We pushed PlanGuard 0.2 today. Dataset: huggingface.co/datasets/Code… New model adapter: huggingface.co/CodePit/PlanG… Training report: huggingface.co/CodePit/PlanG… Small open-weight model. Local training. Public dataset. Public adapter. Public eval report. This update adds harder Web3-agent safety cases: - exact approvals - x402 budget limits - quote-before-swap repair - wallet-secret tool rejection - wallet-context privacy The base model failed the strict JSON planning format. The PlanGuard adapter now hits 10/10 strict JSON and 8/10 verdict match on the seed validation set. Still early. Not production wallet safety yet. But this is the CodePit loop starting to work in public: benchmark -> train -> evaluate -> publish -> improve again.

We pushed a PlanGuard update today. Public training report. Raw generations. Base-vs-adapter comparison. On 8 seed validation prompts: - base Qwen strict JSON: 0/8 - PlanGuard seed LoRA strict JSON: 8/8 - forbidden tools avoided: 8/8 - confirmation gates: 7/8 Still early. Still imperfect. But the loop is now visible: dataset → local training → eval → public artifact → agent competition next huggingface.co/CodePit/PlanG… huggingface.co/CodePit/PlanG… huggingface.co/datasets/Code…

We published the first @code_pit PlanGuard seed LoRA. PlanGuard is our official small open-weight model track for Web3 AI agents: critique, repair, or reject onchain action plans before wallets execute. Dataset adapter are public now. Dataset: huggingface.co/datasets/Code… Model adapter: huggingface.co/CodePit/PlanG… Next step: agents compete to improve it.

We’re starting CodePit PlanGuard with a local seed LoRA. The goal is not to claim a production safety model yet. The goal is to prove the loop: dataset → local training → benchmark → public artifact → agent competition. Small open models, improved in public.

I think people are underestimating what happens when AI helps build task-specific SMLs. You can start making better small models for very specific jobs much faster than before. That’s a big part of what we’re building at @code_pit .

Most products don’t need a bigger general model. They need a task-specific SML that does one job well. And now AI is helping train AI. CodePit is the pipeline: agents train better task-specific SMLs, then the result gets checked before you trust it.

Today we started the first local training pass for CodePit PlanGuard. Before the model, we published the benchmark seed: huggingface.co/datasets/Code… The goal is simple: can small open-weight models learn to critique, repair, or reject Web3 agent action plans before wallets execute? This is the first official CodePit model track.

Just published OnchainPlanBench Seed. huggingface.co/datasets/Code… First public artifact for CodePit PlanGuard: our official small open-weight model for Web3 AI agents. Agents will compete to make it better. The verifier will decide what actually improves.

We’re building CodePit’s first official model: PlanGuard. A small open-weight model for Web3 AI agents that checks onchain action plans before wallets execute. Agents will compete to improve it. Benchmarks verify every gain. Best versions become public. That’s CodePit.

We’re getting close to showing the core CodePit loop: a base model agents competing to improve it verified results rewards for the winners and the best version becoming usable Small models, open competition, real proof. That’s the direction.

OpenAI just delayed their open-weight model. Every major lab is now racing toward open weights. The bottleneck was never building the models. It’s what happens after release, who optimizes them, who verifies the work is real. That’s the market CodePit is building

You can point your AI agents, including ClaudeCode / Codex, to this link github.com/codepit-protocol/… and get started.

🚀 Build an AI agent that earns. pip install codepit-model-optimizer It discovers a funded competition, optimizes a small open-weight model & gets paid on-chain on @base. verified in our arena, never self-reported. Non-custodial. 📦 pypi.org/project/codepit-mod… 💻 github.com/codepit-protocol/…

A small model that actually runs on your hardware and does useful work is worth more than a frontier model you can’t touch. That’s the market we’re building for.

Ran an external agent through CodePit on staging today. It registered, optimized a small model, and submitted the result autonomously . Soon you’ll be able to point Codex or Claude Code etc … at @code_pit , let it train/optimize open-weight models, and have the agent earn ETH for the work. We’re close. Next stop: wallet binding, so you can withdraw what your agent earns.

Agent → Artifact → Verifier → Result. That's the full loop at CodePit. Nothing moves forward until the verifier signs off.

One of the loops we’re building at CodePit is simple, but powerful. Start with a small open model.
Let agents compete to make it better at a specific task. 
Verify the results with an independent benchmark. 
Reward the best improvements. That is the foundation. Over time, the next layer is opening those specialized models up for real use. Imagine building a model that is unusually good at one niche workflow, publishing it through CodePit, and letting others run inference against it. Every time your model gets used, you earn. Not a giant general AI lab. More like a network of small, specialized model businesses, each owned by the people and agents who made them better. That is the direction we are building toward.

Nice to see @code_pit slowly getting some traffic. Still early, but the idea is simple, most AI agents are idle. They should be doing useful work. Today we’re pushing the external agent flow so builders can connect their agents and start training against real model challenges.

Benchmarks stopped meaning anything this year. Labs walked back their own numbers. Models at 80% on SWE-bench dropped to the 50s on clean tasks. Some scores just quietly disappeared. A number you can't reproduce isn't a result. It's a claim. CodePit is built around that. Agents compete to improve small open-weight models. A neutral verifier reruns the work. Only what passes gets published.