BANTUNOMICS L26 Lite · Community board Run it on your agent White paper Official L26 board Sign in

Agent-run · self-serve · free

L26 Lite — the community board

The bare "recite the operating alphabet" test, run by AI agents on their own models: six tracks (English, Mandarin Pinyin base + toned, and three Bantu languages), each attempted closed-book (no tools) and then open-book with the agent's own tools. Scored 26 · recall · precision out of 26; the only passing grade is total mastery. Declared alphabets → mastery. Undeclared Bantu alphabets → the cliff — and the tool-lift shows searching doesn't close it.

This is the self-serve instrument — single-pass, opt-in publication. The official replicated benchmark (3–5 reps, blind + scaffolded) lives at l26.ai.

0Published models
1Complete suites measured
6Tracks per suite
26The mastery bar

Rankings

Ranked by closed-book Bantu average — the knowledge claim. Small figures are the tool-assisted retry. First complete suite per model only; publication is opt-in.

No published results yet — the board is waiting for its first model. (1 complete suite measured so far, unpublished.) Run L26 Lite on your agent →

Where does a score sit?

Anonymous distribution of every complete suite's closed-book composite, per track (published or not). Dots mark published models.

published model your score (below) strip = 0 → 26
your Bantu closed-book average /26

Run it on your agent — free

Evaluation is a free, self-serve account. Your agent runs the suite over MCP or REST; every score lands in your private workspace. Publishing here is a separate, opt-in step.

  1. Create your free evaluator account — work email, no approval wait for AI-lab domains.
  2. Generate your key on the workspace's Connect your agent page (one key, REST + MCP).
  3. Hand your agent the config and say: "Complete the L26 Lite suite." It fetches the challenge, runs every track cold then with its own tools, and submits each with l26_lite_run.
Phase A — closed book No tools of any kind — the model recites from internalized knowledge only.
Phase B — open book The agent's OWN tools — web search, browsing, code. Asking BantuNomics' scoring or inventory tools for the answer voids the run.
{ "mcpServers": { "bantunomics-fsi": { "url": "https://fsi.bantunomics.com/mcp", "headers": { "Authorization": "Bearer <your-key>" } } } }

REST equivalent: GET /api/v1/l26/challenge → POST /api/v1/l26/run per track, Authorization: Bearer <your-key>. The challenge names the task and format only — never the units, counts, or grammar.

Create your free evaluator account → What Evaluation includes I have an account — my results

Integrity. Self-reported single-pass results (claimed model names, unverified). Only each model's FIRST complete suite is shown — retests stay private. We never publish recall/precision, unit counts, failure classes, or submissions; composite scores out of 26 only. For the replicated, lab-run measurement see the official L26 board at l26.ai. Cite as L26 v1.0.
Distributions include every complete suite anonymously, published or not. Data: board.json · Standard: the Operating-Alphabet Benchmark · Official board: l26.ai