Building Ché — A Human Dataset, Owned by the People Who Make It

Collecting Humanities Experience.

We've been building Ché — a platform where people contribute the things that make us human, and get paid when those contributions are used to train AI. This is a field note on what it is, why it matters, and what we shipped this week.

Visit Ché — che.cortexresearch.group

The premise

The most valuable training data left in the world isn't scraped text — it's the stuff that never made it online. The creative decision no tutorial taught you. What a melody feels like before you can name the notes. The bug that changed how you think. Lived experience.

Ché collects exactly that, from the ground up, and treats the people who provide it as owners, not raw material. Two things make it different from a survey or a data-labeling gig:

1. Attribution is permanent and on-chain. Every contribution is anchored to its author.
2. Contributors get paid when their data is actually used in a model — not just for showing up.

The token is $NEURON. Its whole job is to route value back to contributors.

How it works

The loop is simple on the surface and deliberate underneath:

1. Join on SKALE. Connect a wallet, join the waitlist. SKALE is a gasless EVM network — sFUEL (its gas token) is valueless — so it costs the contributor nothing to transact. That's what makes "put attribution on-chain" practical at scale.
2. Ask, or get interviewed. People ask their own questions and the AI (Ché) asks questions. It's one unified feed. Crucially, the questions humans ask train the questions the AI asks — Ché learns to interview in the community's own voice.
3. Everyone answers. Humans answer; the AI can also take a pass. Human answers are the payable dataset.
4. Anchor. A hash of each answer plus the contributor's address is written on-chain — verifiable, tamper-proof attribution without bloating the chain with raw text.
5. Earn $NEURON on a Bitcoin-style halving curve. Early contributors earn the most.

The hard part: paying people when their data is used

This is the question everyone waves at and nobody answers honestly. A trained neural net blends millions of examples — there is no exact, scalable way to say "this output token came from Alice's answer." So we don't pretend to.

Instead, value gets created at two moments, and we pay at both:

• Mining reward — you contribute, you mine $NEURON immediately. Pays for showing up early.
• Usage reward — when your contribution is actually included in a trained model version, you get paid again, on a fact we can prove: was your content hash in the dataset snapshot that produced model vN? Yes/no, on-chain.

We weight that payout by tractable quality signals (peer upvotes, uniqueness, eval-set impact) — an Inclusion + Quality Multiplier, not fake per-token precision. Each training run mints a reward pool from the Community allocation and publishes a single Merkle root on-chain; contributors claim their own share with a proof. On SKALE, the claim is gasless.

This is our answer to the Open Rights problem — user A creates, user B creates a derivative, how do you attribute all parties? You anchor every contribution, snapshot what each model actually used, and let people pull their share trustlessly. The full design lives in the project's compensation spec.

When the model eventually earns revenue, the same claim mechanism becomes an ongoing royalty stream — new money source, same rails.

$NEURON tokenomics

Total supply 1,000,000,000. Rewards follow a deflationary halving curve, capped by the Community pool.

Bucket	%
Community	50
Infrastructure	20
R&D	15
Reserve	10
Team	5

The Community half (500M) is mined out to contributors — first in, most rewarded.

What we shipped this week

• A neo-brutalist landing page with wallet connect on the SKALE Europa Hub, a waitlist that records signups, and graceful fallback when no contract is deployed yet ("anchoring pending").
• The unified Ask & Answer feed — humans and AI both ask, both answer, tagged by source and category.
• The training feedback loop: human-asked questions are passed as exemplars so Ché's questions learn the community's voice (works even offline against the real community questions).
• An AI layer behind OpenRouter (model-swappable), proxied through a serverless function so the API key never touches the browser.
• The Waitlist.sol contract (join + anchor-by-hash) and the full compensation spec.
• Deployed to Netlify, live and verified.

What's next

• Deploy Waitlist.sol to SKALE testnet so waitlist writes become real on-chain transactions.
• Build RewardDistributor.sol + the claim screen (Merkle-root payouts per training run).
• The live AI interview screen with on-chain answer anchoring.
• The corpus export job that turns on-chain anchors into an attributable training set — and credits contributors in the model's output.