Orrery Introduces 2-Part Architecture to Reconstruct Past AI Runs

1 articles · Updated · O'Reilly Media · Oct 8

Orrery proposes preserving AI systems' execution-time bindings so investigators can reconstruct what a system actually used months later, even when rerunning it would yield a different result.
The design targets a preservation gap in modern AI stacks: deployments, traces and model registries may survive, but not the runtime-selected model endpoint, policy, retrieved data, tool contract or configuration tied to one execution.
Its core requirement is twofold—binding integrity and resolution integrity—so systems record which dependency states an execution used and keep those historical identities resolvable for a defined retention period.
The article argues preservation failures need explicit semantics, especially when external effects are committed but usage records are not durably stored, turning reconstructability into only best-effort evidence.
AIGov Core is cited as a partial implementation, using a hash-chained evidence ledger and replay checks labeled Ready, Partial or NotReady to test reconstructability, but only for bindings captured in the first place.