Triggered by recognition, not by a threshold
Event recorders are not new — flight recorders, substation fault recorders and pre-event video buffers all capture a window around an incident. What differs here is what pulls the trigger.
The escalation is not fired by a channel crossing a limit or by an operator pressing record. It is fired by the system's own behavioural recognition: a genome matching a library signature above threshold. That means the high-resolution record starts while the episode is still developing and every individual channel is still inside its normal band — which is precisely the interval a threshold-triggered recorder misses, and precisely the interval a later investigation most wants to see.
What a frame contains
A conventional recorder preserves what the instruments read. This one additionally preserves what the machine concluded, second by second, so the record answers why the system did what it did and not only what the sensors saw.
- The full channel set for that second
- The genome computed at that instant
- The matched signature and its similarity score, if any
- The risk level then in force
- The enclosure heading — which ties the frame to the sweep in progress
- A camera frame, where one is available
Bounded, and cleared by the platform
The burst runs at one vector per second for up to five minutes, and both the duration and the return to baseline are governed by the hosted platform rather than by the field hardware.
That placement is deliberate. A field-side timer is a thing that can fail on, and a node stuck in forensic cadence is a node generating sixty times its normal traffic indefinitely. Putting the bound on the server side makes the escalation self-clearing without depending on field logic or on an operator remembering to switch it off.
A report as a by-product, not a feature
Because the frames carry reasoning as well as readings, the sequence can be rendered as a narrative: how the event developed, what the system concluded at each stage, when it rotated, what it located, and what it recommended — with the tabulated evidence beside it.
That produces a portable hazard report for an insurer, an inspector or a safety review, generated as a consequence of how the system already operates rather than as a logging feature bolted on afterwards. The distinction matters, because a report assembled after the fact from whatever happened to be retained is a different and weaker artefact than one captured at full resolution while the episode ran.
Why the baseline is slow on purpose
Ten seconds between vectors is not a limitation being apologised for. The hazard classes this system targets have precursor windows spanning five orders of magnitude — seconds for an open flame, weeks for a failing bearing — so any single fixed cadence is either too slow for the fast events or wastefully fast for the slow ones.
An adaptive cadence is the only honest answer to that spread: sample slowly enough to be storable indefinitely, and escalate on evidence. The system spends its data budget where the evidence says it is worth spending.
