mixle.reason.anchor_harness module

Runnable geoscience harness for cross-modal target inference.

The harness infers a synthetic subsurface ore-grade target from two heterogeneous evidence modalities: a categorical geochemistry observation and a potential-fields gravity reading. It composes existing reasoning primitives into one measured workflow:

• mixle.reason.modality represents geochemistry and gravity as typed ModalityView objects.

• mixle.reason.cycle_consistency fits conditional transports and checks whether a hop should abstain.

• mixle.reason.belief_walk composes the gravity-to-density and density-to-grade hops by Monte Carlo belief propagation, with calibration checked by hop count through coverage_by_hop_count().

• mixle.reason.task_projection projects the same ore-grade belief for two receivers: one that needs an exact tier and one that needs a binary drilling decision.

The data are synthetic and seeded, using a linear-Gaussian generative process with coefficients hidden from the fitted transports. The report is checkable against ground truth and records measured coverage, abstention, and error.

class AnchorHarnessReport(modalities, hop_names, coverage_by_hop, abstained_site_ids, abstain_rate, driller_projection_components, scout_projection_components, driller_readout, scout_readout, frontier_mae, walk_mae, frontier_is_calibrated, walk_is_calibrated, notes=<factory>)

Bases: object

Measured report for the cross-modal geoscience harness.

Parameters:

• modalities (list[str])

• hop_names (list[str])

• coverage_by_hop (dict[int, dict[str, float]])

• abstained_site_ids (list[int])

• abstain_rate (float)

• driller_projection_components (int)

• scout_projection_components (int)

• driller_readout (str)

• scout_readout (str)

• frontier_mae (float)

• walk_mae (float)

• frontier_is_calibrated (bool)

• walk_is_calibrated (bool)

• notes (list[str])

summary()

Render a compact anchor-harness evaluation summary.

Return type:

str

run_anchor_harness(*, n_train=2000, n_test=200, seed=0)

Run the cross-modal harness and return the measured report.

Requires torch (the transport fits do); raises the underlying ImportError if it is absent, same as other neural transport fitting paths.

Parameters:

• n_train (int)

• n_test (int)

• seed (int)

Return type:

AnchorHarnessReport