DSM · trust & audit layer

    Prove what your AI agent did.

    Append-only, hash-chained, tamper-evident execution trails. The proof layer you hand to an auditor.

    $ pip install daryl-dsm

    The problem

    Your agent decided. Something went wrong.

    Now prove what happened. The usual tools can't.

    Logs are mutable

    Anyone with access can edit or delete them. No integrity guarantee.

    Vector DBs reconstruct

    They approximate context probabilistically — they don't preserve decisions.

    Frameworks track calls

    They follow tool calls, not verifiable proof of what actually executed.

    The proof

    Tamper detection, live

    One command recomputes the entire hash chain. Alter one byte, verification fails.

    $ dsm verify --shard sessions
    total_entries: 248  verified: 247
    tampered: 1  chain_breaks: 1
    status: FAILED — entry #182 altered after the fact

    Tamper-evidence: DSM detects modification, reordering and truncation. It does not prove the truth or correctness of the content — by design, and that's what makes it credible.

    DSM — Daryl Sharding Memory

    The product.

    Off-chain trust, memory and audit layer for AI agents. The provider proposes; DSM validates, audits and replays — never the authority itself.

    01

    Shard

    Each entry is appended to an append-only shard. Nothing is modified or deleted.

    02

    Hash-chain

    SHA-256(content + prev_hash). Every entry is linked to the one before it.

    03

    Verify

    Replay the chain, recompute every hash, confirm integrity — independently.

    How it works

    From action to proof.

    Five deterministic steps, from the agent's action to an independently verifiable proof of integrity.

    01

    Agent acts the action intent is appended to an append-only shard.

    02

    Entry is hashed SHA-256 chained to all prior entries.

    03

    Entry is signed Ed25519 signature proves authorship (optional).

    04

    Shard is sealed archivable with a cryptographic tombstone.

    05

    Anyone verifies anyone holding a copy of the trail replays the chain and confirms its integrity.

    Why it matters

    Use cases.

    Wherever you must reconstruct what an agent did — verifiably, not just plausibly.

    EU AI Act

    High-risk systems must keep decision traceability. DSM provides a tamper-evident audit trail — a building block, not a certification.

    Accountability

    Loan approval, patient triage, trade execution: reconstruct what happened with proof of integrity, not a log grep.

    Multi-agent governance

    Which agent did what, when, and was it authorized? Verifiable identity, dispatch and cross-agent receipts.

    Developers · open source

    Start in one line.

    The DSM engine is open source (MIT). Record and verify agent trails today.

    CLI & library

    Record actions, then verify the whole chain with one command.

    $ dsm verify --shard agent_memory

    What's included

    Append-only engine, hash chain, signing, attestation, receipts, sealing, multi-agent governance.

    MIT1500+ testsPython 3.10+Goose MCPEd25519SHA-256

    Roadmap

    Toward third-party-verifiable proof.

    DSM today is a strong tamper-evidence layer. What's next hardens proof against a fully privileged adversary.

    Shipped

    Tamper-evidence

    Hash chain, signing, attestation, receipts, sealing.

    In design

    External anchoring

    Signed checkpoints and independent witnesses.

    Planned

    Witness / MMR / STH

    Append-only and non-equivocation proofs.

    Future

    API & dashboard

    Hosted verification and compliance reporting.

    Daryl does not become a blockchain. It stays off-chain; on-chain anchoring is deferred and would only prove existence, timestamp and integrity of hashes — not truth.

    Vision · origin

    From fiction to function.

    DARYL began as a character. Today it aims to become the standard for verifiable agent execution — the equivalent of digital signatures, but for AI decisions. Calm, precise, honest about its limits.