policy_ir
The routing language for a single model call: a six-element sigma-pol/v2 term.
The routing language for a single model call. A policy is a six-element JSON array: the tag plus five operational slots: filter, rank, select, mutate, fallback. You author the filter, rank, and select; mutate and fallback have stable defaults.
[
"policy",
["and", ["meets_req"], // filter
["not", ["is", "disabled"]],
["has_cap", "supports_tools"],
["cmp", "bench_intelligence", "ge", 0.5]],
["neg", ["normalize", ["field", "price_out"]]], // rank: cheapest first
["argmax"], // select: top survivor
["id"], // mutate: pass through
["always", { "action": "next_candidate" }] // fallback: next survivor
]The five slots
There is no filter/rank/select wrapper keyword. The tag plus five slots, defined purely by array position. (sigma-pol/v2 dropped the old evidence slot; it never affected a decision.)
1 · "policy"
The tag. Stamped with the grammar version,
sigma-pol/v2.2 · filter Pred
The gate: which candidates are eligible. Pure predicates. Anything that fails is out, not silently downgraded. A policy can only narrow the host floor, never widen it.
3 · rank Scorer
Score the survivors. A pure scorer orders the models that cleared the floor: cheapest-first with
["neg", ["normalize", ["field", "price_out"]]], or by intelligence, latency, or any weighted blend.4 · select Selector
Take the single top of the ranked list.
argmax picks the highest-scoring survivor; the cheapest only when the scorer ranks by price.5 · mutate Xform
id is the identity transform: prompt and params pass through; routing swaps the model, never your call.6 · fallback FailPlan
A declarative table over a closed action vocabulary, never user code. On any failure, move to the next candidate; every hop is written to the trace.