Console & fleet

Feedback becomes a training candidate

When you tell the console that a decision was right or wrong, that verdict does not silently become training truth. It becomes a candidate that has to pass a check first. This page explains the loop: how your feedback is captured, the fidelity gate it must clear, and why nothing trains automatically.

~6 min read · Console & fleet

This builds on cloud-mode incident review; a private device records its decisions locally and reports none of this.

Your verdict starts the loop

After an incident you can record a verdict on it, and only two are accepted:

  • Attack: the system was right to act.
  • Legitimate: it was a false alarm.

Recording a verdict marks the incident as a feedback candidate. It changes no status on the device, triggers nothing on the machine, and is invisible to the daemon. You can also correct a verdict you gave earlier.

The fidelity gate

A verdict only becomes a usable candidate if the evidence behind it is trustworthy. Each candidate is checked, and dropped if it fails:

  • Has a diff: a candidate with no stored redacted change is dropped.
  • Faithful diff: the stored change must provably correspond to the text that was actually judged; an unrecognised or absent fidelity marker is dropped.
  • Counted, not hidden: dropped candidates are tallied by reason, never silently discarded.
A missing fidelity marker is never treated as faithful. An older device that reported no proof of what it judged is left out, not trusted by default.

A candidate, not a training run

Passing the gate does not train anything. The step that gathers candidates is read-only: it writes nothing back to the device and nothing back onto the record. Your verdict still has to pass the same independent re-check as any other label before it can enter a dataset. There is no automatic training, no model change, and no silent flip.

What a candidate carries

A candidate is the incident reduced to what a re-check needs:

FieldMeaning
pathThe file the incident was on
redactedDiffThe judged change, redacted and added-only
diffFidelityThe provenance marker the gate checks
dangerThe graded danger value, when present
verdictYour judgment: attack or legitimate
submittedAtWhen the verdict was recorded

Where the incidents and verdicts live: incidents and audit. What the redacted diff actually contains: what Cloud mode sends. The hold your verdict is about: user confirmation prompts.

Was this helpful?

Your judgment counts, and is still checked.

Feedback becomes a re-checked candidate, gated by fidelity, never blind training. Reviewed in the console.

Sign me up