Evidence / full-game protocol

Protocol leaderboard

Every result has its protocol attached: the model plays Black only in full games against Stockfish 18 with UCI LimitStrength. Read each cohort's evidence state separately; these local measurements are not a cross-cohort ranking.

Data cutoffAugust 8, 2026 at 8:00 PM UTC

Scope. Black-only model games · Stockfish 18 with UCI LimitStrength · 200-ply limit. W-D-L and scores include clean completed games only; provider-error records are excluded.

Ranked by joint local Elo

Luna

Publication authorization: not grantedTerminal · frozen

top2

Top 2

2 randomized, unranked proposals

Local protocol Elo2832Profile 95% interval: 2680–2992

Clean W–D–L41–11–8N=60 clean completed games · 77.5% score

Current opponent anchor2820N=20 · W-D-L 4–9–7 · 42.5%

Terminal data · frozenVisited 1320 · 1720 · 2120 · 2420 · 2820

top3

Top 3

3 randomized, unranked proposals

Local protocol Elo2431Profile 95% interval: 2317–2548

Clean W–D–L43–22–15N=80 clean completed games · 67.5% score

Current opponent anchor2520N=20 · W-D-L 5–11–4 · 52.5%

Terminal data · frozenVisited 1320 · 1720 · 1920 · 2120 · 2320 · 2520

top5

Top 5

5 randomized, unranked proposals

Local protocol Elo1830Profile 95% interval: 1713–1950

Clean W–D–L28–14–18N=60 clean completed games · 58.3% score

Current opponent anchor1920N=20 · W-D-L 4–7–9 · 37.5%

Terminal data · frozenVisited 1320 · 1520 · 1720 · 1920

top10

Top 10

10 randomized, unranked proposals

Local protocol Elo1127Profile 95% interval: 805–1407

Clean W–D–L2–2–6N=10 clean completed games · 30.0% score

Current opponent anchor1320N=10 · W-D-L 2–2–6 · 30.0%

Terminal data · frozenVisited 1320

regular

Regular

No engine proposals

Local protocol Elo1067Profile 95% interval: 705–1351

Clean W–D–L0–5–5N=10 clean completed games · 25.0% score

Current opponent anchor1320N=10 · W-D-L 0–5–5 · 25.0%

Terminal data · frozenVisited 1320

Non-comparable cohorts. Do not rank the cohort values against one another. These are local protocol-specific measurements, not FIDE, online-platform, or general model ratings.

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.

Shared vertical scale · interval bars show profile 95% · line joins point estimates only

Table / text view

This native table lists every cohort and variant represented by the curve data, including the separate regular interface. Frozen cohorts report local protocol Elo and profile 95% intervals; interim cohorts report working estimates and working 95% intervals. “Not published” identifies a missing estimate; “No interval or bound published” identifies missing uncertainty data.

Inverse candidate curve values for every public model cohort and variant.
Model / cohort	Variant	Candidate count	Estimate value	95% interval or bound	Estimate games	Estimate W-D-L	Evidence state
Luna	Regular interface (regular)	Separate interface	1067	705–1351	10	0–5–5	Frozen
Luna	Top 10 (top10)	10	1127	805–1407	10	2–2–6	Frozen
Luna	Top 5 (top5)	5	1830	1713–1950	60	28–14–18	Frozen
Luna	Top 3 (top3)	3	2431	2317–2548	80	43–22–15	Frozen
Luna	Top 2 (top2)	2	2832	2680–2992	60	41–11–8	Frozen
DeepSeek V4 Flash	Regular interface (regular)	Separate interface	973	467–1264	10	0–2–8	Interim working
DeepSeek V4 Flash	Top 10 (top10)	10	1248	998–1466	10	3–1–6	Interim working
DeepSeek V4 Flash	Top 5 (top5)	5	1390	1171–1618	10	5–1–4	Interim working
DeepSeek V4 Flash	Top 3 (top3)	3	1479	1276–1703	12	7–1–4	Interim working
DeepSeek V4 Flash	Top 2 (top2)	2	1852	1596–2190	14	12–0–2	Interim working

Top 101127Local protocol Elo

Profile 95%805–1407Frozen 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.

Protocol scope

Black-only
full games.

Opponent: Stockfish 18 with UCI LimitStrength
Model color: Black only; 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

Interpretation limits

Read each number with its boundary.

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

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

interim boundary — DeepSeek terminal data and publication authorization are separate states: terminal rows can be shown while the freeze-bound rating remains withheld until current release gates authorize publication.

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

Read the rating doctrine ↗Open the C005 rating boundary