§02f Jupiter — Post-Hoc Forensics — Reinforcement-action layer
Mars® Spec › §02f Jupiter — Post-Hoc Forensics › Reinforcement-action layer
← Temporal-chain forensics · Section index · Post-hoc artifact attribution →
7. Reinforcement-action layer
The post-hoc forensic layer selects a typed reinforcement action bound to a recomputable causal attribution, routes it to a named governance target, and emits a re-derivable finding-to-attribution-to-action linkage.
Typed action selection bound to attribution. The reinforcement action is selected as a function of, and bound to, the causal attribution produced under §4. The governed action taxonomy includes {sync-knowledge-base, coach-user, review-prompt, swap-model, tune-guardrail, investigate}, which is non-limiting and extensible. The action is a function of the attribution, not of a raw condition; a condition-to-action mapping that does not consume a causal attribution is not a reinforcement action under this section.
Named-target routing. The action is routed to a named governance target: a specific knowledge base or data source, a specific user or group, a specific prompt template or workflow, a specific model or use case, a specific rule or guardrail engine, or a specific incident. Named-target routing is a structural requirement, not an optional annotation; an action without a named target has not been routed.
Re-derivable finding-to-attribution-to-action linkage. The full chain — finding, attribution, action — is emitted such that an independent inspector re-derives why this action followed this finding, from the finding, its attribution, and the retained witnesses. A mismatch reveals an unsound action. The linkage is a first-class forensic artifact, not a log entry.
Urgency and escalation. Each action carries a recomputable, attribution-bound urgency: immediate, next-cycle, or backlog. The urgency is itself bound to the attribution and recomputable; an urgency that is not derivable from the attribution and the retained witnesses is structurally incomplete.
Investigate as the non-automatable default. Where the causal attribution does not soundly determine an automatable action — because the attribution is uncertain, multi-dimensional, or requires human judgment — the reinforcement-action layer selects investigate (human review) rather than asserting an automated action. The investigate default is structurally required; asserting an automated action on an attribution that does not soundly support it is a false-grounding condition.
Distinction from gap-locus artifact reinforcement. Artifact reinforcement — synchronizing a knowledge base, amending a model — may be selected as a reinforcement action under this section, but only when selected as a function of a causal attribution under §4, routed to a named target, and carrying a re-derivable finding→attribution→action linkage. Artifact reinforcement selected on the basis of a gap locus alone — without a causal attribution binding — is not a reinforcement action under this section; it is a gap-remediation act governed by §03. The two are not equivalent and are not interchangeable: the gap-remediation act operates on a gap in an artifact relative to a specification; the reinforcement action under this section operates on a forensic finding attributed to a causal dimension. An implementation that conflates the two — routing a gap-locus signal directly to a sync-KB action without a causal attribution — is not implementing the reinforcement-action layer and does not satisfy the re-derivable linkage requirement.
← Temporal-chain forensics · Section index · Post-hoc artifact attribution →