AI resolution

AI resolution associates discovered observations with existing program scopes. It interprets messy crawler output under the scope rules Sonar already stores; it does not invent new engagement policy.

Place in Inventory

program + valid scopes + instructions
                    │
discovered observations
                    │
                    ▼
             AI resolution
                    │
        ┌───────────┴───────────┐
        ▼                       ▼
resolved scope items      needs review / unmapped

The global recency phase limits the programs considered by normal resolution runs. A program is skipped when its latest saved resolution already covers the current program version.

Inputs

For one program, resolution receives:

• valid scopes and their instructions;
• the program version;
• discovered asset or scope-like values that need association;
• any previously completed chunks that can be resumed safely.

Outputs

Resolution produces a saved resolution plus individual items. An item records its value, type, source, candidate scope IDs, confidence, reasoning, and review state. Items can be accepted through deterministic handling, sent to verification, or left for human review.

Accepted items raise the materialization event that creates the appropriate Inventory association. The resolution record retains provenance and review counts; it is not merely a transient model response.

Execution behavior

• Programs run concurrently within configured limits.
• A failure isolated to one program does not discard healthy sibling work.
• Quota failures halt the run because sibling requests would hit the same ceiling.
• Saved chunks support resume; poisoned parse-failure chunks are removed so a later run can ask the model again.

Model boundary

Resolution uses the shared chat-provider registry and a stable system prompt. Provider and model selection are deployment configuration. They can change without changing the Inventory concept or the scope rules applied.

Resolution is a candidate-association stage. A model response alone is not proof that a value is in scope; verification and human review remain explicit outcomes.