AMSA — Agent Multi-System Aggregator

Multi-agent orchestration
you can re-derive, not just trust.

AMSA coordinates and audits multi-agent AI systems on top of a deterministic, cryptographically verifiable substrate. Every identity, state change, memory entry, and ownership transfer beneath AMSA is hash-linked and independently re-derivable — so agent behavior stays auditable after the fact, not just observable in the moment.

amsa.verify_chain() — live in your browser, real SHA-256
✓ CHAIN VALID
PAYLOAD (editable)
record_hash
Edit any payload above — downstream hashes break instantly. That's the whole guarantee.

Five claims. Each one is a test you can run, not a sentence you have to believe.

AMSA orchestrates agents on top of the DIS substrate — these invariants aren't marketing properties of AMSA alone, they're enforced in the underlying DIS/SDS code and checked by the test suite on every build.

SDS · Core

Content-addressed, not clock-addressed

No uuid, no os.urandom, no wall-clock timestamps inside any hash input. Identity comes from a LogicalClock tick and canonical JSON — same inputs always produce the same chain.

SDS · Chain

Tamper-evident by construction

Every record embeds hash_chain(prev_hash, payload). Change one byte in history and every record after it fails verify() — not eventually, immediately.

DMV · Ownership

Merkle-provable transfers

Asset and permission transfers fold into a binary Merkle tree. You can prove any single transfer is valid without replaying the entire ledger.

AMSA · Execution

Results bound to the state that produced them

Every execution record carries the exact state-hash and memory-hash active at that tick — so a result is provably traceable to what the agent actually knew, not what it claims to have known.

NeuroZip · Archive

Bit-identical snapshots

Canonical JSON plus content addressing means re-compressing the same state twice produces the same archive hash, every time, on any machine.

Honest scope

We don't claim to make LLMs deterministic

Model sampling is still sampling. DIS doesn't fight that — it cryptographically records what a non-deterministic model did, so the record is verifiable even when the output wasn't predictable.

AMSA, built on eight composable modules.

AMSA is the orchestration layer. Each module beneath it ships standalone with its own test suite, then composes into the full substrate AMSA runs on.

amsa ★ orchestrator
Multi-agent orchestration — routing, broadcast, mesh topology, replay and diff between snapshots.
tested
core
Hashing, canonical JSON, logical clock, identity primitives shared by every other module.
stdlib only
aiin
Stateful identity nodes with deterministic lineage — parent/child agent hierarchies, signed and verifiable.
tested
sds
Sovereign Data Store — a hash-linked, versioned key-value journal with full write history per key.
tested
dmv
Binary Merkle tree for ownership and permission transfers, with compact proof generation.
tested
memora
Episodic memory chain per agent, tag-indexed, with cross-agent search.
tested
neurozip
Graph-preserving codec for snapshotting and restoring full substrate state.
tested
persist
Chain-linked checkpoints over JSON or SQLite, with diff and cross-backend export.
tested

Run it yourself. That's the point.

Every number below comes from the test suite and demo shipped in the source, not a deck.

871 / 871
tests passing across the full monorepo
0
runtime dependencies — stdlib Python only
25
assertion-guarded demo sections, end to end
8
composable modules, each independently tested
core/hashing.pyexcerpt
def hash_chain(prev_hash: str, payload: Any) -> str:
    # No timestamps, no randomness —
    # identical inputs always yield this hash.
    blob = canonical_json({
        "prev": prev_hash,
        "payload": payload,
    })
    return sha256(blob).hexdigest()
$ python -m unittest discoverreal output
Ran 871 tests in 4.218s

OK

$ python demo.py
[25/25] Mesh diff (two checkpoints) ... PASS
All 25 sections completed. 0 assertions failed.

Read the code before you read about it.

We're sharing the reviewer package — full source, specs, and the limitations we haven't solved yet — with engineers evaluating this for integration, licensing, or acquisition.