External validity / full-game ladder

A leaderboard
with its protocol attached.

One score is useful only when its measurement conditions travel with it. Luna is terminal and likelihood-rated. DeepSeek remains a clean-only live readout—not a finished Elo claim.

Evidence snapshotAugust 5, 2026 at 5:12 AM UTC

Ranked by joint local Elo

Luna

Terminal cohort
01
top2

Top 2

2 randomized, unranked proposals
Local protocol Elo283226802992 profile 95%
Clean W–D–L4111860 across all visited anchors · 78% score
Current anchor2820497 · 43% · N=20
Terminal · frozenVisited 1320 · 1720 · 2120 · 2420 · 28200 provider-error records excluded
02
top3

Top 3

3 randomized, unranked proposals
Local protocol Elo243123172548 profile 95%
Clean W–D–L43221580 across all visited anchors · 68% score
Current anchor25205114 · 53% · N=20
Terminal · frozenVisited 1320 · 1720 · 1920 · 2120 · 2320 · 25200 provider-error records excluded
03
top5

Top 5

5 randomized, unranked proposals
Local protocol Elo183017131950 profile 95%
Clean W–D–L28141860 across all visited anchors · 58% score
Current anchor1920479 · 38% · N=20
Terminal · frozenVisited 1320 · 1520 · 1720 · 19200 provider-error records excluded
04
top10

Top 10

10 randomized, unranked proposals
Local protocol Elo11278051407 profile 95%
Clean W–D–L22610 across all visited anchors · 30% score
Current anchor1320226 · 30% · N=10
Terminal · frozenVisited 13200 provider-error records excluded
05
regular

Regular

No engine proposals
Local protocol Elo10677051351 profile 95%
Clean W–D–L05510 across all visited anchors · 25% score
Current anchor1320055 · 25% · N=10
Terminal · frozenVisited 13200 provider-error records excluded

Candidate-width fingerprint

The curve runs backward.

Among candidate-assisted conditions, the observed point estimate rises as the visible set narrows. Select a point to inspect the uncertainty—not just the line.

Point estimates and 95 percent intervals for top 10, top 5, top 3, and top 2 candidate conditions. Interactive controls follow the same positions.500100015002000250030001127top 101830top 52431top 32832top 2
Shared vertical scale · interval bars show profile 95% · line joins point estimates only
Top 101127Local protocol Elo
Profile 95%8051407Frozen Davidson WDL likelihood
Estimate input2–2–610 clean games at the synthesis cutoff

Frozen Luna evidence. These protocol-specific estimates do not identify a candidate-count effect: sets were separately generated, games were unpaired, and anchors were adaptive.

Measurement contract

Same harness.
Visible boundaries.

Opponent
Stockfish 18 with UCI LimitStrength
Colors
LLM always Black; no color-balanced claim
Clock
0.1 seconds per Stockfish move
Game limit
200 plies maximum
Reasoning
high effort
Candidate display
randomized; scores and ranks hidden

How to read this page

Four boundaries prevent a number from becoming a myth.

04
01

These are local protocol-specific measurements, not FIDE, online-platform, or general model ratings.

02

Anchors were selected adaptively. Earlier blocks remain part of the frozen Luna likelihood.

03

DeepSeek has no published Elo until its suites are terminal and the joint likelihood is frozen.

04

Candidate conditions are complete agent protocols, not isolated estimates of base-model chess strength.

W-D-L and scores include clean completed games only; provider-error records are excluded.

Read the rating doctrine