What a receipt evidences about a responsible-AI commitment
Can a Decision Receipt show that an AI decision was fair, accurate or correct?
No. A receipt shows which policy version and context governed the recorded assessment, and it lets another party verify that record afterwards. It does not establish that the inputs described the world correctly, that the ruleset encodes a defensible standard of fairness, or that the decision was right. Where a model produced a recommendation upstream of a governed decision, the receipt evidences the decision, not the model.
In more detail
A responsible-AI commitment and a decision record are different kinds of thing, and the documentation separates them carefully. Some commitments are statements of intent, describing a disposition an organisation holds. Others can be evidenced at the moment a specific decision is made, because a record can be produced showing the claim held for that decision. Neither kind is weaker than the other, but reaching for a decision-time record to support a commitment it cannot carry is a category error, and it produces confident-looking evidence for something nobody measured. What the record does carry is narrow. It shows that a decision was evaluated against a particular ruleset, bound to the context that was supplied. Whether that ruleset encodes a defensible standard is a question about the policy, answered by whoever wrote and approved it. Whether the supplied context described the world is a question about the inputs, which the record cannot settle because it binds what it was given.
Limits
The accountability commitment is evidenced only partly. The record binds the actor identifier the calling application supplied; it does not authenticate it. See actor-identity.
The record is narrow by design, and its value comes from being exactly as narrow as it says.
Several responsibilities stay with the organisation and no release delivers them: whether the policy encodes a defensible standard, whether the supplied actor is the accountable party, the quality of any model or human judgement upstream, retention and publication, and how verifying parties obtain trusted roots.
This entry states what a record evidences. It makes no statement about what is enabled in any particular deployment.
Where it applies
Sort a list of responsible-AI commitments into the ones a decision record can evidence and the ones it cannot, before the record is offered as evidence for any of them.
Answer a regulator, auditor or customer who asks whether a signed record shows the decision was correct.
Keep a fairness or accuracy question with the people who own the policy and the inputs, rather than deferring it to the artefact that records their use.
Sources
MeshQu docs - Responsible-AI evidence (opens in a new tab)
The documentation owner's account of which responsible-AI commitments a decision record can and cannot evidence. Every claim on this entry is quoted from it.
MeshQu docs - Concepts overview (opens in a new tab)
The explicit non-goal that MeshQu is not an explainability tool, which is the model-reasoning boundary stated from the product side.
MeshQu docs - Actor attribution (opens in a new tab)
Why the accountability commitment is evidenced only partly: the actor identifier is bound but not authenticated.
Related answers
The limits of verification
What does verifying a Decision Receipt actually prove?
What a receipt records about who acted
Does a Decision Receipt prove who took the action?
What MeshQu is not
What is MeshQu not, and which problems does it deliberately not solve?
Decision Receipts: what the record contains
What is in a Decision Receipt, and what does it prove?