Builder AI
Produces code, a document, image, interface, dataset, scene, or mixed package.
AI pipeline integration
Your AI did the work. Check it.
Put TUC after the builder and before the work is accepted. Send the completed artifact, original task, requirements, and evidence. The next step gets a verdict it can act on.
TUC
Checks the submitted work against its task, requirements, constraints, and evidence.
Decision
pass, fail, or insufficient_evidence.
Next AI step
Repair a supported finding, collect missing evidence, or release the work.
The contract
Keep the builder. Add one independent decision.
Artifact
The completed work—not a description of what the agent hoped to make.
Intent
The original task, requirements, constraints, success criteria, and references.
Evidence
Files, captures, test output, hashes, or other proof needed to support a decision.
Result
A verdict, findings, recommended fixes, and verification steps in JSON.
Minimal handoff
One request. Three explicit branches.
Python · synchronous evaluation
import os
import httpx
response = httpx.post(
"https://ultimate-critic.fly.dev/api/v1/evaluate",
headers={"X-API-Key": os.environ["TUC_API_KEY"]},
json={
"artifact_type": "code",
"artifact": {
"content": generated_code,
"metadata": {"builder_model": builder_model},
},
"intent": {
"prompt": original_task,
"requirements": requirements,
"constraints": constraints,
},
"domain": "software",
"evaluation_profile": "code_review",
},
timeout=30,
)
response.raise_for_status()
result = response.json()
match result["overall"]["verdict"]:
case "fail":
repair_agent(result["findings"])
case "insufficient_evidence":
collect_evidence(result["metadata"]["evidence_issues"])
case "pass":
release(result["evaluation_id"])Route supported findings and their verification steps to the repair stage.
Collect the missing proof instead of treating uncertainty as a pass.
Let your own policy decide whether the work can move forward.
Try from a terminal
Run the handoff before you buy.
No key · Python standard library only
curl -fsSLO https://ultimate-critic.fly.dev/tuc-preview-gate.py python tuc-preview-gate.py generated.py \ --type code \ --task "Reject unsafe dynamic execution." \ --requirement "Do not call eval" # exit 0 pass # exit 1 fail # exit 2 insufficient_evidence # exit 3 request or input error
Download the gate and point it at one text, code, or JSON file of up to 10,000 characters.
The public route uses fixed checks, calls no external model, and does not save the submitted input or result.
The file is sent to the public TUC service for processing. Do not submit credentials, secrets, private source, or sensitive customer data.
Boundaries
Add judgment without replacing the rest of the stack.
Keep your builder
TUC receives the output after generation. It does not choose or operate the upstream model.
Keep your tests
Test output can be supplied as evidence. TUC does not turn a failed test into a passing result.
Keep human approval
A TUC verdict is an input to your release policy, not a guarantee or a substitute for accountable review.
See it work
Two checked repair loops.
Unsafe code → repaired code
A reproducible sample webhook fails on unsafe execution, then passes after authenticated JSON parsing.
Logo concept → production package
TUC’s own concept fails five production requirements; the repaired six-variant identity package passes.
Founding access
Connect one real pipeline.
$249 once includes one tenant key, 90 days, 1,000 review credits, and up to two hours of first-integration support. Standard checks use 1 credit, extended packages use 5, and full mixed-package reviews use 10. A full review is not limited to a diff.
$249
one-time · no automatic renewal