# Five Gates Between a Guess and a Governed Answer > Five questions to ask of a business answer: what supports it, what it means, what was checked, which version is live and how the agent used it. Published: 2026-07-23T09:00:00Z Updated: 2026-09-10 Author: Archie Norman (Founder, MLX) Category: Twin Tags: twin, governance, approved-meaning, data-validation, publication, activation, evaluation, enterprise-ai Canonical URL: https://mlx.systems/blog/five-gates-to-a-governed-answer ## TL;DR - Source coverage, explicit meaning, validation, activation and answer evaluation support different claims. - The five gates are questions to inspect, not a literal release pipeline or five manual approvals. - Human review applies where required; automated evaluation and policy admission do not prove universal correctness. --- An assistant reports monthly recurring revenue. The figure looks plausible, but five questions still matter: which records supported it, which definition applied, what was checked, which version was live and whether the agent used it correctly. These questions need different evidence. A successful query cannot prove that the business definition is suitable, and a reviewed definition cannot prove that today's sources are complete.
Figure 01. Remove evidence at each control to see which claim it supports. These are five questions to inspect, not a literal five-step release pipeline.
## Five Questions to Keep Separate The gates below are a way to inspect a governed answer. They are not five approval screens or a complete description of the current implementation. MLX's current authoring path combines source discovery, compilation, execution checks, independent semantic evaluation and policy admission before installation and activation. It does not require a human to click approve on every definition. Human review applies where policy requires it or a business decision remains unresolved. Keep that distinction in mind as we follow a recurring-revenue definition through the controls. ## Gate One: Is the Evidence There? Before defining recurring revenue, establish that the relevant subscription and billing records exist, belong to the right organisation and cover the intended period. A connector may be configured while its latest sync has failed. A table may contain subscriptions without the cancellations needed to interpret them. A relationship diagram can look complete while important records are missing. Discovery should record what it found and what it could not establish. For our example, that includes the available contract periods, plan states, billing amounts and source freshness. This supports a candidate definition. It does not decide which plans the business should count as recurring revenue. ## Gate Two: Is the Meaning Explicit? A useful definition states how annual contracts contribute to a monthly figure, how discounts are treated and whether trials or paused subscriptions count. It also states the row meaning, often called the *grain*: one subscription per month, for example. These choices need an accountable policy. Some can be established from existing definitions and source evidence. Others need a finance owner to resolve an ambiguity. Automated semantic evaluation can check a proposed definition independently of its author. It cannot invent a missing business decision. Where approval is required, record who gave it and what they accepted. A signature establishes accountability; it does not make an incorrect definition true. ## Gate Three: Does the Implementation Pass Its Checks? The definition must become something executable. A Twin view contract names the fields and permitted calculations, while compilation and validation check the data structure and result. For recurring revenue, useful checks might cover duplicate subscription-month rows, missing currencies, invalid periods and known subscriptions that should be included. The checks should follow the risk in the definition. If a containment test uses a hand-verified reference sample, that sample must be independent of the result it is testing. A definition generating its own expected answer would only confirm agreement with itself. A passing check supports its declared claim. Including every record in a small reference sample does not establish that every other record is correct or that the measure is suitable for every report. ## Gate Four: Which Version Is Allowed to Answer? A validated definition and an active definition are different states. Publication gives the contract an identity; activation makes a particular version available for governed use. An operator should be able to establish which version is active and inspect the evidence that allowed it to be admitted. Missing or inconsistent required evidence should stop that promotion. For the reader, this answers a practical question: did this report use the definition before or after the change to discount treatment? Without version identity, a changed number can look like a changed business even when only the calculation changed. ## Gate Five: Does the Agent Use It Well? A correct contract can still be used badly. An agent might choose the wrong reporting period, omit a material caveat or describe an aggregate as though it had transaction-level evidence. A supported-question evaluation tests those behaviours against a named definition and version. Cases can include expected values, evidence requirements and requests the agent should refuse. This is separate from the semantic evaluation used while authoring a contract. Not every published product has an additional supported-question suite. Where one exists, its verdict applies to the cases and versions tested; it does not certify every future answer. Evaluation may happen before wider use or during a controlled rollout. Its position at the end of this article should not suggest that testing must wait until after activation. ## Inspect the Claim, Then Its Evidence When someone says a result is governed, ask what they can show: source coverage, an explicit definition, validation results, the active version and evidence from the actual answer. Each contributes something useful. None should stand in for the others. That separation makes it possible to investigate a disputed figure. You can ask whether the source changed, the definition changed or the agent used the right definition badly, and fix the part responsible.