The Composition Energy Law: A Pre-Registered, On-Silicon Measurement of When Many Always-On Capabilities Share One Frozen Decision's Cost — and When They Do Not
The Composition Energy Law: A Pre-Registered, On-Silicon Measurement of When Many Always-On Capabilities Share One Frozen Decision's Cost — and When They Do Not Randolph James Ferlic, M.D. and Kimberly Kate Ferlic — Fieldstone Analytics, LLC, Austin, TX, USA Preprint · Zenodo DOI: 10.5281/zenodo.24426652 · CC-BY 4.0 · Community: spiral-domain-encoder-campaign · a pre-registered, on-silicon characterization of the energy economics of composing many always-on capabilities on one previously published frozen single-token decision. The base encoder, its competence monitor, the multi-token / token-ladder mechanism, and the geometry-gated drift adaptation are previously published and separately patent-pending; this work measures their composed energy and introduces no new claimed matter. The per-deployment selection of which capabilities to compose and how is retained as trade secret — see the notice below. Abstract A frozen, deterministic, class-discriminant single-token decision encoder — a fixed projection and a nearest-of-≤256-centroid lookup that reduces a window of sensor observations to one 8-bit token — is not a single capability but a substrate. The same token already powers a whole family of always-on functions: the base decision, its nearest-centroid competence monitor, an on-device safety interlock, a streaming anomaly monitor, a multi-head token ladder, an amplitude-invariant ordinal front end, and a geometry-gated drift adaptation. A companion study measured the energy of one decision. This paper asks the deployment question that follows and that a careful evaluator asks first: what does running the whole stack cost on real silicon, and does each capability pay for itself? Under pre-registration — with Go/No-Go bands fixed before any measurement — and on the microcontrollers the method targets, we find a sharp and simple law. Composition is sub-additive: the full always-on Tier-0 stack costs 1.002× a single decision on the host op decomposition, because every capability rides the one feature-extraction-plus-token computation and adds only a readout (a threshold on a distance the encoder already returns) or one bounded recompute; the ratio of K separate per-capability models to the shared-token composition is 3.7× against a conservative feature-fed comparator and up to 29× against a realistic raw-signal detector at K = 4, and that frugal multiple is linear in K and unbounded (measured 2.0×, 4.0×, 8.0×, 16.0× at K = 2, 4, 8, 16, the composed stack staying flat near 6,080 ns while K separate models grow in proportion). The mechanism is not a definitional trick: measured call-counters show the composed K = 4 stack invokes the front end and the token encode exactly once each per decision against a naive stack's four and four — a measured reuse ratio of 3.87× ≈ K. The crux is that the shared front end is 92% of the base decision, so the expensive part is paid once and each reused capability is nearly free. The marginal energy of a capability is governed by which of four classes it belongs to, and the classes carry four measured signatures: a stateless-free readout costs 0.037% of the base decision, a bounded-re-derivable recompute (the drift adaptation) 0.128%, an extra token on the ladder one more projection-and-search (about 7.7%), and a needs-persist keep-warm commit — the one genuinely expensive class — 860 µJ per commit on the device's non-volatile memory, about 65× a base decision, measured on silicon and dominated by erase latency. The token ladder, the one capability that materially moves the cost, has a measurable return-on-investment crossover: on the two modalities with accuracy headroom on a matching partition axis it pays fast (ECG +7.3 points to a crossover depth D* = 2; human-activity recognition +1.9 points to D* = 3) and saturates fast (every doubling past the knee buys < 0.5 points), while on a saturated signal (bearing fault, already 0.996) or a mismatched axis it is dead weight — the "No Winning Ticket" of energy, with a sharp knee. The composed sizing law follows: the decision-only minimum storage capacitor rises only to 1.28× the single-token capacitor at a useful ladder depth, and the full boot-to-first-decision capacitor to 1.0007×, because the platform system-wake dominates the per-power-up budget by ≈313× and the capabilities are 0.07% of the deployment capacitor. Everything is reproduced on two cores — a Nordic nRF54L15 (Cortex-M33) metered with a Power Profiler Kit II and an nRF52840 (Cortex-M4F) — with the composed decision bit-exact across both (0 of 256 token flips, 0 of 54 ladder flips on each core, 0 cross-core flips). On silicon the single dominant energy lever is confirmed to be the shared front end: one front-end pass measures 139 µJ, so the ladder's separate front end for depth 4 is roughly 1.1 mJ — which dwarfs the ladder decision itself (57 µJ) and is exactly what a shared-spectrum ladder removes. We report the honest boundaries with equal prominence: