Mechanism
Partial state handling is the set of structured mechanisms the inference-time semantic execution substrate provides for situations in which the admissibility gate cannot render a definitive determination, the cumulative rejection rate exceeds a threshold, or the inference process encounters a semantic boundary it is not authorized to cross. Rather than forcing the inference engine to either abort or push ahead with a degraded result, the substrate exposes three first-class responses to an incomplete or unresolvable situation: decomposition, deferral, and safe non-execution. Each is a defined outcome of the governed inference loop, not an exception handler bolted onto it.
These mechanisms exist because the substrate treats inference as semantic execution rather than token generation. A proposed mutation that cannot be admitted as a whole is not the end of the process. The substrate asks whether the mutation can be broken into parts that can be admitted, whether its evaluation depends on information that may arrive later in the same inference, or whether the correct response is to stop and report rather than to generate inadmissible content. Partial state handling is the name for that three-way choice.
Decomposition
Decomposition breaks a proposed mutation that is too coarse-grained for atomic admissibility evaluation into finer-grained sub-mutations. It is triggered when the admissibility gate determines that a proposed mutation contains both admissible and inadmissible components: a single proposed transition that bundles several semantic changes, some of which pass the gate and some of which do not. The decomposition module separates the components, submits the admissible components individually, and either rejects or recursively decomposes the inadmissible components.
Decomposition is bounded. A maximum decomposition depth is specified in the policy reference field of the semantic state object, so recursive separation cannot continue without limit. This lets the inference advance through fine-grained steps when a coarse-grained transition is inadmissible as a whole but admissible in parts, while keeping the bounding of the work under policy control rather than leaving it to the inference engine.
Deferral
Deferral suspends evaluation of a proposed mutation whose admissibility depends on information not present in the semantic state object and not obtainable through anchor resolution. Rather than reject the mutation outright or admit it on incomplete grounds, the deferral mechanism records the deferred mutation in a pending evaluation queue, annotated with the specific information deficiency, and continues inference along an alternative path.
The deferred mutation is not discarded. If subsequent admitted transitions supply the missing information, the deferred mutation may be re-evaluated against the now more complete semantic state. If inference concludes without resolving the deferral, the deferred mutation is reported as unresolved in the lineage. The absence is therefore carried forward as an explicit, recorded condition rather than being silently filled or silently dropped.
Safe Non-Execution
Safe non-execution terminates the inference process without producing a complete output when conditions for continued inference cannot be met. It does not produce nothing. Safe non-execution produces a partial output comprising the admitted semantic content, a structured termination report identifying the triggering condition, and a complete lineage record. The consumer of that output therefore receives both what was established before termination and an explicit account of why the process stopped.
The treatment of non-execution as a valid, first-class outcome is an architectural property of the substrate. The system treats silence as the correct response when the alternative is generating inadmissible content. This inverts the default behavior of a probabilistic inference engine, which always produces a continuation because producing a continuation is what it does. Here, stopping is a governed result with its own recorded justification.
Relation to the Admissibility Gate
Partial state handling is reached from the semantic admissibility gate, which evaluates each proposed mutation and returns admit, reject, or decompose. Decomposition is itself one of the gate's three outcomes: a mutation that bundles admissible and inadmissible components is broken into sub-mutations that are each independently re-submitted to the gate. Deferral and safe non-execution are reached when the broader inference loop cannot proceed: the gate cannot render a definitive determination, the cumulative rejection rate crosses a threshold, or an unauthorized semantic boundary is encountered.
The same handoff is reached from entropy-bounded admissibility. When a candidate transition exceeds the current entropy bounds it is rendered non-executable and the engine is instructed to select an alternative candidate; if no alternative candidate satisfies the entropy bounds, the inference process transitions into partial state handling. Anchor resolution feeds the same structure: an unresolvable anchor causes its mutation to be rejected, and an ambiguous anchor causes its mutation to be decomposed into alternatives, one per candidate referent, each submitted independently.
Lineage Recording
Every outcome of partial state handling is recorded. The semantic lineage maintains an ordered record of every admitted transition, every rejected transition's rejection rationale, and every decomposition event, written into the lineage field of the semantic state object. A decomposed transition records the sub-mutations into which it was broken. A deferred mutation that is never resolved is reported as unresolved. A safe non-execution emits a complete lineage record alongside its admitted content and its termination report.
Because each admissibility determination is deterministic, the lineage supports reproducibility as well as audit: given the same initial semantic state object, inference engine, and input, the same sequence of determinations, including the same decompositions and deferrals, would be produced. The pattern of rejections and decompositions also functions as a structured diagnostic of where the inference process most often fails, without requiring retraining of the underlying model.
Distinction from Conventional Handling
A conventional inference pipeline confronted with an input it cannot fully evaluate has two common responses: abort the call as an error, or proceed by substituting a default, a prior, or a learned imputation. The abort discards partial work that could have been usefully completed in part. The silent substitution discards the fact that the substitution occurred, so a downstream consumer reasons about the result as if the missing information had been present. Partial state handling provides structured alternatives to both. Decomposition completes the part of the work that is admissible. Deferral carries the absence forward as a recorded, possibly resolvable condition. Safe non-execution stops while still returning the admitted content and a structured account of why it stopped.
The governing principle is that an absence or an unresolved dependency is preserved as explicit, recorded state rather than absorbed into the output. This is what distinguishes the substrate from a system that merely declines to answer and from a system that answers regardless: here the partial result and the reason for its partiality are both first-class, recorded artifacts.
Disclosure Scope
Partial state handling, comprising the decomposition of coarse-grained mutations into finer-grained sub-mutations under a policy-bounded maximum decomposition depth, the deferral of mutations whose admissibility depends on information not yet present in the semantic state object and not obtainable through anchor resolution by recording them in a pending evaluation queue annotated with the specific information deficiency, and safe non-execution that terminates inference while producing a partial output of admitted semantic content together with a structured termination report and a complete lineage record, is disclosed in the cognition filing (U.S. Application No. 19/647,395 and its international counterpart) at Section 8.12. This article describes that disclosed mechanism. The scope extends to its reachability from the admissibility gate's admit, reject, or decompose outcomes, from entropy-bounded non-executability, and from unresolvable or ambiguous anchor resolution, and to the recording of every such outcome in the semantic lineage, including deferred mutations reported as unresolved when inference concludes without resolving them.
