Commissioning Without Labels: a Frozen Single-Token Codebook's Label-Hunger is the Projection, not the Naming — and a Few-Label Screen Says When Unsupervised Structure Suffices

Commissioning Without Labels: a Frozen Single-Token Codebook's Label-Hunger is the Projection, not the Naming — and a Few-Label Screen Says When Unsupervised Structure Suffices Randolph James Ferlic, M.D. and Kimberly Kate Ferlic — Fieldstone Analytics, LLC, Austin, TX, USA Preprint · Zenodo DOI: 10.5281/zenodo.23775800 · CC-BY 4.0 · Community: spiral-domain-encoder-campaign · This record characterizes a property of a previously filed and published method (the class-discriminant single-token codebook and its deployment-sizing law); it discloses no new patentable subject matter. Abstract A frozen, deterministic single-token encoder reduces a window of inputs, by a fixed computation, to one 8-bit token indexing a bounded ≤256-cell codebook and a per-cell lookup; its runtime is bit-exact and runs in microseconds and sub-milliwatts. A prior program has characterized the encoder's reach and its deployment-sizing law, with one standing weakness: the encoder is label-hungry at commissioning time, because the codebook is built by two supervised steps — a Fisher-linear-discriminant projection and a majority-label per-cell lookup. This record asks, with a pre-registration and an exhaustive hardening battery, whether the codebook can be commissioned with few or no human labels, keeping the frozen runtime untouched, and maps a bounded answer. Pure label-free commissioning — cluster the commission set into pseudo-classes, fit the projection and the lookup from those — does not recover the supervised codebook in general (battery-mean gap +0.118 AUROC across 23 real multivariate datasets; +0.131 on the separable subset, far above the pre-registered band). It is free where the dominant unsupervised structure coincides with the decision axis (cwru4/cwru10, EMG, Epilepsy — one anchor per class suffices) and collapses where it does not (mfpt3, ECG, RacketSports, HAR), including at high local purity. The label-hunger is specifically the projection, not the naming (naming saturates at ≤4 anchors per class, leaving a ~+0.10 geometry floor), and it is clustering-independent (k-means, Gaussian-mixture, agglomerative agree). The program's separability screen (kNN-purity) fails to forecast where label-free works (its bootstrap 95% CI on the recovery correlation straddles zero), but a few-label cluster–class Adjusted Rand Index — from the handful of anchors already spent naming — is a significantly better, calibrated, stable, and out-of-sample-generalizing (pooled Spearman 0.84, CI [0.51, 0.97]) ranking screen that flags the high-purity collapses kNN-purity misses. A few anchors used passively cannot close the geometry floor (iterative self-refinement, a coalesce fitting the discriminant from a few anchors, shrinkage, over-clustering, and active selection all fail); a constrained (anchor-pinned) semi-supervised clustering recovers near-supervised accuracy at a modest budget (within 0.017 AUROC at 64 anchors per class), but a go/no-go gain test proves this is a *label-budget* effect, not a mechanism advantage — plain few-shot supervised on the same labels matches or beats it on six of seven large-n datasets. The synthesis is a deployment recipe: commission label-free where the few-label ARI screen clears (recalibrated per domain, validated on a small holdout), and fall back to supervised otherwise. The record discloses no new patentable subject matter. Highlights · Pure label-free commissioning is bounded. Battery-mean supervised-minus-label-free gap +0.118 AUROC (recovers ~41% of the supervised above-chance margin); free where unsupervised structure is the decision axis, collapses where it is not — including on high-kNN-purity datasets (mfpt3 0.936, ECG 0.960), so local purity is not the condition. · The label-hunger is the projection, not the naming. Cluster-naming saturates at ≤4 anchors per class; a ~+0.10 geometry floor (the Fisher-LDA projection) remains, unchanged by fair bijective naming. Clustering-independent across k-means, Gaussian-mixture, and agglomerative pseudo-labelers. · A few-label screen says where it works. kNN-purity is not even significantly correlated with recovery (bootstrap CI [−0.03, 0.82]); a few-label cluster–class Adjusted Rand Index is (CI [0.48, 0.95]), significantly better (paired difference CI [0.002, 0.71]), calibrated, stable, generalizes out-of-sample (pooled Spearman 0.84), and fixes the high-purity collapses. · The semi-supervised limit is a label budget, not a mechanism. Constrained anchor-pinned clustering recovers near-supervised at a modest budget (within 0.017 AUROC at 64 anchors/class), but a pre-registered go/no-go gain test shows few-shot supervised on the same labels matches or beats it on 6/7 large-n datasets — no prior-art-distinct gain; no new subject matter. · A deployable recipe with its negatives drawn. Cluster label-free → name with ~1 