# HOUETOR

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Pierre Florent BOLOUVI](https://cerebralvalley.ai/u/Justpio)
- **GitHub:** https://github.com/bolouvipf/tessercore
- **Demo video:** https://youtu.be/Xn9O6wTGuG8
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4147

TESSERCORE — AI-generated code you can trust, that becomes free through reuse. Push the Frontier. Indie solo.

Problem: LLMs can write small tools, but teams block them — unverified code is unsafe, self-graded tests miss bugs, and regenerating the same function wastes calls. We make software stay tiny and learn verified capabilities on demand.

What it is: Minimal kernel (4.12 KB gz, engine 677 KB gz once) that loads pure-function modules (JSON Schema contract + run(input)). Permissions=[] in v1: no network/DOM/storage/clock. Request "convert 50000 FCFA to euros" resolves T0 cache (offline, 0 LLM) -> T1 signed registry (hash+Ed25519) -> T2 Router Muse Spark Open -> T3 Generation (contract + 2 impls). Verified modules are signed, cached and shared — next request is 0 LLM, <3 KB.

Trust: Contract first. Two implementations (Muse Spark Open temp 0.2/0.7, different prompts) must agree on 200+ auto-generated inputs within 1e-9, pass properties (e.g. round-trip), examples and human confirm of assumptions. One repair round, then fail. Sandbox: QuickJS-WASM in Worker, deadline 200ms, memory 8192KB, output 4096B, textContent only. Attacks blocked: fetch/DOM-> no access, loop->197ms interrupt, mem->MEMORY_LIMIT, 1-byte tamper->HASH_MISMATCH, subtle_wrong->REJECTED, injection->cannot grant permissions.

Meta AI: All runtime calls are Meta models via server allowlist. Primary: muse-spark-open-1026 on Baseten (open-weight, 10K bonus) for router/contract/impl A; impl B = same model sampled differently (honest). No non-Meta runtime.

Results (bench/results/*.json, Node 24 win32): M1 4.12KB gz, M6 evil_loop 197ms/MEM 2.3ms, M4 45 req -> 80% 0 LLM (simulated session), second profile 0 LLM/3KB, 23 tests pass.

For: teams building agents that need runtime tools they can trust. Offline after first load, private registry per org.

Limits: signature=origin not safety; deadline not fuel; 2 models can share error; assumptions shown not verified; generation needs connectivity; request text sent to gateway.

Before: github.com/bolouvipf/houetorwasm — research study on WASM plugin hosting, only as separate named reference, no code copied (AGENTS.md#1). Built Oct29-Nov4: kernel, sandbox, contract, cache/offline, registry, router/generation, verification, 4 seeds (fcfa/unit/loan/cement demo), hostile fixtures, bench. Repo: github.com/bolouvipf/tessercore. Video <3:00.

## More from Meta Global AI Developer Hackathon

- [Apni Gati](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4144)
- [Bypass Labs](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4145)
- [cjrackley](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4146)
- [Silmari AI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4148)
- [Pneumonion](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4149)
- [Stupid](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4150)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/4147. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt