Add SUBIT-INK to community projects

[sciganec]

This PR introduces SUBIT-INK to the Community Projects list under the Ink Detection tools.

Contribution:
We have implemented a 2D U-Net with a dual classification head based on Belnap's four-valued logic (SUBIT-TOPOS foundation). The model outputs two independent probabilities (yang and yin), mapping to four semantic states: INK (confident), BOTH (ambiguous/edge), VOID (background), and UNKNOWN (lacuna/damage).

Value to the ScrollPrize community:

1. Higher text yield: When the ambiguous BOTH pixels are included in the final mask, the F1 score reaches 0.8243, outperforming the standard binary baseline of 0.8197.
2. Hallucination prevention: Pixels with unstable 3D signals are routed to BOTH instead of being misclassified as false ink, ensuring cleaner training data for downstream OCR.
3. Human-in-the-loop efficiency: The BOTH map explicitly highlights letter boundaries and uncertain areas, allowing experts to focus verification efforts exactly where needed.
4. Damage handling: The UNKNOWN state correctly marks missing data, preventing models from inventing text in lacunae.

The code is fully open-source (MIT) and implemented in PyTorch, making it immediately usable by other teams working on ink detection for the Herculaneum papyri.

[vercel]

@sciganec is attempting to deploy a commit to the scroll Team on Vercel.

A member of the Team first needs to authorize it.