Add objective Overall / Value / Capability score columns to the models table

[huncho-tensei]

Summary

Adds three sortable, transparently-computed score columns to the models table — Overall, Value, and Capability — plus a dynamic rank (#) column that renumbers with the current sort.

The goal is to let people rank the whole catalog by objective criteria without leaving the table they already use. Scores are derived entirely from existing catalog fields — no benchmarks, no hand-grading, no external data.

New column layout:

#  |  Provider  |  Model  |  Overall  |  Value  |  Capability  |  … (all existing columns unchanged)

The table defaults to Overall, descending. All three score columns sort like every other column, so users can switch lens with one click.

How the scores are computed

Four normalized 0–100 components are calculated from objective fields, then blended with weights kept in one documented place (packages/web/src/score.ts):

Component	Source fields	Direction
capability	tool_call, reasoning, structured_output, temperature + input/output modality breadth	higher = better
cost	blended cost.input + cost.output ($/1M), log-scaled then inverted	cheaper = better; free = top; unknown = neutral 50
context	limit.context + limit.output, log-scaled	higher = better
recency	release_date	newer = better

Each component is min-max normalized across the whole dataset (a missing/unparseable field collapses to a neutral 50, so it never silently wins or loses). The three lenses are just different weightings:

Lens	capability	cost	context	recency
Overall	0.40	0.30	0.20	0.10
Value	0.35	0.50	0.10	0.05
Capability	0.60	0.15	0.20	0.05

The weights are the only opinion in the change and are isolated in a single WEIGHTS object so they're trivial to audit or tune.

What the score does — and does NOT — measure

This is important and stated up front: the catalog has no quality/benchmark field, so the score cannot and does not measure model "intelligence." "Capability" here means breadth of declared features and modalities, not how good a model's outputs are.

A direct consequence, visible in the live data: broad, cheap, omni-modal models (and meta/auto-routers, which declare every modality at low listed cost) rank at the very top, above expensive frontier models. That is correct given these inputs — it's a spec-breadth-per-dollar ranking, not a smartness ranking.

This is also the reasoning behind shipping it as sortable columns rather than one decreed ranking: the data stays neutral, and the user chooses the lens that fits their use case. If a future schema ever adds an objective quality signal, it drops straight into the existing component blend.

Scope

• Web package only. The canonical api.json / core data is unchanged — scoring is a presentation-layer concern and does not pollute the data API.
• Files touched: score.ts (new), shared.ts, render.tsx, index.ts, index.css.

Validation

• bun validate passes.
• cd packages/web && bun run build succeeds; rendered HTML contains all new columns and computed scores across the dataset.
• Sort invariant verified: 28 sortable headers ↔ 28 sortValues entries.
• Score distribution sanity-checked (spans ~22–90; image/TTS models correctly rank lowest).

Happy to adjust the weights, drop to two columns (Overall + Value), or gate this behind discussion if a built-in ranking isn't a direction you want — feedback welcome.

[huncho-tensei]

Rationale and design discussion in #1893.