced three inward (an anticipation 'ceiling' → near-ceiling; a 'precursor law' → qualitative observation; a 'near-full-model' fusion claim → a within-regime bound). The corrected claims, not the originals, are what the paper reports. What this record contains · Manuscript_Paper42.pdf — the manuscript (with the five figures embedded) and Manuscript_Paper42.docx, the editable source. · PAPER_42_ZENODO_ARCHIVE.zip — the reproducibility archive: the five frozen pre-registrations (including the reviewer-hardening and trustworthiness pre-registrations), the eight experiment runners (forecasting, anticipation, sharp-onset, the forecasting/anticipation hardening pass, complementary fusion, the fusion hardening pass, the trustworthiness trilogy, and the temperature-scaling recalibration), the five figure-rebuild scripts, the per-experiment result JSON files, and the five figures, with a README. All paths and identifiers are scrubbed and leak-scanned per the campaign deposit discipline; no raw benchmark data is redistributed (all datasets are public; sources and a DATA_ROOT convention are in the archive README). Cite as R. J. Ferlic and K. K. Ferlic, "The predictive reach of a decision token: an honest map of forecasting, anticipation, and complementary-modality fusion on a single-token substrate," Zenodo, 2026, doi: 10.5281/zenodo.22736921. License and patent notice Released under the Creative Commons Attribution 4.0 International License (CC-BY 4.0). Consistent with that license, no patent, patent-application, or other intellectual-property right of the authors is licensed, waived, granted, or otherwise conveyed by this deposit. This work characterizes previously described methods in prediction settings and discloses no new algorithmic subject matter; the complementary-modality posterior-fusion result is a reduction to practice of a previously filed method. The methods described — including the class-discriminant single-token codebook encoder, its nearest-centroid-distance novelty/degradatio