ai-convention

Confidence Signal

locked

An honest signal of how reliably Whisper can help with a specific bounded job. Not a permission slip and not a global AI score.

ERMA
E M A
Brief §18 Updated 2026-05-16

Definition

A Confidence Signal expresses how strong Whisper is for a specific kind of bounded job in the current context.

It is attached to the job, not to “the AI” in general. It helps the human decide whether to ignore, review closely, delegate, or take over.

Confidence answers “how strong is Whisper here?” It does not answer “is Whisper allowed to do this?”

That second question belongs to policy and human judgement, not confidence.

What it includes

  • a job-specific strength signal
  • a clear distinction between capability and permission
  • honest uncertainty rather than inflated certainty
  • signals that can improve over time as Hearth learns where Whisper is strong or weak

First-pass complaint examples

Bounded jobModeConfidenceWhy
Prepare acknowledgement draftprepare-onlyhighPatterned, reversible, easy to review
Chase missing internal timelinedelegablemediumOperationally useful, but depends on local context and tone
Suggest similar prior response patternprepare-onlymediumHelpful retrieval task, but analogies may mislead
Propose final complaint outcomehuman-onlylowJudgement-heavy and materially consequential
Flag systemic issue pattern across complaintsprepare-onlylow-to-mediumUseful signal, but false grouping risk remains

First-pass bands

The signal can start simple:

BandMeaning
highWhisper is usually strong here; normal review still applies
mediumPotentially useful, but inspect more closely
lowWhisper may still assist, but should stay cautious and secondary
unknownNot enough evidence to present a strong signal

These bands are deliberately coarse. Precision theatre is the wrong move this early.

What it is not

  • Not a permission model
  • Not a legal or policy waiver
  • Not a global product trust score
  • Not an excuse to hide the work itself
  • Not a replacement for human judgement on consequential work

Design tests it implies

  • Is the confidence attached to a bounded job rather than to Whisper as a whole?
  • Can the user distinguish confidence from permission?
  • Does low confidence make Whisper quieter rather than louder?
  • Does the signal help the user decide what to review, delegate, or take over?
  • Does the signal stay subordinate to the whisper-charter rather than trying to override it?