Skip to content

Diagnostics — Overview

xPRIMEray Research

xPRIMEray produces several categories of diagnostic output from each fixture run. This page describes what each diagnostic measures, what it is useful for, and what it is not.

All diagnostics are derived from Pass 1 stored-hit data or from approved debug captures. No diagnostic modifies or overrides the Pass 1 transport classification.

Transport diagnostics contact sheet showing classification overlays

Diagnostic Categories

Diagnostic What it measures Primary use
Curvature heat maps Cumulative turn angle per pixel Identify transport-structure zones, high-bending annuli
Tile coherence Edge recall and direction-fidelity per tile geometry Choose sampling texture for a given diagnostic goal
Phase coherence Per-pixel coherence score relative to neighbourhood Locate phase-boundary transitions, correlate with banding
Domain ownership Transport-regime cluster membership per checkpoint Identify interpolation-safe zones, domain boundary positions

Reading Diagnostic Outputs

Coverage table (coverage.json)

The primary hermetic validation output. Fields:

  • classified_coverage_ratio — must be 1.0 for a passing run
  • escaped_no_hit — rays that left the scene without classifying; must be 0
  • budget_exhausted — rays stopped by step budget, not by a hit; must be 0
  • Per-class pixel counts: geom_hit, portal_hit, throat_event, throat_entry, throat_exit, throat_shell_transform, throat_inner_absorb

Derived metrics (derived_metrics.json)

Computed from the stored-hit table:

Field Description
opl_mean, opl_max Optical path length statistics
portal_hit_density Portal hits per pixel
throat_event_density Throat events per pixel
crossings_per_pixel Portal/throat crossings per pixel
segments_per_crossing Integration segments consumed per crossing — high values indicate expensive traversal
average_segments_per_ray Mean integration segments per ray

Anomaly scores (bridge_anomaly_scores.json, anomaly_detection/scores.json)

Three independent anomaly measures: Euclidean z-score in standardised feature space, isolation forest, and local outlier factor (LOF). All three ranked post_throat_backstep_01 (the bridge) as the top anomaly in the six-checkpoint ladder (z = 4.40, isolation forest = 0.616, LOF = 1.35).

What Diagnostics Are Not

  • Heat maps are not observable flux. A curvature heat map shows cumulative turn angle, not pixel brightness, emissivity, or redshift-weighted intensity. See heatmaps.md.
  • Coherence previews are not validated truth. Phase-coherence and tile-coherence diagnostics are interpretive layers. They explain structure; they do not override classifications.
  • Clustering is not ground truth. The k = 3 regime decomposition (ARI = 0.5946) recovers the manual regime labels imperfectly. It is evidence for domain structure, not a definitive boundary map.

Cross-References