Field monitor
Is the deployed Reflex still seeing the world it was tested on? Compare outputs, replay exact branches, and know what to check next.
1. Choose the reference Reflex
Loading Reflex and replaying its reference inputs…
2. Bring a decision log
Try an exact replay before connecting a device.
Choose a Reflex, import its timestamped outputs, then compare them with its reference panel. No fleet account, cloud upload or guessed confidence score is needed.
Simulation controls · synthetic drift, not device evidence
Ramps from zero to this fraction of each tested feature span.
Chance of drawing probes from the first reference class.
Uniform perturbation added to the reference inputs.
Firmware log format & what this can prove
Decision log CSV: t_ms,class_id,path_id, optionally followed by f0,f1,…,fN in the Reflex’s feature order. Each record is one decision, timestamps in milliseconds, class and path as integer IDs. A header row is required; # comments are ignored. Resolve timestamp resets before import.
t_ms,class_id,path_id,f0,f1 0,0,0,0.25,-0.5 1000,1,1,1.2,0.8
This example has two features; actual logs must match the selected Reflex. In the Reflex graph ABI, path_id is the terminal node index and equals class_id. Different branch sequences may reach the same terminal. Inputs enable exact route replay and range checks; without them the monitor cannot recover path tests or assess input drift.
These are distribution/streak warnings, not accuracy measurements. Field ground-truth labels are absent. A matching reference Reflex hash must be verified in your firmware release manifest; this minimal CSV does not carry a Reflex hash.