deterministic AI specification compiler

Describe a problem. Compile a verified AI spec.

UPEC turns a structured problem model into a validated, explainable, target-specific AI specification — with lint diagnostics, a deterministic quality score, and an execution contract that states exactly what the target model may do.

source_hash
reproducible
quality_score
0–100
contract
enforced

Illustrative compilation example — not live output

POST /api/premium/compile200 compiled

quality_score

92

readiness=Readydone=Done
COMPILED AI SPECIFICATION
target=coding_agent accuracy=verified risk=low

ROLE
Operate as a backend-systems agent.
Resolve the problem to its verified done-state.

STRATEGY
  1. Normalize universal problem model
  2. Route intents to specialists
  3. Assemble tool contract
  4. Render acceptance gates
build_fingerprint
a1f3.9c22.7de1
contract: Ready

the pipeline

From a raw problem to a verified specification

Every compile runs the same deterministic stages. The same problem model always yields the same source hash, prompt hash and build fingerprint — so you can diff specs the way you diff code.

Click any stage to inspect what UPEC does there.

execution contract

A specification that knows its own boundaries

UPEC resolves every tool against the target model's capabilities and the granted permission set. The compiler refuses to certify a build that asks a model to do something it isn't allowed — or able — to do.

  • Per-tool capability and permission resolution
  • Target-compatibility checks before execution
  • Permission-consistency verdict across the request
  • Machine-readable reasons for every blocked tool
execution_contractstate=blocked

code_search

perm=read_only

ready

apply_patch

perm=edit

ready

run_migration

perm=deploy

blocked
PERMISSION_MISSING — run_migration requires 'deploy'

Ship AI specifications you can actually trust

Start compiling in the browser. Wire the same API into your pipeline with an organization-scoped key when you're ready.