The Data Internet.
Corpora become owned assets, and retrieval enters consensus.
Retrieval, recomputed
A grounded job binds a model, a corpus and a query. The committee converges on two hashes: the output, and the retrieval manifest of exact chunks used.
- 01 / Ground
Embed the query
Tokenize, then embed with integer kernels pinned in the AIOS Module. Same vector on every node.
- 02 / Retrieve
Exact search, top-k
Flat kNN or BM25 over the committed index, ties broken by content hash. Never approximate ANN.
- 03 / Generate
Grounded on the chunks
The model runs on exactly the retrieved context, deterministically.
- 04 / Clear
Two hashes agree
Output hash and retrieval manifest: work-weighted >2/3 on both is canonical.
Corpora as assets
One Merkle root per corpus
A corpus registers like a model: an address, an owner, a per-call price and a verification mode, plus a Merkle root over its chunked, integer-embedded contents and a pointer to where the bytes live. Every grounded proof names that root, so the exact corpus state is provable at any height.
Usage proven, royalty paid
A grounded call's fee splits between the Miners, the model owner, the corpus owner and the protocol burn. The data-royalty is additive, funded by the grounded price, never carved from the Miner slice and never a mint. Earnings on real usage, never framed as yield.
Verified Memory
One visibility flag. Public corpora are listed, priced and verified. Private memory is yours: encrypted, scoped and self-verifiable.
Versioned, under your keys
Your memory is chunked, embedded and indexed as a personal vector store, append-only and committed, so you can prove its exact state at any height. It is organized into session scopes you define, each a retrieval boundary and a permission boundary. The network holds commitments and encrypted pointers, never plaintext.
Verified.
- Listed, priced and browsable
- The committee recomputes the retrieval
- Earns a data-royalty on every grounded call
- Anyone can check the proof
Self-verifiable.
- Encrypted under your keys, unlisted
- You recompute any recall yourself
- Optionally attested by a TEE, never committee-verified
- Recall your own memory for the inference fee only
What verified data means
Provenance, not truth
A verified retrieval proves the bytes match the commitment and were retrieved correctly. It never certifies that the data is true, clean or unbiased.
Integrity, not relevance
Consensus proves every Miner got the same top-k, not the best top-k. It is a claim of agreement, never of correct retrieval.
Deterministic retrieval is unproven
Float embeddings and approximate indexes are nondeterministic. Any corpus that cannot be made byte-deterministic verifies by sampling, never exact-hash.
Not a storage layer
Small corpora sit in chain state; large ones are content-addressed off-chain. AIOS binds data to execution and proves the retrieval. It runs no live oracles or data feeds.
One proof: model, corpus, chunks
Register a corpus and earn on every grounded call, or ground your own calls from code.