Cross-cutting

Operating modes

Assistive vs Decisioning. The default is Assistive — DAVA proposes, humans dispose. Decisioning mode requires a contract amendment because the model is now acting on data, not just commenting on it.

Assistive (default)

Every product's output is a suggestion. Norm produces a cleaned file, Dedup proposes clusters, Connect proposes edges, Agent answers with citations — but nothing writes back to the customer's system of record without a human approving it. The Trust Layer enforces this: the policy decision point (PDP) returns allow only for read actions when the org is in Assistive mode.

Decisioning

The agent (or a Norm/Connect pipeline) is allowed to take actions directly — write back to the source system, dispatch a notification, flag a transaction. This is per-org and per-product, and requires a signed contract amendment because the liability profile changes.

In Decisioning mode, every action goes through a reviewer queue first if the policy is configured for human-in-the-loop. The reviewer workflow lives at /admin/reviews in the dashboard.

What the gate looks like

Action	Assistive	Decisioning
Read row / lineage / audit	✅ allow	✅ allow
Suggest a clean version of a row	✅ allow	✅ allow
Apply the cleaned version back to source	❌ deny	✅ allow (subject to PDP rule)
Send a notification or webhook based on a finding	❌ deny	✅ allow
Auto-resolve a flagged audit case	❌ deny	✅ allow (with reviewer queue if configured)

How to enable

Email support@avaresearch.ai with the customer ID and the products you want in Decisioning mode. We send back a contract amendment listing the specific actions enabled. Once signed, AVA staff flips the per-org PDP rule and audits the change on the immutable chain.