mixle.reason.anchor_harness module

Measured two-modality scientific-reasoning anchor.

The harness is deliberately lightweight enough for a focused release check, but its evaluation boundary is real:

• independent training, calibration, and test splits;

• gravity and geochemistry both enter the inferred grade posterior;

• all coefficients, residual scales, and the gravity-only baseline are learned from training data rather than copied from the hidden generator;

• calibration and abstention thresholds are selected without test outcomes;

• abstained predictions are actually withheld and selective risk is reported;

• receiver-specific projections operate on each inferred test posterior.

Bootstrap fits create an empirical posterior over the two learned stages gravity -> density and (density, geochemistry) -> grade. Component variance propagates both stages’ residual uncertainty. This is not a claim that the synthetic task is a frontier application; it is a fast, falsifiable integration contract for the reasoning surfaces.

class AnchorHarnessReport(modalities, hop_names, coverage_by_hop, abstained_site_ids, abstain_rate, answer_rate, accepted_mae, all_answer_mae, driller_projection_components, scout_projection_components, driller_readout, scout_readout, projection_decision_agreement, gravity_only_mae, both_modalities_mae, gravity_ablation_penalty, test_coverage, coverage_target, coverage_lower_bound, walk_is_calibrated, training_rows, calibration_rows, test_rows, notes=<factory>)

Bases: object

Measured train/calibration/test report for the anchor task.

Parameters:

• modalities (list[str])

• hop_names (list[str])

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

• abstained_site_ids (list[int])

• abstain_rate (float)

• answer_rate (float)

• accepted_mae (float)

• all_answer_mae (float)

• driller_projection_components (int)

• scout_projection_components (int)

• driller_readout (str)

• scout_readout (str)

• projection_decision_agreement (float)

• gravity_only_mae (float)

• both_modalities_mae (float)

• gravity_ablation_penalty (float)

• test_coverage (float)

• coverage_target (float)

• coverage_lower_bound (float)

• walk_is_calibrated (bool)

• training_rows (int)

• calibration_rows (int)

• test_rows (int)

• notes (list[str])

property walk_mae: float

Compatibility name for the two-stage, two-modality posterior mean MAE.

summary()

Render the measured evaluation boundary and results.

Return type:

str

run_anchor_harness(*, n_train=1200, n_calibration=None, n_test=200, seed=0)

Fit on training data, calibrate on a disjoint split, and evaluate once.

Parameters:

• n_train (int)

• n_calibration (int | None)

• n_test (int)

• seed (int)

Return type:

AnchorHarnessReport