anchor/class → measure the few-label ARI → recalibrate and holdout-validate → commission label-free where it clears, else supervised. The transferable lesson The frozen token's value is its cheapness, determinism, and auditability; this record extends that premise to the decision to commission it without labels — a choice that can be made from a handful of labels, cheaply and before committing. The methodological lesson is the discipline of a go/no-go test: the one candidate mechanism recovered near-supervised accuracy and tripped a pre-registered intellectual-property gate, and a non-obviousness test against a few-shot-supervised baseline then showed the recovery is a label-budget effect, not an advance. A map with its negatives drawn — the bounded recovery, the failed variance screen, five failed label-efficient levers, and the gain-test clearing — is worth more than an over-claimed positive. What this record contains · Manuscript_Paper67.pdf / .docx — the manuscript (comprehensive; six figures; the pre-registrations, the main battery, the eight hardening tiers, the semi-supervised limit, the go/no-go gain test, the screen, the regime map, prior work, and the deployment recipe). · PAPER_67_ZENODO_ARCHIVE.zip — the reproducibility archive: the frozen pre-registrations, the campaign runners (the main battery, the eight hardening tiers, the coalesce probe, and the go/no-go gain sim) plus the program's shared machinery, the per-probe result records (JSON), the six figures and their builder, and a README. Public data are not redistributed; absolute paths and handles are scrubbed. Cite as Ferlic, R. J., & Ferlic, K. K. (2026). Commissioning Without Labels: a Frozen Single-Token Codebook's Label-Hunger is the Projection, not the Naming. Zenodo. https://doi.org/10.5281/zenodo.23775800 License and patent notice Released under CC-BY 4.0. Consistent with that license, the deposit licenses the copyright in the text and figures only, and no patent or other intellectual-property right of the authors is licensed, waived, or conveyed by this deposit. This record discloses no new patentable subject matter. It characterizes a property — the extent to which, and the regime in which, a previously filed and published method (the class-discriminant single-token codebook and its deployment-sizing law) can be commissioned with few or no labels. Pseudo-labeling, cluster-then-label, Hungarian assignment, the Adjusted Rand Index, Gaussian mixtures, agglomerative clustering, seeded and constrained (COP-type) clustering, shrinkage discriminant analysis, and active selection are established prior art or standard tools, used only as instruments or comparators. The one candidate mechanism (constraint-based label-efficient commissioning) was pre-registered behind an intellectual-property gate, tested against a few-shot-supervised baseline, and found to confer no non-obvious gain over simply labeling a few examples; no new filing results. The per-deployment configuration-selection procedures are retained as trade secrets and are not disclosed. No third-party partner, customer, or commercial collaborator is named; the public datasets named are cited for reproducibility. © 2026 Fieldstone Analytics, LLC and the authors. Inquiries: randolphf@fieldstoneanalyticsllc.com. Companion deposits (spiral-domain-encoder-campaign) · The Deployment-Sizing Law — "size K down" and the label-hunger this record localizes to the projection (doi:10.5281/zenodo.23323323). References. · When Does a Frozen One-Byte Token Suffice? — the separability law whose screen this record shows does not forecast label-free commissioning (doi:10.5281/zenodo.23288053). Is part of. · The Service-Life Trajectory — companion sizing/service-life study (doi:10.5281/zenodo.23321165). References. · "The Frozen Token" — the four-laws predictive theory (doi:10.5281/zenodo.23002239). References. · The Composed Platform — the multi-capability frozen-token stack (doi:10.5281/zenodo.23401467). References. · A Deterministic Task-Discriminant Token Ladder — the program's accuracy lever on the multi-token axis (doi:10.5281/zenodo.22003179). References. · Class-discriminant single-token codebook — base encoder (doi:10.5281/zenodo.20788187). References. Data and reproducibility All datasets are public and are not redistributed: the UEA and UCR time-series classification archives and the standard CWRU bearing, MFPT, NATOPS, UCI-HAR, MIT-BIH ECG, and robot force–torque and EMG sets. The reproducibility archive contains the frozen pre-registrations, the runners and shared machinery, the result records (JSON), the six figures and their builder, and a README; absolute paths and personal identifiers are scrubbed; no company or partner names appear. This deposit continues the program's characterization of its filed and published frozen-token methods — here, a property of the deployment-sizing law's label-hunger — and discloses no new subject matter. Keywords frozen token; single-token encoder; class-discriminant codebook; label-fre