Archipelo has released Salmon, an Execution Verification Infrastructure (EVI) that records every action taken by AI agents, humans, and automation as cryptographically signed events. By turning execution steps into a machine‑consumable evidence trail, engineers can verify what actually ran, rather than relying on the resulting state alone.
How Salmon Captures Verifiable Execution
When an actor—whether a person, an AI model, or an automated script—issues a command, Salmon creates a signed event that includes:
- The identity of the actor
- The specific action performed
- The system state before the action
- The system state after the action
- A cryptographic signature linking the event to the actor
These events are chained together into a Verifiable Execution Record, preserving the lineage from the original request to the final state. If an execution step cannot be captured, Salmon records a gap instead of fabricating continuity, ensuring the evidence set remains trustworthy.
Architectural Implications
Integrating Salmon requires a few concrete changes to existing pipelines:
- Event Generation Layer – Any component that invokes tools, APIs, or modifies infrastructure must emit a signed event to Salmon. This may involve adding lightweight SDK calls or middleware hooks.
- Secure Signature Management – Actors need access to private keys or signing credentials. Key rotation and protection become part of the security baseline.
- Immutable Store – The execution records must be persisted in a tamper‑evident store (e.g., append‑only log, blockchain‑style ledger, or write‑once storage). The store should be accessible to downstream security, safety, and governance services.
- Verification Service – Consumers—such as intrusion detection, compliance audit, or autonomous supervision—must be able to retrieve and cryptographically verify the record before acting on it.
Because Salmon does not replace existing IAM or runtime controls, it sits alongside authorization checks, providing a post‑fact audit trail rather than an access gate.
Operational and Security Considerations
From an operations perspective, teams will need to address:
- Performance Overhead – Generating and signing events adds latency. Engineers should benchmark the impact on high‑frequency agent loops and consider batching where appropriate.
- Storage Growth – Continuous event capture can generate large volumes of data. Retention policies and compression strategies must be defined.
- Gap Management – When evidence is missing, Salmon explicitly records a gap. Operators should monitor gap frequency as an indicator of integration gaps or potential tampering attempts.
- Key Hygiene – Compromise of signing keys would undermine the trust model. Regular rotation, hardware‑based key storage, and audit of key usage are essential.
Security teams gain a new data source for detecting unauthorized credential use, tool invocation, or state changes that were not part of an approved workflow. The cryptographic signatures make it difficult for a compromised agent to forge a clean execution trail.
Related CloudNinjas coverage: DevOps.
What This Means For Practitioners
Practitioners should start by mapping existing automation points to the event model Salmon requires. Identify where agents invoke external services, modify infrastructure, or delegate work, and insert signing calls at those boundaries. Evaluate your key management process to ensure it can support per‑actor signing without introducing bottlenecks. Finally, set up a verification pipeline that consumes the Verifiable Execution Record and feeds it into your security monitoring, compliance audit, and autonomous supervision tools. By doing so, you gain a tamper‑evident execution log that can be used for rapid incident response, forensic analysis, and continuous governance of AI‑driven operations.
