Whisper Charter

Definition

The Whisper Charter is the authority boundary for AI inside Hearth.

It defines three things:

1. what Whisper may do without human confirmation
2. what Whisper may do only when explicitly delegated
3. what must always remain human

Its job is to stop every Whisper feature from becoming a one-off judgement call.

Confidence is not authority. Capability is not permission.

What it includes

The first-pass Hearth charter is a three-zone model:

Zone	Meaning	Whisper may…	Typical examples
allowed	May act unaided	perform internal, non-binding, reversible assistance that does not speak for the human or commit the organisation	summarise, surface context, flag deadlines, detect blockers, suggest next movements, cluster similar prior work
delegable	May act only when explicitly delegated, while staying visible and interruptible	perform bounded, low-consequence operational work that remains reviewable, stoppable, and policy-allowed	prepare a chronology digest, draft an acknowledgement, assemble a review pack, chase missing internal information
human-only	Must never act alone	make consequential judgements, commit the organisation, or perform irreversible/public acts	decide an outcome, send external communications, approve final wording, alter rights/obligations, close a concern on behalf of the human

The governing rule is not technical possibility. The governing rule is whether the action is:

• reversible
• bounded
• visible
• non-binding or low-consequence
• private vs public
• judgement-light vs judgement-heavy

If the action fails those tests, it moves toward human-only.

Complaint exemplar

For the first complaint prototype:

Job	Charter state	Why
Surface that acknowledgement is overdue	allowed	Internal signal only
Prepare an acknowledgement draft	delegable	Useful and bounded, but still requires review
Chase an internal service owner for missing timeline	delegable	Operationally useful, but should only happen when the human chooses it
Suggest similar prior response pattern	allowed	Retrieval and suggestion, not commitment
Decide the complaint outcome	human-only	Consequential judgement
Send the final complaint response	human-only	External voice and organisational commitment

What it is not

• Not a capability list
• Not a model-confidence score
• Not a permission system by itself
• Not a promise that low-risk work is always safe
• Not a substitute for product policy, audit, or review

If a feature claims Whisper should do something because “the model is good enough”, that is a Charter violation unless the authority boundary still permits it.

Design tests it implies

• Is the proposed Whisper action clearly allowed, delegable, or human-only?
• If delegable, can the user see it, interrupt it, and take over immediately?
• Does the action avoid speaking for the human without consent?
• Would a regulator, buyer, or frontline worker understand why this action belongs in its zone?
• Does the Charter remain stricter than the model’s raw capability?