Evaluation
A performance run is scored in two steps, in order: correctness first, then fuel. A fast wrong answer is no answer at all, so a solution earns a performance result only once it is known to be correct.
Correctness
Section titled “Correctness”The harness builds the solution with the manifest’s required
[build] commands, loads the wasm module, and
runs its entry function against each declared
[[case]] input, checking the returned output
against that case’s expected answer. A submission that fails to build, does not
export the contract’s entry point, exceeds the
sandbox limits, or
produces a wrong answer on any input is incorrect and earns no performance
score — correctness is a gate the solution must pass before its efficiency means
anything. (A correct answer produced just over the fuel ceiling is a distinct
“over the ceiling” outcome that still does not pass but is measured — see
Overshoot below.)
Smoke tests gate the stress cases
Section titled “Smoke tests gate the stress cases”The held-out set is run in two phases (set by each case’s
kind). The smoke tests run first: tiny
instances that each exercise one behavior in isolation, graded on correctness
alone (their fuel is not scored). Only if every smoke test reproduces its
expected answer do the stress cases — the large scored instances whose fuel
total is the result — run at all. If any smoke test fails, the stress cases are
not run and are counted as failed. This catches a broken solution in
milliseconds instead of after burning through the large instances, and it makes a
failure legible: the run’s Results tab shows which behavior the solution got
wrong, in its own section, before any fuel is spent. A run is correct only when
every case — smoke and stress — passed.
For a correct solution, the fuel consumed running the inputs is the
performance result. Fuel is wasmtime’s deterministic measure of work done: it is a
function of the code and its input, not of the host, so the same solution posts
the same number wherever it runs. Lower fuel is better — between two correct
solutions, the one that consumed less fuel did less work and is the better
implementation. Larger inputs dominate the total, which is the point: that is
where an O(log n) solution pulls decisively ahead of an O(n²) one (see
Overview).
Because the measurement is deterministic and reproducible, performance results are directly comparable across runs and models in a way wall-clock timings never could be.
Overshoot: the fuel runway
Section titled “Overshoot: the fuel runway”The fuel ceiling ([sandbox].fuel_limit) is the pass/fail line, but a solution
that only just misses it and one that misses it by 10× are very different, and
without help the harness cannot tell them apart — wasmtime traps exactly at the
ceiling, so an exhausted run reports no fuel at all. A case may therefore grant a
per-[[case]] fuel_runway: the solution is
allowed to keep running past the ceiling, up to fuel_limit * fuel_runway, purely
to get a reading.
This adds a middle outcome between pass and fail:
- Pass — correct answer produced within
fuel_limit. Its fuel is the score. - Over the ceiling — correct answer, but produced only past
fuel_limit(on the runway). It does not pass and earns no comparable score, but its consumed fuel is recorded as the overshoot, and its factory is still playable — so you can see exactly how, and how far, an inefficient-but-correct solution went over. The Results tab marks it “over ceiling” with the percentage over. - Incorrect — wrong answer (any fuel), or it exhausted even the runway.
The runway never moves the pass line; it only buys visibility into a failure. The multiplier is scaled down for larger inputs, because their verification is costlier per unit of fuel — so a small input can afford a wide runway (10×) while a large one keeps it tight (2×).
No human review
Section titled “No human review”A performance run is graded entirely by the harness. Correctness plus the fuel
number is not merely the decisive signal — it is the whole result. A performance
case therefore declares no scoring [[domain]] and no [[review_item]]
checklist, and a performance run carries no
review, rating, or writeup: there is nothing
for a reviewer to add that the bit-exact checksum gate and the fuel total have not
already settled. In the console a performance run’s detail page opens on a
Results tab (where a reviewed run shows its Verdict) carrying the recorded
correctness and fuel breakdown, and such runs never enter the unreviewed worklist.
Because fuel alone gives no sense of how good a number is, the Results tab places a correct run against the field: its rank and efficiency percentile among every model’s best correct run of the same case, version, and variant. The field is per-model-best — each model counts once, at its lowest fuel — so re-running a model does not skew the standing, but the run being viewed is placed as itself, so a slower duplicate still sees where it lands (and that its model already has a better run). The same per-model-best ranking drives the case’s Leaderboard tab, which for a performance case ranks models by the fuel of their best correct engine instead of by a reviewer score it does not have.