Observe decisions without inventing accuracy
The Field monitor inspects imported device decision records against a selected Reflex and a reference distribution. It can help identify shifts in class frequency, path usage and input coverage. It is a browser diagnostic workflow, not a hosted fleet-management system. Remote ingest, device enrollment and persistent fleet infrastructure need connected service capabilities.
A changed class distribution can reflect a real change in use, a sensor fault, a preprocessing change or a model problem. Without authoritative labels it is not a new accuracy estimate. An unseen path is a reason to inspect the event, not proof of an error. Separate data drift, implementation disagreement and task failure before choosing a repair.
Log a real decision contract
t_ms,class_id,path_id
12000,0,17
12020,1,29A decision log has a millisecond timestamp, class ID and actual path ID. Keep firmware, device identity, Reflex hash and class/path mapping with the file. Timestamps since boot are not UTC: record a boot/session identity and synchronization if logs span restarts. The field schema differs from the raw sensor capture grammar, whose value columns are channel measurements.
The reviewed standalone predictor exports uint16_t <runtime_name>_legal_predict(const float *). It has no path-ID return or trace API. To populate path_id, add a reviewed trace adapter bound to the exact graph and define a stable path mapping in the release metadata. Do not set it equal to the class, a random value or an invented runtime function. Deploy’s optional logging line stays disabled until a real path-ID source is supplied.
Compare against the right baseline
Choose a baseline from the same graph, feature contract and device operating context. Preserve its time period and sample population. A catalog probe set is a verification fixture, not a representative field distribution. Class balance from local training can also differ intentionally from deployment prevalence; a shift between those two populations is not automatically a defect.
| Observation | Useful next check |
|---|---|
| Class-frequency shift | Compare use/context and gather labels for affected events. |
| Unseen path | Replay its finite feature vector against the same graph. |
| Out-of-range feature | Check units, scaling, wiring and coverage of the reference set. |
| Different Reflex hash | Select a new baseline; do not merge path IDs across graphs. |
Input range diagnostics need logged inputs. A file with only class and path IDs cannot establish which feature went out of range. If you log inputs, retain the manifest order and float32 interpretation, minimize sensitive fields and review the privacy consequences. Never assume dataset rows were uploaded merely because monitoring is enabled.
Turn an observation into a reviewed change
Open suspicious events in Results where the graph and finite input vector are available. Replay the whole decision and check the actual branch conditions. Ask whether the vector is physically plausible and whether acquisition or preprocessing changed. Path replay explains what the classifier did, not why the real-world condition occurred.
Gather independently labeled examples before deciding to retrain. Keep new cohorts separate from the data already used for model selection. A retrained Reflex needs a new package, device parity checks and target measurements if its resources or charge matter. Preserve the old baseline rather than overwriting it with the new model’s behavior.
For a C1 measurement, keep logging outside the active decision interval. Field instrumentation adds work and may expose private information; it is not free observability. The contract includes a planned authenticated monitor-ingest endpoint, but its presence in an endpoint registry is not proof that a fleet service is available. Use local imports today and check service readiness before designing remote automation.
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.