Vibe
Experiment ledger · 2026-07-10T13:43:17-07:00
Hive engine research · live status

Vibe strength and iteration dashboard

A source-backed view of playing strength, correctness, and the next action in the NNUE engine flywheel.

Current

Implement incremental NNUE accumulators and expand the calibrated match pool.

Next: Run equal-time v0.2-vs-v0.1 confirmation, then generate deeper v0.2 self-play for the next residual.

Estimated Elo
470Source: dashboard/elo_estimates.json Bradley-Terry pool
95% interval -170–1110; 68 games. Nokamute d7 is fixed at 1800.
Current champion
v0.2.0Source: dashboard/experiments.json latest experiment
Promotion requires paired-match evidence.
UHP rule suite
21/21Source: Nokamute uhp-debug output
Deterministic interoperability cases passing.
Fuzz positions
52,100Source: Nokamute uhp-debug random differential search
Random Base+MLP positions with bidirectional move agreement.
Search throughput
348k n/sSource: target/release/vibe bench 4
Depth 4 in 0.493s on the fixed midgame fixture.
Replay positions
6,866Source: data/selfplay/stackfix-depth3-v0 generated shards
Self-generated positions only; no external play data.

Estimated Elo by promoted iteration

Absolute scale anchored to Nokamute depth 7 = 1800; bands are 95% intervals.

-500 1350
No point is drawn until actual paired games produce an estimate. Candidate-vs-champion deltas drive promotion; the absolute pool is for orientation.

Experiment ledger

Newest first; exact evidence source is retained for every row.

EngineEloGamesResultSource
v0.2.047040Promoted: 29.5/40 (73.8%) vs v0.1 at equal depth; +179 Elo direct (approx. 95% +59 to +299)artifacts/matches/candidate-blend25-vs-champion-seed20260717.json + models/current.nnue
v0.1.0-nnue-stackfix—20Not promoted alone: 10/20 vs Nokamute d1; retained as a 25% residualartifacts/matches/candidate-stackfix-small-vs-nok-d1-seed20260715.json
v0.1.0-nnue-v2—20Rejected: 5.5/20 (27.5%) vs Nokamute d1; postmortem found stack-order feature corruptionartifacts/matches/candidate-v2-small-vs-nok-d1-seed20260714.json
v0.1.0-nnue-small—20Rejected: 3.5/20 (17.5%) vs Nokamute d1; game-held-out validation 0.316artifacts/matches/bootstrap-small-vs-nok-d1-seed20260713.json
v0.1.0-nnue-b0—20Rejected: 5.5/20 (27.5%) vs Nokamute d1; 254k n/s; 4,472 shallow training positionsartifacts/matches/bootstrap-nnue-vs-nok-d1-seed20260712.json + models/bootstrap.nnue
v0.1.03472017/20 points (85%); provisional +301 Elo relative to Nokamute d1, absolute calibration pendingartifacts/matches/vibe-d3-nok-d1-seed20260711.json
v0.1.034720 wins, 1 draw, 1 loss vs Nokamute d1 at equal depth; 25% score, sample too small for ratingartifacts/matches/d1-vibe-white.log + d1-vibe-black.log
v0.1.03470Rules baseline; depth-4 search 0.352s after make/unmake + TT; Elo calibration pendingcargo test + Nokamute UHP suite + target/release/vibe bench 4

Metric definitions

Estimated Elo is fitted from paired-color game outcomes. UHP cases are deterministic protocol/rule fixtures. Fuzz positions count only completed positions where both engines accept every move from the other and agree on move count.

Methodology

Every candidate must pass correctness and throughput gates before paired matches. Nokamute d7 is assigned 1800 exactly; lower settings and Vibe iterations are inferred relative to that anchor.

Limitations

There is no Elo estimate until calibration games complete. This explicit empty state prevents implementation progress from being mistaken for measured strength.

Next step

Run equal-time v0.2-vs-v0.1 confirmation, then generate deeper v0.2 self-play for the next residual.

Decision queue

After the first end-to-end replay shard, compare 128- and 256-channel queen-relative NNUEs at equal wall time.

Freshness. Ledger updated 2026-07-10T13:43:17-07:00. This page reloads every 30 seconds while open.