Struct CompiledLogicProgram

Source

pub struct CompiledLogicProgram { /* private fields */ }

Implementations§

Source §

impl CompiledLogicProgram

Source

pub fn prepare_conditioned( &self, py: Python<'_>, prob_source: &str, memory_mb: Option<u64>, ) -> PyResult<CompiledConditionedProgram>

Compile this program’s accepted epistemic evidence and prob_source once.

The returned handle can be evaluated repeatedly while independent probabilistic fact priors are changed atomically. It does not accept caller-supplied input relations, matching evaluate_conditioned.

Source

pub fn evaluate_conditioned( &self, py: Python<'_>, prob_source: &str, memory_mb: Option<u64>, ) -> PyResult<EpistemicEvalResult>

Run this epistemic program on the GPU and condition prob_source on what it knows.

The compiled program must contain epistemic operators (know, possible, …) AND lower to a single-component epistemic plan: ordinary Datalog programs are rejected, because there is no accepted world view to condition on. Only facts declared in the epistemic program’s own source are used to build that world view.

Both epistemic modes are reachable here. FAEEL programs and non-recursive G91-compatibility programs (#pragma epistemic_mode = g91) both lower to a single-component epistemic plan and condition normally; epistemic_evidence().epistemic_mode names the mode, and the trace’s accepted_faeel_world_view_evidence_consumed / accepted_g91_world_view_evidence_consumed pair says which one actually supplied the evidence. Only the recursive G91 shapes (positive possible cycles that need tuple-level compatibility) compile to a dedicated G91-compatibility plan and are rejected at plan level, alongside split, stratified and WFS plans.

LIMITATION: unlike evaluate, this method does not accept dlpack_inputs. Caller-supplied input relations are NOT consulted — if the epistemic program depends on a relation that is normally supplied at call time via evaluate(dlpack_inputs=...), that relation is empty here and no world view is accepted. This method then RAISES RuntimeError (“Unsupported epistemic construct: accepted GPU world-view evidence … probabilistic evidence requires non-empty accepted GPU final output”); it does NOT fall back to the unconditioned prior. That is fail-closed by design: a conditioned query that silently became unconditioned would be indistinguishable from a successful one, which is exactly the failure the trace counters exist to make visible. To test for the case without catching an exception, call epistemic_evidence() first — it reports accepted_world_views == 0 without raising.

The returned trace must show a non-zero gpu_conditioned_evidence_facts — otherwise the conditioning did not reach the GPU exact path. gpu_conditioned_evidence_facts is the total the engine itself validates; the per-class counters (gpu_conditioned_know_evidence_facts, gpu_conditioned_possible_evidence_facts, gpu_conditioned_not_known_evidence_facts, gpu_conditioned_not_possible_evidence_facts) break it down. A possible-only or negated-evidence program conditions correctly with the know counter at 0, so check the total, not the know class alone.

Source

pub fn epistemic_evidence(&self, py: Python<'_>) -> PyResult<EpistemicEvidence>

Run this epistemic program on the GPU and report what it accepted.

Diagnostic counterpart of evaluate_conditioned: it answers whether the know-broadcast happened at all, without involving the probabilistic tier.

LIMITATION: like evaluate_conditioned, this only ever sees facts declared in the epistemic program’s own source; it does not accept caller-supplied input relations. A program that depends on such a relation reports accepted_world_views == 0, accepted_candidates == 0 and final_output_rows == 0 here — without raising. The operator censuses (know_operator_count, possible_operator_count) are read off the plan, not the execution, so they stay non-zero: it is the accepted/consumed family that goes to zero, not “every counter”. Calling evaluate_conditioned() on that same program RAISES rather than returning an unconditioned result, so this method is the non-raising way to probe for the case first.

Source §