Releases·Tool Release
911 dispatcher builds Arcaeon 0.9.0, a tamper-evident log tool that verifies AI agent records without trusting the agent
A developer who runs a 911 dispatch floor released Arcaeon 0.9.0, an MIT-licensed tool that makes AI agent logs verifiable by third parties. Agents write their own records of tool calls; Arcaeon adds an external, checkable chain so anyone can detect edits without the agent's cooperation. It targets the gap where an agent's self-reported 'story' is the only audit trail.
5 engagement1 sourceFri, Sep 25, 2026, 07:06 AM
AI Overview
- A 911 dispatch-floor operator released Arcaeon 0.9.0 (MIT-licensed, zero dependencies), a tool for tamper-evident AI agent logs
- Arcaeon's hash-chain 'verify' flags any log edit: changing one word yields 'BROKEN first_break: line 1: chain mismatch', and checks need no account or key
- Arcaeon includes a proxy that records agent tool calls without agent cooperation, a reconcile tool comparing agent vs. tool records, and an MCP-server source checker for dangerous patterns like leaked secrets
- Arcaeon proves only that a record wasn't altered after writing, not that it's true — an agent can faithfully record a lie
- The only paid piece is an outside witness holding the log's latest state off-site, closing the gap where someone truncates the record; code is on github.com/Arcaeon-io/arcaeon
Entities
Related
Developments·Paper
Hearsay: Can an Auditor Trust the Record a Deployed Agent Harness Writes?
An arXiv paper, Hearsay, asks whether auditors can trust the records agent harnesses write of their own runs. Across sixteen deployed frameworks, none writes a record a third party can verify without trusting the writer. The authors show an externally kept append-only log of harness-model exchanges catches all fabricated or omitted entries, where hash chains over the harness's own record do not.
An arXiv paper, Hearsay, asks whether auditors can trust the records agent harnesses write of their own runs. Across sixteen deployed frameworks, none writes a record a third party can verify without trusting the writer. The authors show an externally kept append-only log of harness-model exchanges catches all fabricated or omitted entries, where hash chains over the harness's own record do not.
0 engagement1 sourcearxivSat, Sep 26, 2026, 11:37 AM
Products·building
AgentAudit — tamper-evident audit trail for AI agent actions (EU AI Act oriented)
AgentAudit is a Python SDK that records what AI agents do — LLM calls, decisions, tool calls, inputs/outputs — into a hash-chained, tamper-evident audit trail. It targets teams building agents who need to debug failures and satisfy EU AI Act traceability requirements. The builder is seeking feedback and testers.
AgentAudit is a Python SDK that records what AI agents do — LLM calls, decisions, tool calls, inputs/outputs — into a hash-chained, tamper-evident audit trail. It targets teams building agents who need to debug failures and satisfy EU AI Act traceability requirements. The builder is seeking feedback and testers.
1 engagement1 sourceredditThu, Sep 10, 2026, 09:44 AM
Social·Community
The log I trust after a week unattended is the one the agent never wrote
A developer describes why agent-written traces and compaction summaries make untrustworthy records of what an autonomous run actually did, and how their reef setup on a Raspberry Pi instead keeps runtime-written records of every exchange and report. A harness-evolution tutorial run also emits a replay.html showing each step's verdict and tree diff. The system records but does not enforce; enforcement must live where writes or sends happen.
A developer describes why agent-written traces and compaction summaries make untrustworthy records of what an autonomous run actually did, and how their reef setup on a Raspberry Pi instead keeps runtime-written records of every exchange and report. A harness-evolution tutorial run also emits a replay.html showing each step's verdict and tree diff. The system records but does not enforce; enforcement must live where writes or sends happen.
12 engagement1 sourceredditWed, Sep 30, 2026, 04:00 AM
Products·launched
Verify AI agent actions from terminal without trust
A lightweight accountability layer for AI agents that lets anyone verify what an agent did from their terminal in 15 lines of code, with no signup required. It uses blockchain-based finality semantics (settled, blockNumber) to provide tamper-evident records, solving the trust problem for agents acting with real authority.
A lightweight accountability layer for AI agents that lets anyone verify what an agent did from their terminal in 15 lines of code, with no signup required. It uses blockchain-based finality semantics (settled, blockNumber) to provide tamper-evident records, solving the trust problem for agents acting with real authority.
9 engagement1 sourceredditTue, Jul 21, 2026, 04:47 PM
Products·building
Verify AI agent actions with tamper-evident proofs
Cinchor is a lightweight tool that lets operations teams share verifiable records of autonomous agents' context, instructions, and output. It uses a custom L1 blockchain to anchor proofs in a tamper-evident manner, solving the problem of trusting first-party logs from AI agents.
Cinchor is a lightweight tool that lets operations teams share verifiable records of autonomous agents' context, instructions, and output. It uses a custom L1 blockchain to anchor proofs in a tamper-evident manner, solving the problem of trusting first-party logs from AI agents.
2 engagement1 sourcehackernewsThu, Jul 23, 2026, 03:01 PM