CHAMPION ACCURACY

each step up is a coronation — a challenger provably beat the reigning model and took the crown

{{ kingPct }} {{ sinceGenesisTxt }} ♛ {{ coronationsTxt }} {{ duelsTotalTxt }} duels fought · {{ acceptRateTxt }} accepted

accuracy = exponential moving average over fresh research tasks · blocks ≈ 12 s each · exact values per coronation in the succession table below

LINE OF SUCCESSION

GENESIS {{ genesisPct }} base model

→

♛{{ c.n }}{{ c.hotkey }} {{ c.pct }} after crowning won by {{ c.lcbTxt }} vs {{ c.deltaTxt }} req. {{ c.when }} ·

THE ARENA

every duel, newest first — a challenger's 99.9% lower confidence bound (score) must clear the adaptive threshold (needed) to take the crown ▸ click any row's OUTCOME to expand per-task results & eval data · build v5 · live auto-refresh 30s

♛ crowned ◈ near miss — paid from arena pool ✕ lost duel ⊘ failed intake/probes ↺ stale parent

OUTCOME

ROUND

WHEN

CHALLENGER

CHECKPOINT

SCORE lcb

NEEDED δ

MARGIN

VERDICT IN

eval-data provenance — feed these to the replay tool and the recomputed digest must match the on-chain verdict

{{ d.replayHint }}

no per-task detail was recorded for this duel (produced before diagnostics were added)

no duel matches "{{ q }}" — check the full hotkey or the first 8 characters of your checkpoint digest

QUEUE {{ queueCountTxt }}

admitted challenges waiting for the next round — chain order, first revealed duels first

#

MINER

REPO

DIGEST

BUILT ON

REVEALED

WAITING

The queue is empty — every admitted challenge has been dueled. Submissions revealed now enter the next round.

MINERS

ranked by crowns won, then near misses, then best score — losing narrowly is paid, not punished

#

MINER

CROWNS

NEAR MISSES

ATTEMPTS

BEST SCORE lcb

ARENA CREDIT

LAST DUEL

THE MACHINERY

no human hands — intake, duels, verdicts and payouts run themselves

SUBMISSION PIPELINE · {{ funnelTotalTxt }} checkpoints submitted

queue right now: {{ queueDepthTxt }}

EMISSION SPLIT · effective vs configured

{{ e.effTxt }} / {{ e.confTxt }}

fresh-crown bonus ×{{ bonusFactorTxt }} · reign decay ×{{ decayFactorTxt }} per epoch

burn phase — this split is computed but nothing is paid out yet

VALIDATOR QUORUM

to crown≥ {{ thetaTxt }} of validator stake must accept bootstrapmin {{ bootstrapMin }} evaluators verdict timeout{{ timeoutTxt }} awaiting verdict{{ pendingCountTxt }}

{{ p.hotkey }} · revealed {{ p.when }}

VERDICT SPEED · {{ slaSamplesTxt }} verdicts sampled

median (p50)

p95

target: verdict within {{ slaTargetTxt }} — currently using {{ slaUsageTxt }} of that budget

CALIBRATION

harness noise floor{{ noiseFloorTxt }} threshold clamp≥ {{ deltaClampTxt }} LLM-judge reliance{{ judgeRateTxt }}

noise is measured by dueling the king against itself; the win threshold is clamped above it — verdicts are never coin flips. Scoring is programmatic; the LLM judge is a rare fallback.

RESEARCH TASKS · release {{ taskgenRelease }}

per duel{{ nPubTxt }} public + {{ nPrivTxt }} private tasks private poolepoch {{ poolEpoch }} · {{ poolTasksTxt }} tasks · next publish{{ poolRotationTxt }}

private pools are published in full when rotated — delayed transparency: you can't game tasks you can't see, and you can audit every task afterwards

REALITY CHECK · external benchmark, runs at every coronation

latest run{{ anchorAccTxt }} external · {{ anchorEmaTxt }} internal EMA divergence{{ anchorDivTxt }} runs recorded{{ anchorRunsTxt }}

no anchor run recorded yet — the first fires automatically at the next coronation

every new king is scored on a pinned public benchmark with the benchmark's own corpus behind the tools. Observational only — it can never grant a crown — so a reign whose exam score rises while its anchor stays flat is publicly visible here, and inflating the anchor buys nothing