> Canonical page: https://posttrainllm.com/benchmarks/arena

Candidate arena · 2 rating families · 0 hidden conversions

# Rate the match. Show the doubt.

Head-to-head games earn Arena Elo. Independent games earn paired score evidence. Neither becomes a universal intelligence number.

Rating console

## One arena, honest vocabularies.

 Chess 2048

 Competition family Head-to-head

 Rating Arena Elo

 Evidence 4 games · 2 forfeits

 1200 1300 1400 1500 1600 1700 1800

 1585 Qwen3 4B unrated

 1415 Qwen3.5 9B unrated

The 170-point gap is a regularized diagnostic from four games—not a qualified rank. Two 4B wins came from 9B notation forfeits. This pool-specific Arena Elo is not FIDE, human, engine, Chess.com, or Lichess Elo.

 Qualification UNRATED

- Total games 4 / 30
- Per policy 4 / 10
- Forfeit rate 50% / ≤10%
- Color balance 2 white · 2 black
- Connected pool yes [Inspect every chess move →](https://posttrainllm.com/benchmarks/chess)

 Competition family Independent episodes

 Rating Paired score

 Why not Elo No interacting opponent

 −20 −10 0 +10 +20 +30

 +10.7 Claude Opus 4.8 unqualified

 -1.0 Claude Sonnet 5 unqualified

Paired delta compares each model with random legal play on the same seed. The interval crossing zero means no demonstrated advantage.

 Qualification UNQUALIFIED

- Claude Opus 4.8 3 / 30 pairs
- Claude Sonnet 5 4 / 30 pairs
- Universal Elo not applicable [Inspect every 2048 run →](https://posttrainllm.com/benchmarks/game-2048)

Evidence ledger

## A number is the last column, not the first.

 candidate-only-not-a-public-rating

 01 · Match evidence

### Arena Elo

Only two interacting policies, win/draw/loss outcomes, one frozen pool, and match-level uncertainty.

 02 · Score evidence

### Paired delta

Single-player policies share seeds with a baseline. Mean score and uncertainty stay in the game’s own units.

 03 · Qualification

### Estimate ≠ rating

The math can produce a diagnostic before the sample can support a public rank. We show both states.

Extension contract

## Add a game without rewriting the arena.

- 1 Choose a family. Head-to-head result or independent paired score.
- 2 Normalize evidence. Participant identity, outcomes, failures, trace hashes, and source paths.
- 3 Freeze qualification. Sample, balance, connectivity, and failure limits before scoring.
- 4 Reuse the report. The validator, uncertainty engine, and dashboard remain unchanged.

 Download arena evidence [Browse benchmark archive](https://posttrainllm.com/benchmarks)