One task
You describe the change and the proof you expect.
Smart Harness
Don't ask the model twice.
Ask for proof once.
Smart Harness catches candidates that look finished but fail your checks, gives the model a bounded chance to recover, and applies the verified winner. You see every attempt, every token, and the final verdict.
Opt-in · bounded policies · fail-closed outcomes · complete attempt metering
01
candidate generated
ready02
verifier: 1 test failed
retry03
bounded recovery attempt
metered04
7 / 7 tests passed
verifiedVerified rescueWinner applied once · receipt attached
What it does
A small loop with a hard boundary
The public contract is simple. Your checks—not the model's confidence—decide when the task is done.
Bounded attempts
The selected policy controls how far recovery can go.
Your verifier
Tests or checks decide whether a candidate is acceptable.
Truthful outcome
A verified winner is applied once—or the result says verification failed.
No secret sauce diagram. Just the customer-visible contract: bounded work, independent checks, and an honest result.
Same model. Same task.
The difference is what happens after the first answer
Step through a realistic boundary-bug repair. The left side keeps going when the verifier finds evidence; the right side hands that loop back to you.
Interactive task replayStep 1 of 5
Bounded verification loop
Smart Harness
Task accepted
Fix the inclusive-date billing bug
Same repository, same model, same failing test, and the same definition of done.
1 failing · 6 passing
Same model · no harness
Single-pass agent
Task accepted
Fix the inclusive-date billing bug
The single-pass agent starts from the identical prompt and repository state.
1 failing · 6 passing
Controlled benchmark evidence
More tasks ended verified
On a controlled 1,140-task BigCodeBench run using production infrastructure—not organic customer traffic.
627 of 1140 passed
817 of 1140 passed
190 more verified completions
billing-invariant violations
Controlled benchmark result. It does not predict the lift on every repository, model, policy, or task.
Real product evidence
The result comes with receipts
These are direct captures from a real Verdict Desktop code-task completion. They show the finished workspace and final response receipt—not a Smart Harness A/B screenshot.
Honest accounting
Quality lift is not free compute
A failed first candidate can trigger more calls. Smart Harness makes that trade visible instead of hiding it behind a success animation.
Retries are bounded
Your selected policy caps the number and kind of additional attempts.
Every attempt is metered
Tokens, wall time, and the relevant cost or energy signal roll into the outcome.
No blanket savings claim
The proven benefit is more verified completions and less manual retry choreography—not universal token or time savings.
Questions, answered
The short version
Is Smart Harness another model?
No. It is a verification and recovery layer around the model and task you choose.
Does it retry forever?
No. Every policy has a bounded attempt budget. You can choose a faster or more thorough policy, and the outcome reports what actually ran.
Does it always save tokens or time?
No. A failed candidate can trigger extra attempts. The gain is a higher rate of verified completions and less manual retry choreography; every attempt is metered.
Where is the full engine available?
The repository-aware engine runs in Verdict Code and in Verdict Desktop portable coding tasks. Other surfaces may expose catalog controls without running the full code-task engine.
Stop shipping the first plausible answer
Give the model a finish line it can prove.
Start locally for $0. Turn on Smart Harness when the task deserves a verified result.