Overview

One canonical task spec per prompt, compiled into model-specific variants, measured, and routed by quality − λ·cost.

Models online
0/0
all providers keyed
Prompts
0
canonical task specs
Runs recorded
0
playground + eval executions
Total spend
$0
measured from provider usage

Pipeline

all prompts →
  1. Ingest — paste a raw prompt; Opus 4.8 extracts the canonical task spec (role, task, constraints, output contract).
  2. Render — five transformation rules compile the spec into per-model variants (XML for Claude, Markdown for GPT, flat for DeepSeek, bare for R1).
  3. Evaluate — test cases with programmatic checks + optional LLM judge produce a quality score per model.
  4. Route — quality − λ·cost under a quality floor picks the serving model and fallback chain.

Recent runs

fleet detail →
No runs yet — create a prompt and try the playground.