import Link from "next/link"; import { ScopeTrace } from "@/components/ScopeTrace"; import { fmtParams, getRuns, latestPerTarget, limitOf, type Run, } from "@/lib/results"; export const dynamic = "force-static"; function Metric({ label, value, unit, tone = "ink", hint, }: { label: string; value: string; unit?: string; tone?: "ink" | "signal" | "reference" | "dim"; hint?: string; }) { const toneClass = { ink: "text-ink", signal: "text-signal", reference: "text-reference", dim: "text-dim", }[tone]; return (
{label}
{value} {unit && ( {unit} )}
{hint && (
{hint}
)}
); } function PrimaryCard({ run }: { run: Run }) { const v = run.verdict; const limit = limitOf(run); const perf = run.perf?.points ?? []; const single = perf.find((p) => p.concurrency === 1); const peak = perf.length ? perf.reduce((a, b) => (b.output_tps_total > a.output_tps_total ? b : a)) : null; return (
in production
primary target

{run.target.display_name}

{run.host.hardware} · {fmtParams(run.target)} ·{" "} {run.target.quant?.toUpperCase()} · {run.target.serving.engine} · ctx{" "} {run.target.serving.max_model_len?.toLocaleString()}

{perf.length > 1 && (

Throughput scales{" "} {peak && single ? `${(peak.output_tps_total / single.output_tps_per_stream).toFixed(1)}×` : "—"}{" "} from one stream to {peak?.concurrency ?? "—"}, but per-stream decode falls to {peak?.output_tps_per_stream.toFixed(1) ?? "—"}{" "} tok/s. Batch work and interactive work want opposite settings on this box.

full score card →
)}
); } export default function Home() { const runs = getRuns(); const board = latestPerTarget(runs); const primary = board.find((r) => r.target.tier === "production") ?? board[0]; const lastMeasured = runs[0]?.timestamp?.slice(0, 10); return (

Does this model earn
its place on our hardware?

Lumbridge Bench is the evidence layer for Lumbridge Compute. Two numbers decide a self-hosting call: whether a model is good enough at the work we actually do, and whether it is fast enough on the box we would run it on. Most leaderboards report only the first. Every card here reports both, measured together, on the same machine.

{board.length} target {board.length === 1 ? "" : "s"} {runs.length} run {runs.length === 1 ? "" : "s"} {lastMeasured && ( last measured {lastMeasured} )}
{primary && }

The board

latest run per target
{board.map((run) => { const v = run.verdict; return ( ); })}
target host signal reference tok/s ×1 peak tok/s measured
{run.target.display_name} {fmtParams(run.target)} ·{" "} {run.target.quant?.toUpperCase()} {run.host.id} {v.signal_score != null ? v.signal_score.toFixed(3) : "—"} {v.reference_score != null ? v.reference_score.toFixed(3) : "—"} {v.single_stream_tps?.toFixed(1) ?? "—"} {v.peak_throughput_tps?.toFixed(0) ?? "—"} {run.timestamp.slice(0, 10)}

Signal scores come from private tasks built from our own workloads. They are the measurement. The set is not published — a public test set gets scraped into the next training run and stops measuring anything.

Reference scores come from public benchmarks, unmodified. They are calibration, not a ranking: if one lands far from its published value, our harness is wrong, not the model.

); }