Open specification · multi-agent LLM systems

Reasonledger

A notation for transferring reasoning between AI systems that leaves an auditable record of the transfer.

What it is

Multi-agent systems are easy to run and hard to inspect. When a planner hands work to an executor, and that output feeds a third agent, the reasoning behind the final answer is scattered across opaque messages. Plain prose loses structure at each hop; plain JSON carries data but no record of what was checked versus assumed.

Reasonledger encodes model-to-model messages as named blocks, each carrying an explicit instruction for how the receiver should process the enclosed content. Across a chain of agents, those blocks and a three-way epistemic status accumulate into an inspectable record of what each model claimed, what it verified, and what it accepted on trust. That record is the ledger.

It is a payload format. It rides inside a transport such as MCP or A2A; it does not replace one. It is not a blockchain and involves no tokens, chain, or consensus — “ledger” refers to the audit trail the format produces, nothing more.

Epistemic status, preserved across hops

Every claim carries one of three states, kept distinct across model boundaries rather than collapsed into a single answer.

confirmedverified at this step
openasserted, not yet checked
deniedruled out

A small example

@FORMAL [kepler_third]
T² = (4π² / GM) · a³

@DERIVE [circular_velocity]
(1) For a circular orbit, gravity supplies the centripetal force
(2) G·M·m / r² = m·v² / r
(3) From (2): v² = G·M / r
→→ v = sqrt(G·M / r)

@TAXONOMY [orbit_properties]
bound_orbit: confirmed
exact_eccentricity: open
parabolic_escape: denied

After a transfer hop, the receiver's record carries the audit trail: which agent asserted what, which steps were verified, which were trusted, and where a claim's status changed along the way.

From the paper

An LLM auditor's false-positive rate on identical clean reasoning traces moves from 60/60 to 0/60 on prompt wording alone.

Under a decision rule fixed before data collection, the notation gives no failure-detection advantage over information-matched JSON; its value is standardization. The stronger, separate finding is the prompt sensitivity above, reproduced on held-out traces under a design anchored before any live auditor call. These are author-run results; independent replication is invited and has not yet been performed. Reasonledger makes no claim about model cognition.

When the Auditor Cries Wolf: Prompt-Dependent False Positives in LLM Reasoning Auditors — specification, reference checker, evaluation harness, study protocols, and raw results are released.

Resources