When something goes wrong, the first instinct is to reconstruct: pull the logs, line up the timestamps, and work out what happened. For decisions made by AI agents, that instinct leads somewhere uncomfortable. By the time anyone asks, much of what you'd need to know is gone, or has quietly changed.
Decision provenance
Researchers have a name for the thing you actually want. In "Decision Provenance: Harnessing Data Flow for Accountable Systems" (IEEE Access, 2019), Jatinder Singh, Jennifer Cobbe and Chris Norval propose using provenance techniques to expose decision pipelines: the chain of inputs that led to a decision, the nature of the decision, and its effects downstream. Their argument is that accountability in systems-of-systems depends on understanding these data flows, not only on how a particular algorithm works.
Agents make that argument sharper. An agent's "pipeline" is assembled at run time: whatever it read, whatever tools it called, whatever another agent handed it. No design document captures it in advance. The only reliable time to capture it is the moment it happens.
Why reconstructing later goes wrong
Databases learned this lesson long ago. A bitemporal table keeps two times for every fact: when it was true in the world (valid time), and when the system recorded it (transaction time). Without both, you can't answer the question an investigation really asks: what did the system believe then, before the later corrections came in?
Reconstructing an agent's decision after the fact runs into the same problem, and a few others:
- State moves on. The file the agent read has been edited since. The permission it held has been revoked. The policy has a new version. Today's state answers a different question.
- Existence isn't access. A log shows a document existed. It doesn't show the agent read it, or that it could have and didn't. Treating "it was in the system" as "the agent knew" is how reviews invent knowledge.
- Hindsight leaks in. Once you know how things turned out, it's very hard not to read that outcome back into the evidence. A record written before the outcome can't be contaminated by it.
- Logs belong to the system under review. An auditor has no reason to trust logs written and stored by the system being investigated.
Write it down before acting
Decision Horizon applies decision provenance at the one point where an agent's action can be both seen and stopped: the tool call. Before each consequential action it writes a record of what the agent had observed, with content hashes. It also records which of that content reached the action, what was accessible but not read, the authority in force, and the rule that decided. Where it can't establish something, it says so rather than filling the gap.
Two choices follow from the lessons above:
- Time-bound by construction. The authority and horizon are captured as they were at the instant of the decision, and the record is signed before the action runs. Later changes can't rewrite it without breaking the chain.
- Unknown is an answer. A record that guesses is a record nobody can rely on. Deterministic claims carry full confidence. Estimates are labelled as estimates. Everything else goes into
abstain.
Reconstruction will always have a place: not every decision passes through a point you control. But for the actions that do, the cheapest and most trustworthy evidence is the record you write at the time.
See how a record is structured in the docs.