README / README.md
takakuni-imosuke's picture
org card renders README.md (remove index.html static app)
5e472ff verified
|
Raw
History Blame Contribute Delete
2.49 kB
---
title: TensorCash
emoji: 🔍
colorFrom: indigo
colorTo: purple
sdk: static
pinned: false
---
<div align="center">
<img src="tensorcash-og-dark.png" alt="TensorCash" width="680">
# TensorCash
**A layer-1 blockchain whose proof-of-work *is* verifiable AI inference.**
*Don't trust the provider. Verify the machine.*
[Website](https://tensorcash.org) · [Verifier (GitHub)](https://github.com/tensorcash/tensorcash/tree/main/services/verification-api) · [Whitepapers](https://tensorcash.org/whitepapers) · [Blog](https://tensorcash.org/blog)
</div>
---
## What is TensorCash?
Most AI runs behind a black box: you send a prompt, tokens come back, and you have no
way to check *which* model produced them — or whether the work was really done.
TensorCash makes model execution **checkable**. Every generation on the network emits a
compact, replayable **proof of inference** — the chosen tokens, the sampler draws, and
the top-of-distribution logits, bound together so anyone can re-run the claimed model and
confirm the output is genuine. The same proof doubles as the chain's proof-of-work: the
mining puzzle is real inference, not wasted hashing.
## On Hugging Face
- 🔍 **[verify-proof-of-inference](https://huggingface.co/spaces/tensorcash/verify-proof-of-inference)** — browse released proofs and see, side by side, the full GPU-replay verdict, a local CPU transcript self-check, and the cheap distributional screen. It shows exactly *why* a cheap screen isn't enough.
- 📊 **[proof-of-inference-eval](https://huggingface.co/datasets/tensorcash/proof-of-inference-eval)** — a byte-exact corpus of honest and forged proofs (cross-architecture substitution, same-architecture fine-tune, and 4-bit quantization) with full methodology, so you can reproduce the detection results yourself.
In live GPU teacher-forced replay, full verification accepted honest proofs with zero
false-rejects and detected **100% of all three forgery classes** — while a cheap
distributional screen misses same-architecture fine-tunes and quantization. Per-token
replay is the decisive check.
## Build & verify
- Run your own verifier — [walkthrough](https://tensorcash.org/blog/how-to-run-a-verifier)
- Read the method — [*Verifiable Inference*](https://tensorcash.org/whitepapers/verifiable-inference)
---
<div align="center">
*Open · pseudonymous · verifiable*
Educational and technical resources only — nothing here is financial advice or an offer.
</div>