ADR 0005 — Typed decisions through a cascade; Laya as the embedded decision model

• Status: Accepted
• Date: 2026-10-13

Context

Routing used a cheap generative model (a one-word classification); readers used hand-written rules; some safety checks were prompt rules. A hosted "decision model" API (Jev, TypeSafe AI) showed the shape that fits: typed questions answered with calibrated probabilities in one forward pass. It is proprietary and metered per token. Laya (Convai Innovations, Apache-2.0) is an open-weight model that speaks the same System One protocol, with published parity on the typed-decisions benchmark.

Decision

• A typed-decision interface (choice, noul, score → distributions) with pluggable backends and a cascade: rules → System One server (Laya via Ollama/laya-serve, or Jev) → LLM with logprobs. The agent never knows which link answered; answered_by is on the result.
• Laya is the embedded open-source decision model: a compose profile pulls it; the CLI, the MCP server and the runtime's router use it when VERDICT_SYSTEMONE_URL is set.
• The kit also SERVES the protocol (createSystemOneServer) so a rules + LLM cascade can stand in for a decision model behind the same URL.
• Confidence is derived uniformly (distribution concentration; |2p−1|) so links compare; calibration (temperature, ECE, abstention threshold) is measured on the operator's own data.
• The create gate stays rules-only; a decision model may advise a reader on the model-handoff path, never loosen a gate where a false yes performs a write.

Consequences

Routing costs nothing with keywords, ~30 ms with Laya, one short model call only as a last resort. The vendor dependency is optional and replaceable by anything that speaks the protocol.

---

Source: docs/adr/0005-typed-decisions.md — edit it there; the site mirrors the repository.