The System One decision layer — offer ranking, escalation policy, a confidence gate
(Jev/Laya), and a typed PolicyGuard wrapper. Used by shop-via-ucp
when it's available, in preference to raw MCP calls.
The System One decision layer for portage-ucp — see
docs/plans/system-one-decision-layer.md for why this exists.
Sits between the agent loop and the Adapter/client layer, turning judgment
calls that were scattered across skill instructions, portage-cli, and
portage-ucp into typed, inspectable decisions:
Portage::Ucp::Decision::OfferRanking — which offer to pick among several.
A typed wrapper around Portage::Ucp::Support::OfferRanking.
Portage::Ucp::Decision::EscalationPolicy — hand off vs. keep going. A
typed wrapper around Portage::Ucp::Support::Escalation.
Portage::Ucp::Decision::ConfidenceGate — proceed unattended above a
threshold.
Portage::Ucp::Decision::PolicyCheck — a typed wrapper around
Portage::Ucp::PolicyGuard.check!.
The ranking, escalation and policy rules live in portage-ucp core, and
these three wrap them as Verdicts. portage-cli calls the core modules
directly, so it answers the same way with or without this gem. Only
ConfidenceGate and its model backends are unique to this gem.
ModelBackends::Laya needs a Python bridge script (it's HuggingFace weights,
not a hosted API) — see examples/laya_bridge.py and set
LAYA_BRIDGE_SCRIPT; a missing or broken bridge raises
BackendNotConfiguredError/BackendError with a message naming the exact
problem (unset, path doesn't exist, LAYA_PYTHON not executable, non-zero
exit, timeout, unreadable reply). The bridge must answer with the same
{"answers": {name => answer}} body Jev returns, each answer keyed by its
type: {"type": "noul", "noul": <P(yes)>}, `{"type": "choice", "choice":