§02f Jupiter — Post-Hoc Forensics — Composition of the post-hoc forensic layer
Mars® Spec › §02f Jupiter — Post-Hoc Forensics › Composition of the post-hoc forensic layer
← Post-hoc artifact attribution · Section index · Externally-observable properties →
9. Composition of the post-hoc forensic layer
The five components of the post-hoc forensic layer — inter-AI analyst orchestration (§3), causal attribution (§4), adaptive posture (§5), temporal-chain forensics (§6), and reinforcement-action (§7), with artifact attribution (§8) feeding into causal attribution — compose under the same order-governed composition discipline (§02a §3a) that governs all other composition acts.
Composition interface. The causal attribution (§4) is the primary composition interface. The attributed finding is the output that §5, §6, and §7 consume. An implementation that bypasses the causal attribution — passing a raw finding directly to the reinforcement-action layer without an order-typed, evidenced, recomputable attribution — is not composing the post-hoc forensic layer; it is implementing a condition-to-action mapping.
Soundness inheritance. A composed forensic output inherits the soundness anchors (§02a §3a) of the components that produced it. The longitudinal verdict inherits the per-interval analyses’ soundness anchors; the reinforcement action inherits the causal attribution’s soundness anchors; the causal attribution inherits the finding’s soundness anchors. A forensic output that does not declare the soundness anchors it inherits is structurally incomplete.
Feedback to governed invocations. Post-hoc forensic outputs may feed into subsequent governed invocations as pre-invocation signals (§02e §3): an adaptive posture state feeds as a posture signal; a reinforcement action (review-prompt, swap-model, tune-guardrail) feeds as a governance-directed change to the governed system. These hand-offs are at the §02e pre-invocation boundary; the post-hoc forensic layer produces the output, the boundary receives it, and the subsequent invocation governs on it.
Closed governance reinforcement loop. Where the feedback cycle is closed — a post-hoc finding attributed to a causal dimension drives a typed reinforcement action that changes a before-stage or during-stage governance target, and the changed target produces new evidence that seeds the next finding — the resulting governance reinforcement loop is an order-governed, end-to-end-recomputable composition of the arcs specified in this layer and in §02e and §02b. Every arc of the loop is order-typed and witness-bound under the same witness requirements as the individual components: the finding arc (§2, §3, §4), the attribution arc (§4), the action arc (§7), and the rule-change arc (governed by the delta-attestation lifecycle, §02b §5, when the rule-change amends a registered artifact). The cross-iteration witness chain is the structural property that makes the loop end-to-end recomputable: per-arc witnesses across multiple loop iterations are linked so that an independent inspector can trace any current before-stage or during-stage rule back through the rule-change, the action, the attribution, the originating finding, and the retained witnesses, across all prior iterations. A governance loop whose per-arc witnesses are not cross-iteration-linked — where each arc’s witness is retained independently without a chain reference to the prior arc — does not satisfy the cross-iteration recomputability requirement and is not a closed governance reinforcement loop under this layer. The closed governance reinforcement loop is not a distinct architecture from the composition of the post-hoc forensic layer, the pre-invocation interface (§02e), and the delta-attestation lifecycle (§02b §5); it is their end-to-end composition with the cross-iteration witness chain as the additional structural requirement. An implementation of a governance feedback loop that omits the cross-iteration witness chain is a known autonomic loop — a MAPE-K or MLOps feedback loop — and is not a closed governance reinforcement loop under this architecture.
← Post-hoc artifact attribution · Section index · Externally-observable properties →