Permission to contribute a trace should have its own scope

A useful learning loop starts with a precise answer to what an owner has agreed to share. An owner might authorize a failure example for an internal evaluation while keeping the surrounding conversation private. We propose a contribution permission receipt that names the artifact, permitted uses and expiry, with each use checked at the point where it happens. This is a proposed architecture; the inspected public material establishes no current DXAP training-consent service.

The distinction matters because DXAP already has a different authority path. Its policy and execution documentation describes checking account authorization and configured trading policies before submission. Permission to inspect a contributed example would confer no ability to submit an order, revise a mandate or activate a setting. A contribution processor would run without those execution capabilities.

One example, three possible uses

Consider an illustrative owner report: “The agent interpreted my two-position limit as two positions per symbol.” The owner selects the relevant instruction, a redacted state snapshot and the rejected proposal. They approve internal evaluation, decline public publication and leave model-training use undecided. The receipt records those three choices independently. A general “help improve DXAP” button would conceal the distinction precisely where later consumers need it.

The internal evaluator can then build a fixture asking whether a third distinct symbol is rejected. A dataset builder preparing examples for weight updates must stop because that use lacks authorization. A researcher preparing a public article must request a separately reviewable excerpt. Existing approval for evaluation cannot silently carry either operation.

The receipt should bind to the reviewed artifact version. If a curator appends a longer conversation or an account state, the resulting object requires renewed review for the additional material. Otherwise permission can expand through routine preprocessing. A parent-child link helps identify that expansion, while the owner-facing view shows the actual content and use choices.

Check the boundary through attempted misuse

We would test this design with attempted publication, attempted training export and attempted trade submission from an evaluation-only contribution. Expected results are specific denials with the missing authority identified. We would also test an approved evaluation export to ensure the controls preserve the useful path. The denominator includes all attempted uses, including requests the processor cannot classify.

The controls paper companion supplies a relevant engineering precedent: authenticated mandates and policy checks have distinct roles along the action path. Our proposed extension applies that separation to contributed data. The release criterion would be complete use attribution and correct handling of the tested permissions, rather than a claim that more contributed traces already improve forecasts. Owners could help construct sharper tests while retaining a clear boundary around both their data and their live account.