unhardcoded
reference · format

Operators & fields

The sigma-pol/v2 operator signature and the catalog field vocabulary.

The core sigma-pol/v2 operators, by slot; the complete signature lives in the engine's sig.lua. Field names come from GET /x/fields; use a real field, e.g. price_out, not price.

Filter, predicates

and · or · not
Boolean combinators.
meets_req
The request's own implied requirements, auto-derived from the call: tool calls require tool support, image inputs require vision, a JSON response format requires structured output.
has_cap <name>
Require a capability flag, e.g. supports_tools, supports_json_mode.
is <flag>
A boolean model field, e.g. disabled, no_log, has_tee, cap_tools, cap_reasoning, in_image. Image and reasoning gates use is (in_image / cap_reasoning), not has_cap.
cmp <field> <op> <value>
A numeric bound; ops ge · le · eq · ne · lt · gt. Compares a catalog field to a constant.
family_eq <name>
Keep only candidates in one model family. Pin a step to a single provider or model line.

Rank, scorers

field <name>
A raw catalog field as the score.
normalize
Scale a sub-score to 0–1.
neg
Invert, for cheapest or lowest-latency.
scale <w> · add
Weight a sub-score, and sum weighted sub-scores into one ranking.
zero
A constant score, used when the filter already leaves a single survivor (e.g. a family_eq pin).

Select · mutate · fallback

argmax
The single highest-scoring model. sample <t> is a seeded stochastic pick (e.g. ["sample", 0.5]) for ensemble diversity.
top_k <n> <sel>
Keep the top n as an ordered failover cascade.
id · clamp_param
Mutate (Xform): id passes the request through unchanged (the default); clamp_param bounds a parameter on the chosen call.
always {"action":"next_candidate"}
Fallback (FailPlan): on a provider failure, advance to the next survivor; override maps a specific failure reason to a different action.

Catalog fields

bench_intelligence
Intelligence score on a 0..1 scale. The canonical quality floor is ge 0.5.
price_out · price_in
Output / input price. Cheapest-first via ["neg", ["normalize", ["field", "price_out"]]]; ceilings like le 5, le 1, or le 0 (free only).
context
Context-window size, e.g. a long-context floor ["cmp", "context", "ge", 200000].
latency_ms · success_rate
Latency in ms (filter le 2000) and a reliability score, both used in filters and scorers.
bench_agentic · bench_coding
Benchmark scores, with leaderboard ranks bench_agentic_rank, bench_coding_rank (le 5 = top-5).
supports_tools · supports_json_mode
Capability flags read with has_cap.
cap_tools · cap_reasoning · in_image · no_log · has_tee · disabled
Boolean fields read with is.

Plus more input-modality flags (in_audio, in_file, in_video, out_image), capability flags (cap_seed, cap_tool_choice, cap_parallel_tools, …), and the bench_arena score with its rank. GET /x/fields returns the complete, live list for your host.

← Back to docs