The Race & Scoring

A "race" is a timed window during which scenarios arrive on a shared conveyor and each lane processes them. This page covers what a race is, the run modes, and exactly how the score is computed.

Run modes

A race runs in one of six modes, each defining a duration and an arrival cadence (see MODE_CONFIG in src/engine/arrivalPump.ts):

Mode	Duration	Use
blitz	15s	max-frenzy clip
short	30s	social cut
standard	1 min	the default window
extended	5 min	long demo / narration
endless	∞	runs until you stop it; counts elapsed time up
sudden-death	90s	arrival rate ramps until one lane drowns

The mode sets durationSec and arrival: { intervalMs, maxQueue }. The arrival pump emits one scenario per intervalMs; when a lane's queue hits maxQueue, extra arrivals spill into its backlog (the belt jams).

You can switch modes from the lobby before GO, or from the preset switcher in the Stage (which restarts the race at the chosen window).

The controls

During a race, the Controls bar (src/stage/Controls.tsx) offers:

• Live timer — remaining M:SS (or elapsed ∞ for endless).
• Preset switcher (30s / 1m / 5m / ∞) — restarts the race at that window.
• ↺ Reset — return to the lobby.
• ⏹ End — score the race immediately.

There is deliberately no pause: the lanes are live agents, so freezing the clock can't freeze in-flight model calls. To halt a run, End it or Reset.

Pause/resume are real engine methods (GameEngine.pause() / resume()), not a synthesize-and-end shortcut.

How the score is computed

The headline formula (from src/engine/scoring.ts and src/shared/contract.ts):

score = Σ (partial × difficulty × BASE_POINTS)   over cleared items

where BASE_POINTS = 100.

• partial (0..1) — the fraction of the task's fields the agent got right, from the grader. A fully-correct answer is 1.0; a totally wrong one is 0.
• difficulty (1, 2, or 3) — the task's difficulty tier, which multiplies points and (intentionally) shows the speed gap more.
• A wrong answer scores 0 but still consumes the lane's time — the realistic penalty. Throughput falls out naturally: the faster lane simply clears more items in the window, so "throughput × accuracy" needs no separate term.

So a perfect difficulty-3 item is worth 1 × 3 × 100 = 300 points.

Resolved vs cleared

There are three related but distinct counts:

• cleared — items whose output matched ground truth (partial > 0), regardless of the operational verdict.
• resolved — items that were both output-correct and reached the correct verdict, and the case was resolvable (correctOutcome.pass === true). This is the honest "parcels handled correctly" number.
• verdictCorrect — whether the agent stamped the verdict the answer key said it should (accept / reroute / hold / refuse), independent of field accuracy.

The win banner and ROI card report resolved and verdict accuracy, not raw cleared — those are the numbers that reflect real operational quality.

The answer key

Every scenario carries a correctOutcome (src/shared/contract.ts):

interface CorrectOutcome {
  verdict: 'accept' | 'reroute' | 'hold' | 'refuse';
  pass: boolean;      // can this case be resolved at all?
  summary: string;    // one-line readout, e.g. "crushed → refuse"
}

This is first-class metadata, not inferred from the agent's output. It's what the focus card compares against to show ✓ correct / ✗ wrong verdict and should be: REFUSE. See Tasks & Grading.