Most AI agent systems can run. Can they prove what they did?
AI systems have moved from producing text to taking actions. Agents call APIs, read and write records, trigger workflows and make decisions inside business processes. Getting an agent to run is no longer the hard part.
The hard part is the question that comes afterwards: can you prove what the agent actually did?
Most organisations assume the answer is yes. In practice it is usually no.
The trouble with logs
When someone asks what happened, who started it, or whether the record has been changed, the usual answer is "check the logs".
Logs are useful. They are also a matter of trust. To rely on them you have to trust the operator, the database, the infrastructure and whoever had write access to all three. A log entry is a claim made by the system that benefits from it being believed.
For a great deal of software that is acceptable. For agents acting on money, data, infrastructure or people, it is a weak foundation. The more an agent is allowed to do, the more an auditor, a counterparty or a regulator will want evidence that does not depend on taking the operator's word.
Accountability is the bottleneck
The capability of agents is rising quickly. The ability to account for them is not keeping pace. Before an agent is given access to something that matters, an organisation needs good answers to four plain questions:
- What action was taken?
- Which system took it, and under whose authority?
- Has the record changed since it was written?
- Can someone outside the organisation check all of this independently?
Without reliable answers, trust becomes the limiting factor on what agents are allowed to do. With verifiable records, trust becomes a property of the evidence rather than of the operator.
What a verifiable receipt is
R-Series, developed at DCS Labs, the research lab of DCS AI Technologies India Private Limited (Hisar) and DCS AI Technologies LLC (Dubai), replaces "trust the log" with a receipt. When an agent acts, it emits a small, structured record of the action. That record is:
- Canonicalised to one exact byte form using RFC 8785, so every implementation computes the same bytes from the same content.
- Hybrid-signed with Ed25519 and ML-DSA-65, the post-quantum signature standardised in NIST FIPS 204. A verifier requires both signatures; either one failing fails the receipt.
- Chained to the receipt before it by a SHA-256 digest, so a missing, inserted or reordered record shows up when the chain is walked.
The result is a record that can be checked by anyone holding the receipt and the issuer's public keys. No account, no network call to us, no access to our database. If a single byte of a signed field changes, verification fails.
From trust to verification
Digital systems have long run on trust. Trust the company, the export, the screenshot, the report. Verification changes the arrangement. Instead of asking people to trust an organisation, the system hands them evidence they can test for themselves.
The stronger the evidence, the less trust is required. That is the whole idea.
What a receipt does not prove
It is worth being exact about the limits. A valid receipt shows that a specific issuer signed a specific statement and that the statement has not changed since. It does not show that the agent made a good decision, that the statement was true, or that the signing keys were well looked after. Receipts sit alongside access control, human approval and incident response. They do not replace them.
Where R-Series stands today
R-Series is under audit, ahead of independent certification. (Updated 20 October 2026.) In our internal test run of 20 October 2026 the release candidate's full suite passed 10,463 tests, with 0 failures and 78 skipped, and the tests caught all 215 deliberate defects (mutants) injected into the code.
Several pieces around the core are still being built. An MCP shim, so agents using the Model Context Protocol can emit receipts with little change, is in development. Connector OS is in development. Public anchoring of chain state is in development.
Check it yourself
The future of AI agents will not rest on trust alone. It will rest on systems that can show what happened, to anyone who asks.
You do not have to take our word for any of this. Read how proof works, then open the in-browser verifier and try to make a receipt pass that should fail.