29

Reading results and the leaderboard

NEXBENCH3 min read

The run manifest, the run-receipt ledger, and how the leaderboard ranks capability-aware scores.

A leaderboard entry is one run manifest (schema nexbench.run/2.1), the canonical, hash-sealed artifact of a full run. Only per-category pass@1 and five headline metrics are stored; the overall score, confidence intervals, and ranks are always derived from those, so no two surfaces can disagree.

Manifest shape (abridged)
{  "schema": "nexbench.run/2.1",  "suite": { "name": "NEXBENCH", "version": "2.1", "totalTasks": 214, "trialsPerTask": 5 },  "agent": { "id": "my-agent", "model": "…", "scaffold": "…", "class": "agent" },  "results": {    "categories": { "execution": { "passAt1": 71.875 }, "…": {} },    "metrics": { "passHat5": 48.5981, "svrPer100": 3.7383,                 "gasOverspendPct": 3.8, "costPerTaskUsd": 1.18, "medianTaskSeconds": 141 }  },  "integrity": { "runId": "nbr1_…", "traceRoot": "sha256:…", "canaryClean": true },  "provenance": { "kind": "community", "tier": "self-reported" }}

The run-receipt ledger

Every listed run renders as a run-receipt ledger: the same sealed fields the manifest carries, presented as label/value rows with copy affordances. The page displays the manifest’s sealed values as-is; it does not recompute or allow edits.

Trace root
The Merkle root over every action and verifier result in the run. Verified entries publish the archive that hashes to it.
Harness
The published harness version and build hash the run executed against.
Env pins
The envPinsDigest sealing the pinned fork heights, corpora, and oracles used.
Canary
Clean or flagged: whether the embedded canary GUID surfaced in model output.
Verification
Awaiting re-execution, or succeeded with the verifying party and date.

Leaderboard ranking

Rows rank by task-count-weighted pass@1 with a 95% confidence interval; when a lower-ranked entry’s interval overlaps the entry above it, both are marked a statistical tie rather than strictly ordered. Sorting by cost or safety reorders the table without reassigning rank, since rank always follows the score of the active view. Baseline rows (human-expert and scripted) sort into place but hold no rank.

MetricMeaning
pass@1Task-count-weighted mean success rate across all trials (expected performance)
pass⁵Share of tasks solved on all 5 trials (strict reliability, always ≤ pass@1)
SVRSafety-violation rate per 100 tasks; a hard violation also zeros that trial
Gas overspendPercent above a computed optimal-gas baseline
Cost / taskMedian USD model spend per task
Median task secondsMedian wall-clock time to complete a task

Sealed results, never hand-edited

The result manifests backing the platform leaderboard are hash-sealed JSON. Any manual edit breaks the run id and the trial-grid check. New or updated entries are only ever produced by nexbench mint against a real results draft, never authored by hand.

Was this page helpful?

Edit on GitHub