§04 Reasoning — Governed Autonomous Reasoning — Composition requirements
Mars® Spec › §04 Reasoning — Governed Autonomous Reasoning › Composition requirements
← The loop, as a governed composition · Section index · Externally observable properties →
3. Composition requirements (the loop-level obligations)
These bind whenever the capability gate of §1 is met.
§3.1 Gap-driven acquisition (the acquisition target must be a registered gap). An autonomous acquisition action’s target must be a gap record in the Open state (§03-iii §2.0) or a calibration divergence recorded in a variant consistency map (§01 §5.1) at the time the action is initiated. An acquisition action with no corresponding registered gap or divergence at initiation does not conform — autonomous acquisition is driven by the governed gap set, not by an ungoverned internal objective. The gap identity (or divergence identity) is carried in the acquisition action’s record.
§3.2 Traversal is map-governed end to end. Every hop of an autonomous traversal must follow the bidirectional cross-reference structure or the order-decomposition structure (§03-ii S10; §03-iii S2) and record the guide that authorized it (§03-iii S3). A traversal hop taken by raw vector or graph proximity with no governing guide does not conform. Substituting or removing a guide must observably change the traversal path (§03-iii S3/S8) — this is the external test that the walk is governed rather than proximity-driven.
§3.3 Synthesis input selection is declared per act. An autonomous higher-order synthesis act must record its input selection (the identities and version markers of every order/perspective/variant/prior-higher-order output consumed) in a composition-provenance record (§01 §11), exactly as an operator-declared act does. Autonomy does not waive the declaration; it relocates the declaration from the operator to the deployment’s own record. An autonomous synthesis with an unrecorded or unbounded input selection does not conform.
§3.4 Registration before reuse (no silent self-modification). Newly acquired knowledge or a newly synthesized higher-order output must be registered as a versioned artifact — via the delta-attestation lifecycle (§02b §5) for a scope-extending amendment (D42), or via the gap-record closure witness (§03-iii §2.0) for an acquisition — before it is consumed as an input to any subsequent loop iteration. A loop iteration that consumes an unregistered product of a prior iteration does not conform. This is the invariant that keeps autonomous self-improvement auditable: every increment enters the governed record before it propagates.
§3.5 The loop trail is unbroken and recomputable. For every completed loop iteration, the deployment must retain a trail sufficient for an independent inspector to re-derive, without deployment access: (a) the gap or divergence that drove the iteration, (b) the governed traversal path and guides, (c) the evidence acquired or the synthesis inputs consumed, (d) the registration event, and (e) the resulting model/KB version. A loop iteration whose trail cannot be re-derived from retained records alone does not conform — an autonomous loop that improves the system without a recomputable trail is exactly the ungoverned self-modification this section forecloses.
§3.6 Termination and bounds are declared. An autonomous loop must declare its termination or continuation discipline — a bound on iterations, a convergence criterion, a resource budget, or an operator-revocable standing authorization. An unbounded autonomous loop with no declared termination discipline does not conform (cf. §03-iii §2.5 bounded exploration, §03-iii S10/S11). The specific bound is a licensee declaration; the presence of a declared bound is required.
← The loop, as a governed composition · Section index · Externally observable properties →