Evidence-first · protocol-sensitive

Read the protocol. Then the score.

Browse baselines, original-table bests, official-board leaders, and their evidence. Results enter one ranking only when their evaluation protocols align.

Browse 47 benchmarks →Read the alignment rules

47benchmarks

58tasks

345evaluation protocols

56,014source-located result rows

Coverage

22 benchmarks with public numeric results

The other 25 benchmarks keep auditable gap pages. Scores are never estimated to make the catalog look complete.

Numeric results available22

No public numeric result21

Restricted2

No standard numeric protocol2

Four claims, kept separate

Official baselineExplicitly designated by the organizer or benchmark paper.
Original-table bestBest only within the same source table.
Official-board leaderBound to one round and leaderboard snapshot.
Strict cross-work SOTAAllowed only when every critical comparison field aligns.

Find a result

Start from a benchmark or a task

Browse benchmarks

Inspect coverage, tracks, protocols, baselines, and qualified leaders.

Browse tasks

Find evaluations by molecule, output representation, and ML formulation.

Immutable snapshot

2026-08-06

Literature cutoff 2026-08-06. Future updates create a new dated path instead of silently rewriting this snapshot.

View data interfaces