composition is sub-additive only for capabilities that read out the shared token; a capability with its own transform (an ordinal, topological, or recurrence front end) is additive, costing from 2.3× to 95× the base decision. The frequently-cited int8 energy win does not transfer to a core with a hardware floating-point unit — on the Cortex-M33 the int8 base decision is 8% worse — so that lever is specific to floating-point-less microcontrollers, though sub-additivity itself is arithmetic-independent. The law is not specific to the Fourier transform: the frugal multiple is ≈ K for three different shared transforms, and holds on a fourth sensor modality (surface electromyography). The whole result is a characterization of previously published, separately patent-pending matter; it introduces no new claimed mechanism, and the per-deployment selection procedure is retained as trade secret. Highlights · Composition is sub-additive, and the frugal multiple is unbounded. The full always-on Tier-0 stack (base decision + competence monitor + interlock + anomaly monitor + drift adaptation) costs 1.002× a single decision. The ratio of K separate per-capability models to the shared-token composition is 3.7× (conservative) to 29× (realistic) at K = 4, and is linear in K — measured 2.0× / 4.0× / 8.0× / 16.0× at K = 2 / 4 / 8 / 16. · It is not a definitional cheat; measured. Call-counters show the composed K = 4 stack invokes the front end and token encode exactly once each per decision versus a naive stack's four and four — a measured reuse ratio 3.87× ≈ K. The crux: the shared front end is 92% of the base decision, so each added capability is rounding error (together +0.24%). · Four capability classes, four energy signatures. A stateless-free readout costs 0.037% of the base decision, a bounded-re-derivable recompute (drift adaptation) 0.128%, an extra ladder head one more projection-and-search (~7.7%), and a needs-persist keep-warm commit — the one expensive class — 860 µJ ≈ 65× a base decision, measured on silicon and dominated by erase latency. The predicted ordering holds exactly. · The ladder ROI has a sharp crossover. Where there is headroom on a matching partition axis the ladder pays fast and saturates fast (ECG +7.3 points to D* = 2; human-activity +1.9 points to D* = 3); where the signal is saturated or the axis mismatched it is dead weight — the "No Winning Ticket" of energy. · The composed capacitor is ≈ the single-token capacitor. The decision-only minimum capacitor rises only to 1.28× at a useful ladder depth, and the full boot-to-first-decision capacitor to 1.0007×, because the platform wake dominates the per-power-up budget by ≈313× and the capabilities are 0.07% of the deployment capacitor — not K×. · Measured on two cores, bit-exact, with the boundary mapped. Reproduced on a Cortex-M33 (metered) and a Cortex-M4F, the composed decision is bit-exact on both (0/256 token, 0/54 ladder each, 0 cross-core flips). The honest boundaries: own-transform capabilities are additive (2.3×–95× base); a needs-persist commit is expensive; a ladder token past the knee is dead weight; and int8 does not save energy on a floating-point-unit core (8% worse on the M33). The four capability classes carry four energy signatures The marginal cost of a capability is set by which of the four previously-established classes it belongs to, and the measured signatures span four orders of magnitude. The predicted ordering — stateless-free < bounded-re-derivable < extra token < needs-persist — holds exactly; the needs-persist figure is measured on silicon. capability (class) marginal cost, % of one base decision note competence monitor / interlock / anomaly (stateless-free) 0.037% a threshold on a distance the token already returned geometry-gated drift adaptation (bounded-re-derivable) 0.128% a running-mean update over the committed window, no re-search one extra ladder head (extra token) ~7.7% one more projection + 256-centroid search keep-warm commit (needs-persist) ~6,500% (≈65×) 860 µJ on silicon, erase-dominated — the one expensive class The frugal multiple is linear in K, and the reuse is measured The economic case is that the second, third, and Kth capability are nearly free. Sweeping the number of always-on capabilities K, the shared-token composition stays flat while K separate models grow in proportion — a frugal multiple ≈ K with no ceiling — and the sharing is a measured fact (one encode, K readouts), not an accounting convenience. K shared-token composition (ns) K separate models (ns) frugal multiple 2 6,054 12,207 2.0× 4 6,112 24,300 4.0× 8 6,064 48,672 8.0× 16 6,080 97,458 16.0× Measured call-counters over 100,000 decisions: the composed K = 4 stack calls the front end and the token encode 1 and 1 per decision (6,614 ns); a naive per-capabi
Authors
- Randolph James Ferlic (ORCID: https://orcid.org/0009-0006-2367-5205)
- Kimberly Kate Ferlic (ORCID: https://orcid.org/0009-0007-3012-3430)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-21
- DOI
- https://doi.org/10.5281/zenodo.24426651
- Primary Topic
- Energy Efficiency in Computing
- Type
- preprint