TAU: Proof-of-Generalization Subnet

TAU is a Bittensor subnet that replaces static benchmarks with adversarial, epoch-varying evaluation to measure true model generalization. Current subnets reward miners for memorizing fixed test distributions — TAU changes the game by scoring models on how well they generalize to tasks they've never seen.

Each epoch, a deterministic task generator produces evaluation prompts, then an adaptive distribution-shift engine perturbs them (template rotation, context mutation, controlled noise) to simulate real-world drift. Miners submit probabilistic predictions and are scored via a composite formula balancing accuracy, robustness to shift, cross-task consistency, and calibration — explicitly penalizing overconfidence.

The incentive layer includes cosine-similarity-based collusion detection, a slashing engine for pathological behaviors (score collapse, exploit patterns), and stake-weighted reward distribution with EMA smoothing. The entire pipeline is fully deterministic via SHA-256 seeding — any validator anywhere reproduces identical scores.

The repo ships with a working simulator (5 miner archetypes across 10 epochs), 22 unit tests, and a FastAPI server for submission and score queries.

Core insight: true intelligence is not pattern matching — it is pattern transfer.

Repo: https://github.com/karagozemin/TAU-Subnet
Pitch Deck: https://genesis-subnet-fhue8y2.gamma.site

During the hackathon, we designed and implemented the core deterministic evaluation engine for GENESIS Subnet, replacing static collapse thresholds with a volatility-adjusted (μ + kσ) endogenous risk detection model.

We built the full scoring pipeline, slashing logic, and deterministic seed mechanism to ensure validator non-manipulability.

The architecture, attack-surface analysis, and collapse decision tree were formalized and aligned with implementation.

All documentation reflects the actual codebase (no aspirational components), and the system is structured as a research-grade subnet prototype ready for validator-level review.

Not Raising