Translucent Assistance

Definition

Translucent Assistance is the behavioural model for AI inside Hearth.

It means the help is present, visible, and useful, but it does not become the centre of attention or the owner of the work. When the user does not need it, it fades. When the user delegates a bounded job, it stays visible and interruptible.

What it includes

• soft contextual cues
• bounded delegated jobs
• visible progress while Whisper is working
• explicit uncertainty rather than fake certainty
• takeover without drama

What must stay visible

When Whisper is active, the human should be able to see:

Visible element	Why it must be visible
Current job	Prevents vague “AI is helping” theatre
Current step	Makes progress legible
Inputs in use	Helps the human trust or challenge the work
Confidence for this job type	Shows confidence-signal honestly
Policy state	Distinguishes capability from the whisper-charter authority boundary
Takeover controls	Preserves human control

Delegation states

These are the first-pass visible states for delegated work:

State	Meaning	Human can…
ready	Whisper could take this bounded job	delegate or ignore
working	Whisper is actively preparing	inspect, pause, take over
waiting	Whisper needs input or judgement	supply input or take over
paused	Work deliberately halted	resume, redirect, take over
taken-over	Human reclaimed the job	continue manually
complete-for-review	Whisper finished preparation	review, amend, accept, reject

Complaint exemplar

In the complaint prototype, a translucent assist should look like this:

flowchart LR
  A[Complaint concern] --> B[Cue appears: acknowledgement overdue]
  B --> C[Human delegates draft acknowledgement]
  C --> D[Whisper working]
  D --> E[Complete for review]
  D --> F[Waiting: missing tone choice]
  F --> G[Human takes over]

The important move is not that Whisper drafts something. The important move is that the human can always see where the work is and re-enter it naturally.

What it is not

• Not autonomous black-box execution
• Not modal interruption
• Not synthetic certainty
• Not AI taking credit for the outcome
• Not help that becomes a dependency

Design tests it implies

• Does the help fade when unneeded?
• If delegated, can the user see the work happening?
• Can the user interrupt and take over instantly?
• Does the assistance preserve human judgement where the stakes require it?