Human Review Is a Workflow State, Not an Approve Button
Published August 2026 by Solarc Labs
Approval theatre is easy to build
A workflow can display an Approve button and still leave the reviewer with no meaningful control. If the person cannot see the source material, the proposed action, the uncertainty or the consequence of approval, the interface records a click rather than a substantive decision. Across Solarc product work, we treat review as a state with entry conditions. The system should know why an item needs review, what evidence must be visible before the reviewer acts, and which transitions are allowed afterwards. That makes the review point part of the operating model rather than a cosmetic safety layer.
The reviewer needs an evidence packet, not a confident summary
A useful review packet starts with the exact source the workflow inspected and keeps derived facts visibly separate from the source. It should also show deterministic exceptions, model-assisted suggestions where they exist, and any unknown or unsupported state that still needs judgement. This is why several Solarc workflows stop at states such as ready_for_review instead of silently converting a clean extraction or automated check into an approval. The reviewer should be able to compare the source with the proposed interpretation and reject, correct or escalate without reconstructing context from another system.
State transitions make authority explicit
The control question is not simply whether a human appears somewhere in the diagram. It is who is allowed to move the workflow from one state to another and what that transition authorizes. A reviewer may be allowed to approve an extraction but not trigger payment, accept a document but not certify legal sufficiency, or acknowledge receipt without creating an electronic-signature claim. Writing those boundaries into the state model reduces accidental authority expansion. It also gives tests something concrete to verify: an automated component must not bypass the review state, and a reviewer action must not unlock a broader consequence than the product scope promises.
Log the decision boundary, not hidden reasoning
An operational record should preserve the source reference, proposed action, reviewer identity or role, decision, timestamp, relevant exception state and resulting transition. It does not need to expose private model chain-of-thought to be useful. What matters is enough inspectable evidence to explain what the system presented and what the person authorized. That distinction keeps the audit trail practical. The goal is not to preserve every internal computation. The goal is to preserve the externally meaningful control event so another operator can understand how the workflow moved and what evidence was available at that point.
Consequential automation earns autonomy gradually
Solarc starts consequential workflows with bounded human gates and expands automation only when the real operating context produces evidence that a narrower control can be relaxed safely. That may mean exception-only review later, but it should be an earned change backed by observed workflow behaviour rather than a default assumption that fewer clicks automatically means a better product. Human review is therefore not the opposite of automation. It is one of the mechanisms that lets automation become trustworthy enough to expand.
Continue the job