zero-shot-ie-bench: GLiNER 2.5, GLiFormer, GLiClass, Laya, Von and Jev in one web UI

# zero-shot-ie-bench

Zero-shot information-extraction and classification benchmark: a curated
catalog of decision, classification, NER and reranking systems — hosted
APIs and local runner islands — benched against versioned, multilingual
suites, with an append-only run history and generated result tables.

- **Catalog** (`catalog/`): one YAML per system; providers, tracks,
  languages, machines. Schema-validated (`bun run validate`).
- **Suites** (`suites/`): versioned parallel-item pools with content
  hashes. Canary suites are the cheap scheduled nerf-detector.
- **Runs** (`runs/`): append-only; every table below is generated from
  them (`bun run lake && bun run render`). Rebenching adds history; a
  nerf or improvement shows up as drift.
- **Bench** (`bench/`): TypeScript + Bun — provider adapters (one per
  API contract), scorers per track, `run | canary | doctor`.
- **Runners** (`runners/`): one independent Python/uv project per
  dependency island, all speaking one HTTP protocol; a machine that
  benches needs Bun next to uv, nothing else.
- **Board / Playground** (`apps/`): static decision board and a
  try-it UI for hosted + live-local models.

Quickstart: `bun install` · `bun run check` (typecheck + tests + lint) ·
`bun run render` (lake + README tables) · `bun run --cwd bench dev
doctor`. Deeper docs: [methodology](docs/methodology.md) (scoring,
deferred runs, cost model) · [adding a model](docs/adding-a-model.md)
(the 3-step recipe) · [the restructure plan](docs/plan/00-index.md).

The v0 era (2026-09-20 → 2026-10-24, 81 systems / 38 families) is
preserved verbatim at tag `v0-final`; history since lives in
[CHANGELOG.md](CHANGELOG.md) and the run records themselves.

## Mixed-pool spectrum benchmark (all 78 systems)

Every system answers the same 48 mixed c