Read evidence by scope
Evidence labels identify how a result was obtained. Measured means a target-board observation with its recorded instrument and boundary. Verified means a stated property or replay was checked. Local split means a calculation on this browser’s supplied partitions. Illustrative means arithmetic or an example under stated assumptions. Not measured is missing evidence, not zero cost. Stale means an input binding changed. Needs service identifies an unavailable compiler or hosted operation.
| Evidence | Keep beside the number |
|---|---|
| Measured | Board, firmware, workload, instrument, window and baseline policy. |
| Verified | Source/graph hash, numeric scope, toolchain and check performed. |
| Local split | Dataset version, split and selected session. |
| Illustrative | Assumptions and excluded system costs. |
These are different kinds of evidence, not a ladder that makes accuracy become energy. An exactness receipt and a PPK2 trace answer different questions. Both need to refer to the right Reflex.
The published C1 campaign
| Item | Published method |
|---|---|
| Board | Nordic nRF54LM20B DK Measured |
| Processor | Arm Cortex-M33 at 128 MHz, CPU only Measured |
| Boundary | C1: processed float features in RAM → published class Measured |
| Instrument | Nordic PPK2, about 1 kS/s Measured |
| Protocol | 64 phases, 784 windows, frozen artifacts Measured |
| Energy | Measured charge × assumed constant 3.0 V Illustrative |
| Datasets | 5 tasks, 6,288 rows Measured |
| Evidence | Nordic evidence v6.3, 4 Nov 2026 Measured |
C1 starts at processed float features already in RAM and ends at the published class. It excludes acquisition and raw waveform processing. The measured board is the nRF54LM20 DK B revision, with the Arm Cortex-M33 CPU only. It is not an accelerator result or a complete sensor-to-radio application budget.
The stateless cost workloads use fixed equal-weight anchors; historical accuracy uses the embedded panels. Transport uses an ordered stored panel with its recorded reset policy, not the complete original acquisition chronology. Accuracy panels are historical implementation checks, not fresh prospective estimates. The frozen comparator models have unequal quality; ratios do not demonstrate framework superiority at equal accuracy. The frozen TensorFlow Lite Toothbrush model has 24% historical accuracy Measured. Transport’s historical panel does not establish perfect unseen-session recognition.
Measured charge, derived energy
PPK2 current integrated over time gives charge. Published energy is that measured charge multiplied by an assumed constant 3.0 V; the actual DUT rail was not independently recorded. Gross charge includes all current in the active interval. Idle-corrected charge subtracts a measured nearby idle contribution and must remain separately named.
Q[nC] = integral I[µA] dt[ms]
E[nJ] = Q[nC] × V[V] // only for a constant rail
// Published conversion: assumed 3.0 V, not independently verified.
// A varying measured rail requires integral V(t) × I(t) dt.| FICUS Reflex | C1 latency | Gross derived energy |
|---|---|---|
| Bearing | 1.336 µs Measured | 10.998 nJ Illustrative PPK2 charge × assumed 3.0 V |
| Toothbrush | 0.539 µs Measured | 4.412 nJ Illustrative PPK2 charge × assumed 3.0 V |
| Magic Wand | 1.572 µs Measured | 12.852 nJ Illustrative PPK2 charge × assumed 3.0 V |
| Gearbox | 6.648 µs Measured | 54.444 nJ Illustrative PPK2 charge × assumed 3.0 V |
| Transport | 3.326 µs Measured | 28.817 nJ Illustrative PPK2 charge × assumed 3.0 V |
Shorter execution can reduce decision charge without reducing active power. Do not relabel a charge ratio as a power ratio. Any battery scenario is illustrative, compute share only: sensing, sleep, DSP, radio, supply loss and application duty cycle remain additional costs.
Prepare a comparable measurement
Freeze the exact Reflex, firmware, features, compiler flags, board revision and clock. Verify probe classes before timing. Record instrument mode, rail wiring, marker pin, actual sample rate, warmup, calls per window, reset/state policy and rounds. Batch short calls so current windows contain enough samples; quantify GPIO and loop overhead, and keep verbose logs outside the boundary.
The PPK2 runner is a reference outline, not a CI-built measurement system. Validate its sample clock, edges and continuity on hardware before publishing results. Importing a trace does not itself authenticate the firmware. Preserve raw captures and sensitivity to edge/baseline choices; repeatability checks are not confidence intervals for deployment populations.
Memory also needs a scope. The published shared Instinct component is 77,104 B Measured of linked code and read-only data, with 2,320 B Verified caller-owned checked context. These exclude stack, inputs and system memory. Individual-Reflex linked roots, comparator parameter arrays and TensorFlow Lite model blobs are unequal scopes; no total-memory saving factor follows from them. The shared Instinct capture is separate and cannot be multiplied into individual-Reflex comparator factors.
Last reviewed .
Source of truth: published facts in content/public.ts; CURRENT_MEASURED_RESULTS.md; CONNECTED_BUILD_WORKFLOW.md; standalone C emitter and v3 endpoint contract. Reference examples are not physical-board test receipts.