Method

Standardization

For each measure m with raw figures x over the models it lists:

z = (x − mean) / sd, score = clamp(50 + 15·z, 0, 100)

A measure with fewer than two figures is dropped: there is nothing to standardize against.

Confidence

Where a source publishes an error bar e and the measure’s spread is r = max − min:

confidence = clamp(1 / (1 + 16·(e/r)²), 0.15, 1)

No error bar means confidence 1. Absence is not treated as evidence of a wide one.

Weighted mean with shrinkage

For a scope (Overall, a category or a domain) with measures i, weights wᵢ, confidences cᵢ and standardized scores sᵢ:

prior = PRIOR_FRACTION · W_scope, with PRIOR_FRACTION = 0.25 and W_scope the sum of weights available in the scope

score = (Σ wᵢ·cᵢ·sᵢ + prior·50) / (Σ wᵢ·cᵢ + prior)

Missing measures are excluded from the sums, never imputed.

Weights

Board	Overall share	Also shapes
LMArena Text	25	Human preference
Artificial Analysis Intelligence Index	25	General intelligence
LiveBench	20	Reasoning, Coding, Mathematics, … (8 measures)
Terminal-Bench	15	Agentic coding
ARC-AGI-2	15	Abstract reasoning
GDPval-AA, AA-Briefcase	—	Professional tasks, Agentic knowledge work
Any report ✱ measure	—	its domain, at 10 (a board measure carries 25)

Eligibility

• Overall: scored by boards from ≥ 2 independent publishers. Publishers, not boards: Artificial Analysis’s three boards are one voice.

• Category / domain: any model a recognised board has measured in that scope. Reports alone never rank.

Anchored estimate

For a model with no Overall score: on every measure it shares with models that have one, its figure is placed among theirs and their Overall index is read at that position (linear between neighbours). Outside their range the nearest anchor is a bound, shown as ≤ or ≥. Placements are averaged by measure weight, discounted where fewer than six anchors support them. Shown as ~55✱; never a rank.

Identity

Sources spell one model several ways. Names are reduced to a bag of tokens with reasoning-effort tiers, run settings, vendor prefixes and board badges removed; the highest published tier per source is indexed and the exact label kept. Every merge is logged in the build report.

Data contract

index.json — generatedAt, benchmarks[] (id, name, publisher, url, retrievedAt, metric, category, domain, kind, source, group, caveat?), models[] (id, name, org, access?, size?), scores[] (modelId, benchmarkId, raw, stderr?, sourceLabel, scaffold?), excluded[], catalogs[]. The same shape is validated by the pipeline before a snapshot ships.