Why durable state matters in a long scientific review

A long scientific review is not one conversation. It is a sequence of evidence cuts, calculations, interruptions, challenges and approvals. Stateless chat treats each return as a reconstruction problem.

State must live outside the model

ARiDA’s session agent persists per chat and checkpoints between phases. Plans and progress are written to workspace files before and after dispatch. Evidence, calculations and reports are durable artifacts. If the provider call stops, the work does not vanish with it.

This architecture matters because conversational memory is not an audit record. The system needs to know which asset revision was accepted, which files were authoritative, which lanes completed and which decision objects became stale after a change.

Interruptions are ordinary operations

A user can stop an in-flight provider call without destroying the session. New evidence can steer the next phase. Idle sessions suspend and resume. The same runner retains the working conversation while the file-backed state provides the recoverable truth.

For governed work, continuity extends beyond chat. A run has an envelope, lane states, spend limits, output gates and a terminal status. Failure cannot be converted into completion by a friendly final message. The case remains blocked or incomplete until its contract is satisfied.

Persistence creates accountability

NIST’s AI RMF emphasizes lifecycle monitoring, documented roles and systematic records (NIST). Persistent sessions support that standard only when they preserve the right objects. More conversational history is not enough.

The enterprise value is continuity of judgment. A reviewer can return to the prior evidence state, understand why an assumption moved and see who approved the released decision. That is what stateless chat cannot supply.