ADR 017 — Local Deterministic Routing Experiment Exception

1. Context

ADR-016 keeps routing engines, provider adapters and Runtime 2.0 implementation out of AletheIA’s current runtime boundary. A bounded experiment is useful only if it makes its inputs, hard filters and result inspectable without becoming a runtime authority.

2. Decision

Permit one isolated TypeScript function, recommendRoute(input), for local synthetic replay. It uses existing Ajv validation and selects only the first route eligible under the caller-provided preference order. The explicit default-deny request/provider allowlists, route availability, provider availability and every required capability are hard filters. Unknown required capabilities produce no_eligible_route; duplicate IDs and dangling references are rejected. The result preserves the validated input and both version identifiers.

This exception does not authorize a routing engine, scheduler, model/provider switching, tool selection or invocation, network access, kernel change, provider adapter, automatic execution or new runtime contract.

3. Consequences

Positive

• The filtering and rejection rationale can be inspected and replayed from a synthetic fixture.
• No hidden scorer or capability inference is introduced.

Negative / accepted tradeoffs

• Preference is caller-declared, not learned or optimized.
• The experiment cannot establish provider quality, cost or operational value.

4. Alternatives considered

• Implement a provider-aware router. Rejected: it would cross the ADR-016 runtime boundary.
• Document a hypothetical algorithm only. Rejected: it would not validate deterministic rejection behavior or replay fields.
• Add a reusable public contract or schema. Rejected: the local input schema remains experimental and does not change canonical runtime contracts.

5. Relationship

The experiment is subordinate to the Capability Routing Reconciliation: it recommends a declared local route but cannot approve, execute or govern a Work Slice. The public reconciliation and canonical schemas were reviewed; this experiment adds only its own local input schema and does not alter engine/kernel.ts, runtime adapters, harness behavior, policies or schemas used by canonical contracts.

6. Review

Reopen only with measured, source-backed evidence that this local replay experiment cannot express a needed decision and a separately approved boundary proposal defines the minimum safe interface.