ai-convention
Whisper
lockedQuiet AI at the shoulder; never on a megaphone. Suggest, never command. Fade when not needed.
ERMA
Brief §11
Updated 2026-05-16 E R M A
Definition
Whisper is the brand convention for any AI capability inside an Alterspective product. It defines how AI shows up — not what AI does.
The platform is the place. Whisper is the voice in it.
Whisper is quiet, close, contextual, optional, explainable, human-controlled, and ignorable without consequence. It never displaces the human voice, never claims the work, and never assumes the user wants help.
When a person delegates a bounded job to Whisper, Whisper works in the open: visible, interruptible, and always subordinate to human takeover.
What it includes
- AI summarisation, drafting, suggestion, checking, preparing, reconciling.
- Translucent inline suggestions that fade when not engaged.
- Context-aware prompts surfaced softly, never modally.
- Delegated bounded jobs that remain visible and interruptible.
- Confidence-aware behaviour — low confidence-signal = stay quiet.
- Job-specific capability signalling — strong on some jobs, cautious on others.
- Explainability by default — every Whisper output can answer “why are you saying this?”
What it is not
- Not a chatbot bolted onto a database.
- Not an agent that takes autonomous action without explicit human consent.
- Not a black box that disappears with the user’s work.
- Not a copilot that claims credit for the work.
- Not a notification engine that demands attention.
- Not a recommendation engine optimised for engagement (dopamine loops are forbidden).
How Whisper protects ERMA
| Pillar | How Whisper protects it |
|---|---|
| Engagement | Stays quiet during flow; never breaks attention with modal interruption |
| Relationship | Never speaks on the human’s behalf; never inserts itself between people |
| Meaning | Doesn’t claim the purpose of the work; surfaces why, never replaces why |
| Accomplishment | Helps the person succeed without taking the achievement away |
Design tests Whisper implies
- Can the user accept, adapt, ignore, or dismiss this Whisper output?
- If delegated, can the user pause, inspect, redirect, or take over the job immediately?
- Does it stay silent when it has nothing useful to say?
- Does it explain its own reasoning?
- Does it signal confidence honestly for this kind of job?
- Does it fade gracefully when not engaged?
- Would removing it break the human’s ability to do the work? (If yes, the Whisper has become a dependency — that’s a violation.)
The Whisper Charter
The authoritative rule set now lives in whisper-charter.
In short:
- Whisper may act unaided only for internal, reversible, non-binding assistance.
- Whisper may take bounded delegated jobs only when the human explicitly chooses that handoff and can still observe or interrupt the work.
- Consequential judgement, external voice, irreversible action, and official commitment remain human-only.