[lispandfound]

Summary

Replaces the tectonic type rules in source_modelling/focal_mechanism.py with the CatSep plausibility scores of Haynie et al. (2026, SRL 97(2A), doi:10.1785/0220250078), following software release v1.0.0 (DOI-USGS/neic-catalog-separation@1.0.0). Everything sits behind the existing public API.

Within the Slab2 footprint, all ramps are from eq. 1 and Table 1 of the paper:

• P_interface = p_flex · p_sz · p_kagan
  • p_flex is 1 within ±15 km of the slab surface (the "flex zone").
  • p_sz ramps from 1 at the seismogenic zone depth to 0 at 15 km below it. The depths are from Hayes et al. (2018): Kermadec–Hikurangi 47 km, Puysegur 30 km.
  • p_kagan ramps from 1 at a Kagan angle of 30° to 0.25 at 53°. The angle is measured against slip on the slab plane at the region's seismogenic rake (93° for Kermadec, CatSep's 90° default for Puysegur). The term is dropped when there is no mechanism.
• P_crustal = (1 − P_interface) · ramp(depth − slab depth: 1 at −20 km → 0 at +20 km) · p_moho
• P_slab takes the remainder.
• Outside the footprint, events are crustal, as before (CatSep's non-NSHM mode).

Moho: CatSep's CRUST1.0 depth-to-Moho grid (Laske et al. 2013), cropped to NZ (NZ_CFM/crust1_moho_nz.npz, 2.7 kB, 1° cells 54.5–24.5°S, 156.5–189.5°E), is used through MohoModel. Outside that grid there is no Moho (∞): the crust extends to the slab. Not implemented: the overturned-slab supplement, the mantle flag and the A–D quality grades.

API

Stable: CMTClassifier.tectonic_type(centroid, nodal_plane_1, nodal_plane_2), the call workflow's gcmt-to-realisation (and so the realtime simulation workflow) makes, and TectonicType. The planes are now optional. The Kagan angle is the minimum over both planes, so plane order doesn't matter.

New: kagan_angle (via scipy's D2 rotation group), MohoModel, SeismogenicZone, SEISMOGENIC_ZONES, SlabModel(grids, seismogenic_zones), SlabQuery.seismogenic_zone, and an optional moho_model keyword on CMTClassifier (from_faults/load pass the packaged grid). The ramps are np.interp calls on KAGAN_RAMP / CRUSTAL_DEPTH_DIFFERENCE_RAMP.

Removed:

• The NSHM 2022 rule: the old signature of the module-level tectonic_type_probabilities, _is_interface_like and its constants, and the sigma keyword of CMTClassifier.tectonic_type_probabilities.
• The NGA-SUB rule: nga_sub_tectonic_type(_probabilities), SlabZone, SlabZoneQuery and SlabModel.zone.

validation_small_magnitude/scripts/compare_sources.py passes sigma=0.0 and will need updating.

Behaviour

The workflow test events keep their old classifications:

Event	Result
2003 Fiordland	interface
2016 Kaikōura	crustal
2017p795065	slab

On the 81 crustal validation events in simulation_validation (GeoNet CMT centroids and planes):

Rule	Crustal	Interface	Slab
CatSep (CRUST1.0 Moho)	74	5	2
CatSep (no Moho)	75	5	1
NSHM 2022	74	3	4

Six misses are within 10 km vertically of a shallow Hikurangi Slab2 surface, and five have Kagan angles of 20–48° to interface slip. The Moho adds one more: 2020p198433, a 24 km centroid in an 11 km CRUST1.0 cell, drops from 0.88 crustal to 0.43 (slab 0.47). The Moho can only lower crustal scores.

Test plan

• The rule's branches, including the mantle wedge and a Moho below the slab; seismogenic zone and Moho lookups; plane-order invariance; and the no-mechanism / outside-footprint paths
• The np.interp/scipy simplification matches the previous implementation to 1e-15 over 200,000 random inputs (Kagan angles to 1e-12°)
• Kagan angle identities and the ≤ 120° bound
• Full suite: 862 passed; ruff, ty and npdlint clean

🤖 Generated with Claude Code

[lispandfound]

@claude review

[claude]

Claude finished @lispandfound's task in 4m 46s —— View job

---

Review complete

• Gather context (PR body, diff, repo guidelines)
• Read changed files in full
• Review for correctness bugs, security issues, edge cases, regressions
• Check for existing review comments to avoid duplicates (none found)
• Post findings

No CLAUDE.md/REVIEW.md present in this repo to follow beyond the default correctness-focused review.

Findings: 0 blocking, 0 minor

I hand-traced tectonic_type_probabilities (source_modelling/focal_mechanism.py:512-589) against all 9 new parametrized cases in tests/test_focal_mechanism.py (including the mantle-wedge ramp and the "Moho below slab" discontinuity case), and every branch (flex, seismogenic, kagan, above_moho, crustal) reproduces the expected scores exactly by hand calculation — no arithmetic or branch-ordering bugs found.

Specifically checked and confirmed correct:

• flex_top <= depth <= slab_depth + FLEX_ZONE_KM vs. the elif moho_depth <= depth < flex_top mantle-wedge branch don't overlap or leave gaps.
• The above_moho cliff to 0.0 when moho_depth >= slab_depth (focal_mechanism.py:574-579) is intentional and exercised by test_tectonic_type_probabilities[48.0-40.0-None-45.0-...] — np.interp's clamp-to-boundary semantics make the degenerate xp=[m, m] case well-defined here, not a bug.
• _slip_vector (new) matches the standard Aki & Richards eq. 4.88 slip-vector formula used alongside the pre-existing _plane_normal.
• KAGAN_RAMP/CRUSTAL_DEPTH_DIFFERENCE_RAMP are unpacked into np.interp(x, xp, fp) in the correct order.
• No remaining references to the removed nga_sub_*, SlabZone, _is_interface_like, or sigma APIs anywhere else in this repo (the one known external caller, validation_small_magnitude/scripts/compare_sources.py, lives outside this checkout and is already flagged by the author as needing a follow-up update).
• CMTClassifier.tectonic_type_probabilities's new moho_model.depth(lat, lon) call happens even when the point is outside the slab footprint and the result is discarded by the early-return in the module-level function — harmless, not worth a line comment.

Didn't find anything to raise as an inline comment.