§02f Jupiter — Post-Hoc Forensics — Externally-observable properties
Mars® Spec › §02f Jupiter — Post-Hoc Forensics › Externally-observable properties
← Composition of the post-hoc forensic layer · Section index
10. Externally-observable properties
| Property | Observable condition |
|---|---|
| P1 — Per-analyst attribution | Each AI analyst’s contribution to a forensic finding is attributable to that analyst by order-typed finding element; no finding element is attributable only to “the system” |
| P2 — Inter-analyst divergence as a forensic signal | Where two or more AI analysts diverge on the same question, a divergence record is present in the forensic output as a first-class artifact; the divergence is not resolved into an aggregate before recording |
| P3 — Recomputable analyst contributions | An independent inspector re-derives each analyst’s contribution and the recorded divergence from the retained witnesses without executing the deployment |
| P4 — Order-typed causal attribution | Each forensic finding carries a causal attribution tagged by the order(s) under which the finding and its cause are expressed; attribution is not a flat label |
| P5 — Evidenced attribution | The causal attribution cites the witness evidence — methods, divergence signals, derivation steps — that supports it |
| P6 — Recomputable causal attribution | An independent inspector re-derives the causal-dimension assignment from the finding and its retained witnesses; a mismatch reveals an unsound attribution |
| P7 — Order-typed posture state | The adaptive posture state expresses scrutiny tagged by order and policy class, not as a single scalar |
| P8 — Recomputable posture transition | An independent inspector re-derives each posture transition from the prior state, the triggering finding, and its causal attribution against retained witnesses |
| P9 — Attribution-bound posture change | Asked why scrutiny changed, the system returns the bound causal attribution as an evidenced, re-derivable answer |
| P10 — Order-structured longitudinal composition | The longitudinal verdict is the composition of per-interval order-typed analyses across time under the composition discipline, inheriting per-interval soundness anchors |
| P11 — Recomputable longitudinal verdict | An independent inspector re-derives the longitudinal verdict from the retained per-interval witnesses |
| P12 — Attribution-bound reinforcement action | The reinforcement action is observably a function of the causal attribution, not of a raw condition |
| P13 — Re-derivable finding-to-attribution-to-action linkage | An independent inspector re-derives the full finding→attribution→action chain from the retained witnesses |
| P14 — Named-target routing | The reinforcement action names a specific governance target |
| P15 — Investigate default | A finding whose attribution does not soundly determine an automatable action yields investigate, not an asserted automated action |
| P16 — Governing-bound fallback | Given assets but no specification, the forensic layer uses a domain-model-as-spec or reverse-synthesized model-as-spec, declaring the fallback position used |
| P17 — Two-layer asset relation | Asset-relative attribution is resolved at the join of KB governance and data-source instance-facts; neither layer alone is treated as sufficient |
| P18 — Recomputable asset-relative attribution | The asset-relative attribution is re-derivable by an independent inspector against retained witnesses |
| P19 — Sycophancy dual-locus invariance | A post-hoc sycophancy scoring finding carries a content hash linking it to the same registered discrete sycophancy representation as the runtime gate disposition (§02b); a finding lacking that hash is not a governed sycophancy finding |
← Composition of the post-hoc forensic layer · Section index