Skip to content

Evidence vs observability

Teams that run AI agents usually already trace them with an LLM observability tool: Langfuse, LangSmith, Arize Phoenix, Datadog LLM Observability and others. Those tools and HardState both record what an agent does. They are built for different readers, and a regulator, a court or the person affected is not the reader an observability tool is built for.

Observability tool HardState
Reader The team that builds and runs the agent A third party: auditor, supervisor, court, the person affected
Question it answers Why did the agent do that, and how do we make it better? What did the system decide, when, and has anything been changed or removed since?
Granularity Every span: tokens, latency, cost, prompts, evaluations Each decision, each consequential action, each human review
Where the data lives A database its operator administers, and can edit, filter or purge Your institution’s HardState Log: signed, chained and anchored to Bitcoin
Contents Full prompts and outputs, for debugging Only salted commitments; contents stay with you
Who must be trusted The operator of the tool, and whoever administers it Nobody: anyone checks it in the browser against Bitcoin
Shows a deleted record No Yes: the chain breaks, and the heartbeat shows the gap

An editable store is what makes an observability tool good at its job. Retention policies, filters and deletion are features there. They are also why a log kept in one cannot, on its own, prove to an outsider that it is complete and unaltered. Its operator could have changed it, and the outsider has no way to check.

Keep it. HardState does not replace it and does not need it. The two answer different people:

  • your engineers keep the full traces they debug with, for as long as they find useful;
  • the decisions those traces lead to, and who reviewed them, go to the HardState Log, where they can be verified later by someone who trusts neither you nor us.

HardState records decisions on its own, with no tracing tool needed. If your agent already marks its decision points, that is where ai.decision goes. A whole agent run can be sealed as one ai.run record.

HardState is a Zevlat Intelligence venture. © 2026 Zevlat Intelligence.