mirror of
https://github.com/hanzoai/chains.git
synced 2026-10-30 18:46:11 +00:00
zeekayandGitHub
96da029e9a
* hanzo-evm: aiquorum + aimarket precompiles, evmllm (geth AI precompile + e2e), operator/canonical (commit/reveal, int8-quant), aiffi bench. Thinking-chain substrate. * hanzo-evm/evm: wire AI precompiles (inference 0x0300..03 + modelregistry) Registers deterministic int8 inference (Tier-1 in-consensus AI) + ModelSpec registry into the Hanzo EVM. threshold dep aligned to v1.9.4 (go.mod match). Builds green (55M binary). aivmbridge follows luxfi/precompile release.
49 lines
2.5 KiB
Go
49 lines
2.5 KiB
Go
// Copyright (C) 2026, Hanzo AI Inc. All rights reserved.
|
|
// See the file LICENSE for licensing terms.
|
|
|
|
// Package canonical is the SHARED TRUTH for the LLM quorum-settlement wire
|
|
// format. It is the piece that makes a quorum actually FORM: two honest
|
|
// operators running the same ModelSpec over the same engine must produce
|
|
// BYTE-IDENTICAL hashes, or the on-chain precompile never sees agreement and
|
|
// the job fails.
|
|
//
|
|
// # Why this package exists separately from the operator
|
|
//
|
|
// The operator (sibling package, cgo) talks to the native engine and signs
|
|
// payloads. THIS package is pure Go — no cgo, no I/O, no clock, no RNG except
|
|
// where a nonce is explicitly drawn by the caller. Everything here is a total
|
|
// function from bytes to bytes. That is what lets it be table-tested to the byte
|
|
// and cross-checked against the on-chain precompile (precompile/aiquorum) for
|
|
// exact equality. Determinism lives here; effects live in the operator.
|
|
//
|
|
// # The contract with the chain (precompile/aiquorum/aiquorum.go GOVERNS)
|
|
//
|
|
// Every hash in this package is byte-identical to the precompile's recompute on
|
|
// reveal. The precompile is the verifier: at RevealResponse it recomputes the
|
|
// commit from the revealed (output_hash, embedding_hash, nonce) and rejects any
|
|
// mismatch. If a single byte of ordering, width, or endianness diverges here,
|
|
// no honest operator can ever reveal. The functions that MUST match:
|
|
//
|
|
// ModelSpecHash <=> aiquorum.ComputeModelSpecHash
|
|
// Commit <=> aiquorum.ComputeCommit
|
|
//
|
|
// crosscheck_test.go (in the operator module) imports the real precompile and
|
|
// asserts equality on randomized inputs, so on/off-chain agreement is proven
|
|
// against the deployed code, not a transcription of it.
|
|
//
|
|
// # What the chain agrees on vs. what it never sees
|
|
//
|
|
// The chain treats output_hash and embedding_hash as opaque 32-byte values. It
|
|
// only decides whether >= threshold operators submitted the SAME bytes. The
|
|
// MEANING of those bytes is defined entirely here:
|
|
//
|
|
// RAW mode output_hash = keccak256( deterministic UTF-8 model output )
|
|
// GOVERNANCE mode output_hash = keccak256( canonical structured decision )
|
|
//
|
|
// In GOVERNANCE mode the consensus is deliberately narrowed to a structured
|
|
// decision {vote, confidence_bucket} bound to the model_spec_hash. Free-form
|
|
// rationale prose is NOT in the consensus hash — see governance.go for the full
|
|
// justification. Rationale and citations ride along as non-consensus audit
|
|
// metadata, canonicalized by canonicaljson.go (RFC 8785) for storage only.
|
|
package canonical
|