Skip to the mode chooser
VERDICT/ EXECUTION MODES

SEVEN WAYS TO RUN A REQUEST

Choose how Verdict works.

Direct sends your request straight to your model and is the default. The other six modes add steps that check, repair or coordinate first. They are opt-in prototypes: more steps cost more tokens and time, and none has shown a proven gain over Direct. Here is what each cost in one measured snapshot.

Find your starting mode
BIGCODEBENCH · 100 TASKS · ONE SAMPLE PER MODEMEASURED
3.9×
Tokens used by the three experimental modes, against Direct

They also scored lower: 35–51 against 58.

Direct73.0K tokens
Experimental modes (average)285K tokens
Direct and the three standard checked modes landed within a few points of each other.56–59

One 100-task sample per mode on one local setup. A few points either way is sample noise, not a ranking.

See every mode, side by side

01 / YOUR PRIORITIES

What do you need most?

Start with your goal.
See the mode that fits.

RECOMMENDED STARTING POINT

Verdict Proof Mode

Answers directly first, then validates and recovers where the task family has useful checks.

Opt-in modepas Coding or structured output where a wrong answer costs more than extra time.
Where to select it

Verdict Chat, Verdict Server, Cloud Chat and iPhone: open the chat options, then Execution mode. In VS Code, set verdict.chat.executionArchitectureId.

Also consider Reflective Agent System (ras): Starts direct and spends repair effort only when the evaluation shows a failure.

02 / THE MEASURED COST

More steps. What do they cost?

One 100-task BigCodeBench sample per mode, on one local setup. Scores moved several points between samples (a nearby Direct sample scored 60; an older Verdict Proof Mode sample scored 53), so small gaps between the four standard modes are not a ranking. These are component measurements, not a promised gain on your work.

OPT-IN MODE · NO PROVEN GAIN · pas

1.97× Direct

Verdict Proof Mode

Answers directly first, then validates and recovers where the task family has useful checks.

  1. Prompt
  2. Direct-first answer
  3. Validation or recovery
  4. Answer
Score56
Tokens143.5K
Wall time24.7 min
Energy6.53 Wh

Best for: Coding or structured output where a wrong answer costs more than extra time.

WHAT WE ARE NOT CLAIMING

No checked-answer gains on this page.

Early small runs of the check-and-repair comparisons were cut off by their own token limit, so their numbers are not valid evidence and are not shown here. Until a clean run exists, this page claims no lift from checking or repair, and none of these modes has been shown to beat Direct. Retry with a test to check against is a separate opt-in feature, described on Advanced features.

Read about Advanced features

03 / WHERE TO SELECT

One setting, six places.

Pick a mode where you already work.
The default is Direct.

Verdict Chat

Chat options, then Execution mode.

Verdict Server dock

The Execution mode picker beside the model.

Cloud Chat

Chat options, then Execution mode.

VS Code

Set verdict.chat.executionArchitectureId.

iPhone

Chat Options, then Execution mode.

Command line

verdict-code --architecture direct | pas | cas | aas | gas

The command line accepts only direct, pas, cas, aas and gas, and cas, aas and gas may run as a dry run unless enabled. Fast Artifact System (fas) and Reflective Agent System (ras) are chosen in the apps.

THE PRINCIPLE

Start short.
Add steps where a wrong answer costs more.

Direct is the right default. A checked mode is worth trying when the cost of being wrong is higher than the cost of the extra steps, but we have not shown that it answers better. Keep the experimental modes for experiments.

Get Verdict