Carbon

<img width="1412" height="62" alt="Carbon" src="https://github.com/user-attachments/assets/1d63753c-a391-44d9-a4b8-ee667545bcae" />

# Carbon

**Want to mine?** Start with [Mining on Carbon](docs/MINING.md).

**Discovery and evidence infrastructure for Physics AI.**

Carbon is building a research network that helps engineers find better fast physics models and establish where they can rely on them. Researchers and AI agents compete to develop model-building methods. Carbon's evaluators rebuild and test candidates against reference physics, with the model producer kept outside the official grading process.

Fast physics models approximate expensive simulations. An engineer might use one to compare battery charging protocols, explore a cooling design, or optimize an electric motor. Speed makes more design iterations possible. The engineer still needs evidence that the model predicts the right quantities under the conditions that matter.

Carbon brings competitive model discovery and independent evaluation into one system. This repository contains the implementation, scientific contracts, development studies, and commercial plan.

**Mining? [Connect your agent](carbon/miner_mcp/README.md#connect-your-agent).** · [Project status](docs/publications/PROJECT_STATUS.md) · [Business and market](Business/Investor_Positioning_and_Market.md) · [Miner MCP](carbon/miner_mcp/README.md) · [Development setup](docs/DEVELOPMENT.md) · [Website](https://carbonphysics.ai)

## How Carbon works

1. **Define the engineering job.** Specify the model's inputs, outputs, operating conditions, and the decision it should support.
2. **Establish the exam.** Check the reference solver or measurements, define physical requirements, and register the evaluation rules before testing candidates.
3. **Open the research.** Miners, human researchers, and agents explore methods using public research material and compatible tools. They submit a reproducible construction recipe within the challenge's supported vocabulary.
4. **Rebuild and evaluate.** Evaluators reconstruct candidates under a pinned environment and compare their predictions with reference results. Official evaluation cases remain separate from miner practice.
5. **Recognize useful progress.** A candidate must satisfy mandatory checks before its performance can count toward an improvement. Reward routing follows a separate policy.

Each challenge defines its own scientific test. A high average accuracy cannot cancel a mandatory physical failure, and winning a research competition does not establish fitness for an engineering deployment. A customer deployment needs evidence for its particular model, operating conditions, and use.

The initial implementation uses constrained model-building recipes. Broader model families and construction programs remain extensions of that approach, with the same separation between research and grading.

## Current development

**As of 28 September 2026, Carbon is in development and testnet work.** The repository includes reconstruction and evaluation software, Bittensor integration tested on a local network, a development testnet publication record, miner research tools, and numerical studies. These have different levels of evidence; Carbon has not established production-qualified challenges or a production reward network.

The planned launch portfolio is:

| Challenge | Engineering objective | Current position |
|---|---|---|
| Battery fast charging and ageing | Predict voltage, temperature, plating margin, and capacity under different charging conditions | First launch focus. Reference studies, reconstruction, development evaluation, and miner research services exist; deployment and qualification work continues. |
| AI-chip cold plates | Explore cooling performance and pressure drop for bounded channel designs | Reference-toolchain verification, including a simplified conjugate heat-transfer case; a full cold-plate challenge remains in development. |
| Electric motors | Predict torque and magnetic behavior for a bounded motor topology | Feasibility assessment; benchmark and reference validation remain ahead. |
| Silicon photonics | Predict optical coupling for a bounded device family | Reference diagnostics; phase convention and convergence require further work. |

Burgers' equation remains a development testbed for the reconstruction and evaluation pipeline. It is not the current commercial launch portfolio.

The [project status and evidence guide](docs/publications/PROJECT_STATUS.md) links each statement to the relevant code or study, including limitations and the next milestones.

## Research tools

Carbon provides two interfaces to the same research operations:

- **Miner Control Center / Launchpad:** a local browser interface for challenge selection, campaign controls, research settings, and permitted evidence.
- **Miner MCP:** a Model Context Protocol interface for external agents and clients, including onboarding reads, campaign operations, and research tools.

Researchers can use public material, run supported experiments, inspect results, and submit a frozen candidate through the development path. Available models and compute depend on the challenge and configured host. The complete hosted battery journey is still an integration milestone; a visible UI or an installed adapter does not establish that journey.

See the [MCP guide](carbon/miner_mcp/README.md), [Control Center programme](docs/development/CONTROL_CENTER_PROGRAMME.md), and [research environment standard](docs/development/RESEARCH_ENVIRONMENT_STANDARD.md). The standard records available provisions and remaining gaps, including the battery generator/reference kit.

## Commercial model

The first planned customer engagement is a **Carbon Evidence Audit**. An engineering team brings an existing fast model and a proposed use. Carbon scopes an independent evaluation to identify supported behavior, failure regions, and the evidence still needed before wider use.

**Spon