Semantic / intent-based routing: promote from ai#74 stretch goal to a scoped proposal

[jordigilh]

Context

This was previously scoped as ai#715 (transferred to experimental#5 before the repo migration) with a companion proposal PR, ai#750. Both were auto-closed on 2026-08-20 when proposals moved to praxis-proxy/enhancements — closed for process reasons, not rejected on merits. Linking here for full history/discussion.

Today semantic/intent-based routing has zero coverage in praxis, ai, or grid — it's tracked only as a stretch-goal bullet in praxis-proxy/ai#74, never sized or scoped on its own. Rough estimate: 3-4+ weeks for a demo-quality slice.

Design consensus already reached (see closed thread for full detail)

• Classifier, not embedding, is the real problem. Run it as an external/sidecar service reusing the already-shipped Add bounded native /v1/embeddings passthrough configuration and transport coverage ai#476 transport, rather than adding an in-proxy embedding engine (avoids Implement a pinned Candle MiniLM embedding engine for developer use ai#479/Expose the Candle engine as a feature-gated local /v1/embeddings filter ai#480 becoming new ML-lifecycle surface inside Praxis).
• Compliance tier comes from declared metadata only (tenant/session/app) — never content-derived, since classifying content to decide whether it's safe to classify is circular. Consistent with demo(grid) add runnable external-ingress GLB demo grid#7's existing principle that semantic routing stays subordinate to policy/entitlement.
• Concrete accuracy data point: a POC from @cnuland measured ~80% baseline / 98.5% after fine-tuning a 23M-param model on CPU.

Ask

Sign-off to move this to a formal Proposal under the new process (What?/Why? first), so it can get an EPIC issue in enhancements and continue toward an experimental prototype.

cc @shaneutt @cnuland

[praxis-bot]

It appeared this discussion was a duplicate of https://github.com/orgs/praxis-proxy/discussions/1017. If that is not the case, please let us know 🖖