Scoring and Stars

Scoring and Stars

Two metrics, in tension

A pack declares exactly two metrics, never a four-term composite. The tension between them is the game design.

Metric	Direction	Definition
cost	minimize	The sum of turn costs, reported against the pack's do-nothing baseline
reliability	maximize	Delivered relief over MW-hours asked, during order turns

Delivered relief is nominal grid draw minus actual grid draw, capped per turn at the amount asked. Buying yourself out of an obligation you were never given earns nothing.

The two pull against each other. The cheapest day ignores the order. The most reliable day overpays. Every interesting decision in the game lives between them.

Every number ships with its denominator

A cost of $99,144 means nothing on its own. The settlement reports it against the do-nothing baseline, which is the same day bought entirely from the grid at nominal draw. During play the comparison is against the baseline so far, because a full-day baseline is useless at turn three.

Reliability is reported the same way: delivered over asked, with both numbers visible.

Stars

Where an optimum can be certified, your result is ranked against the true best play rather than against other players.

Tier	Rule
Blue Ribbon	The optimal line itself, within one cent
Gold	Within one percent of the optimum
Silver	Within two percent
Bronze	Within three percent

A scenario whose optimum cannot be certified awards no stars and says so.

Where the optimum comes from

Offline, by exact optimization. The pack is solved as a mixed-integer linear program, and the solution is then replayed through the same engine a player uses, to the penny, before publication.

Both halves matter. A solver's answer is a hypothesis about the rules; replaying it through the engine is the check that the hypothesis was about the right rules. A solution that cannot survive replay is a bug in the model, not a score.

This is also why the engine's arithmetic is kept linear. The battery's efficiency convention, described in The Engine, exists in the form it does so that an exact solver can certify an answer.

Replay is contractual

score(replay(episode_log)) === stored_score

The episode log is downloadable from any settled episode, the scoring function is published and versioned, and the equality above is enforced by an automated test that ships with the code.

The purpose is to make disagreement productive. Arguing with the scoring function is the instrument working, provided the argument is against something fixed, inspectable, and versioned.

Versioning

score_version is pinned into an episode when it is created. Changing the scoring function does not restate old results, and it cannot: the old episodes carry the version they were scored under.