model-router

Governed model and service routing for SocioProphet: local vs hosted, small vs large, cost, latency, quality, privacy, fallback, personalization, and eval-confidence policy.

Role

model-router decides where a request should go. It does not own model lifecycle, local model carry profiles, or per-user personalization consent.

Local + personal routing

The router now has a contract surface for routing between:

• base local model profiles from SourceOS-Linux/sourceos-model-carry;
• per-user personalization artifacts governed by SocioProphet/model-governance-ledger;
• higher-quality local fallback profiles;
• hosted fallbacks that require policy approval.

Contract and example:

schemas/local-personal-route-binding.schema.json
examples/local-personal-route-binding.llama32.json
tools/validate_local_personal_route_bindings.py

Agent execution routing policy

The router also owns the canonical policy for model use inside agent execution chains. The policy converts the operating rule into a machine-readable contract:

Use the cheapest lane that can safely complete the next irreversible decision.

The contract defines no-model, local-cheap, cheap, standard, high-end, and pro lanes; task classes; chain stages; escalation reasons; context limits; tool-use gates; and evidence requirements.

Contract and example:

schemas/agent-execution-model-routing-policy.schema.json
examples/agent-execution-model-routing-policy.default.json
tools/validate_agent_execution_model_routing_policies.py
docs/agent-execution-model-routing-policy.md

Budget and resource optimizer

Routing is also constrained by budget and live resource availability. The optimizer chooses the cheapest acceptable lane subject to policy, privacy, safety, quality floor, budget ceilings, local resource health, provider quota, provider health, latency, and deterministic verification availability.

Contract and example:

schemas/agent-execution-budget-resource-optimizer.schema.json
examples/agent-execution-budget-resource-optimizer.default.json
tools/validate_agent_execution_budget_resource_optimizers.py
docs/budget-resource-optimizer.md

Default local posture

The first SourceOS local profiles are:

urn:srcos:model-profile:local-llama32-1b
urn:srcos:model-profile:local-llama32-3b

The 1B profile is the laptop-safe router/triage/summarization default. The 3B profile is the quality fallback when local resources allow it.

Policy invariants

• Local-first routing is default.
• Prompt egress is denied by default.
• Hosted fallback requires policy approval.
• Per-user personalization requires consent and a model-governance-ledger contract.
• High-end/pro agent execution lanes are denied unless an allowed escalation reason and receipt exist.
• Routine execution de-escalates to standard or cheaper lanes after planning.
• Verification defaults to deterministic tools rather than another expensive model call.
• Budget, quota, resource, latency, and provider-health constraints are evaluated before route execution.
• Missing required optimizer signals fail closed.
• Unknown premium quota denies high-end/pro use.
• Exhausted budget denies or downgrades according to policy rather than silently overrunning.
• Evidence records route decisions, escalation receipts, budget decisions, resource snapshots, quota snapshots, candidate sets, runtime health, cost class, context policy, tool policy, and governance references.
• Prompt evidence should be hash-only by default.

Boundary

Repo	Responsibility
SourceOS-Linux/sourceos-model-carry	Local model profiles, service refs, local resource posture, and evidence collectors.
SocioProphet/model-governance-ledger	Per-user consent, data boundary, evaluation, promotion, revocation, model-routing escalation receipts, cost-class evidence, and budget/resource audit trails.
SociOS-Linux/socios	Opt-in orchestration for personalization workflows.
SocioProphet/model-router	Runtime route binding, agent execution model-routing policy, budget/resource optimization, and policy-aware target selection.
SocioProphet/agentplane	Execution-chain evidence and run/replay artifacts for routed agent work.
SocioProphet/guardrail-fabric	Fail-closed policy decisions for tool hooks, model-lane escalation, budget/resource constraints, and write/network gates.
SocioProphet/policy-fabric	Policy packaging, inheritance, validation, release review, and budget/resource constraint governance.

Validation

python3 tools/validate_local_personal_route_bindings.py
python3 tools/validate_agent_execution_model_routing_policies.py
python3 tools/validate_agent_execution_budget_resource_optimizers.py