Version 2.3.0#

pyCSAMT 2.3.0 New Fix API Change Docs Build#

Released 2026-08-14.

A minor release built around a new pycsamt.geology package, a new Dual-Uncertainty Hybrid Inversion (DUHI) pathway, and a substantially smarter processing pipeline. Rock resistivity classification, borehole logs, and (new in this release) structural-geology field measurements move out of pycsamt.interp into their own EM-agnostic package, alongside a pluggable architecture for sourcing rock-property tables. The release deep-fills three previously stub pages of the interpretation user guide, which – as with every doctest-verified page in this project – surfaced and fixed several real bugs along the way. On the inversion side, DUHIInverter2D and a native Occam2D prejudice-file reader/writer connect a trained AI ensemble’s mean/uncertainty to the physical solver’s model-space regularization, alongside a reusable synthetic benchmark-geology generator and a field-calibrated corruption model – exercised end to end against a real compiled Occam2D binary and a real trained ensemble, which surfaced two further real bugs. Taking a real legacy CSAMT survey line all the way from raw Zonge AVG through EDI conversion and the processing pipeline surfaced and fixed six further real bugs, plus a packaging gap that silently dropped the fallback EPSG table from installed wheels. Finally, the pipeline itself gains four method-aware presets with real near-field correction and data-driven QC, a branded dashboard report, and data-driven power-line-harmonic detection, alongside a full rewrite of the config-driven-pipeline tutorial and CLI reference that – once again – surfaced real bugs simply by trying to write accurate examples.

A new geology package New#

pycsamt.interp had accumulated general earth-science domain knowledge – rock resistivity classification, borehole logs – that has nothing to do with electromagnetics, alongside the EM-specific concepts that actually need it (ResistivityModel, ModelCalibrator). pycsamt.geology is the new home for the former: lithology (RockDatabase, RockEntry, StratigraphicLog), borehole (Borehole, Interval), and the literature-compiled table behind RockDatabase.default() (BUILTIN_ROCKS). pycsamt.interp re-exports all of these for backward compatibility, so existing imports keep working; new code should reach for pycsamt.geology directly. This follows the same top-level-compatible split already used for pycsamt.core/pycsamt.transformers.

The rock-property table can now also come from somewhere other than the built-in literature compilation. RockPropertyProvider is a small protocol with two implementations – LocalRockPropertyProvider for a file on disk and RemoteRockPropertyProvider for a URL – wired up as RockDatabase.from_url() and RockDatabase.from_provider(). Remote fetches are cached locally under ~/.pycsamt/rock_db/ with a TTL and quietly fall back to the built-in table if the fetch fails, the same convention already used by pycsamt.ai’s pretrained-model cache. No public dataset of resistivity-labelled rock samples at the scale of the built-in table actually exists (checked against Data Series 595, the Geochemical and Geophysical Characteristics of the Conterminous US dataset, and Macrostrat’s REST API – all either too narrow, resistivity-free, or lithology-only), so the built-in table itself was expanded directly from Palacky, Telford, and Keller instead.

The package also gains field structural-geology primitives, previously an empty placeholder module: StructuralMeasurement records a planar measurement (strike/dip/dip-direction) and cross-validates dip-direction against strike so a transposed field-notebook entry is caught rather than silently accepted; LinearMeasurement records a linear one (trend/plunge); FaultTrace records where a fault crosses a profile, with its downthrown side and sense. StructuralModel collects all three against a profile position with CSV I/O and nearest/within queries – the piece needed to eventually back the structural-continuity review step described in Interpretation workflow with real data rather than leaving it as an unbacked checklist item.

Auditing this split also meant checking whether the classes involved actually follow this project’s own base-class conventions. They mostly did – Interval and RockEntry were already consistent – except ResistivityModel, which predated the convention and was a bare dataclasses.dataclass() with none of it. It now inherits PyCSAMTObject/MetadataMixin like every other model container in pycsamt.interp, picking up auto-repr, to_dict(), clone(), and update() for free, plus a new metadata field for attaching free-form provenance (coordinate reference, backend version, original file paths) to a model after the fact.

DUHI: hybrid AI-physics inversion infrastructure New#

Built out of an in-progress Dual-Uncertainty Hybrid Inversion (DUHI) paper reproduction, this release adds the pieces needed to run a real AI-physics hybrid inversion end to end, not just each half in isolation.

DUHIInverter2D is the connecting piece. Given a completed InputBuilder project and an AI ensemble’s mean and predictive-uncertainty grids, it replaces nominal datum errors with reliability-weighted effective errors (apply_observation_reliability), maps the AI grids onto the Occam mesh’s own parameterization (map_ai_grid_to_occam()), optionally seeds the Occam startup vector with the AI mean, and writes an uncertainty-weighted model-space prejudice constraint through the new OccamPrejudice – a full reader/writer for the sparse OCCAM2MTPREJ_2.0 format, including the target/weight encoding the bundled solver expects (prewt**2 in the penalty Hessian, prewt*premod in the right-hand side, so the quadratic penalty stays centred on the requested target as the weight changes).

Reliability itself is now a first-class, independently testable quantity rather than an inline calculation: dimensionality_reliability() scores phase-tensor skew against 2-D compatibility, and combine_observation_reliability() combines it with measurement quality into one bounded factor (pycsamt.ai.inversion). On the physics side, run_forward() runs the bundled solver’s native forward-only -F mode directly, without a full iterative inversion – useful for generating synthetic truth responses or checking a proposed model’s misfit before committing to refinement.

The synthetic side gained matching infrastructure. New in pycsamt.ai.geology, generate_benchmark_geology() and BenchmarkGeology generate the six in-distribution and six out-of-distribution geological families (ID_BENCHMARK_FAMILIES, OOD_BENCHMARK_FAMILIES) used to build a reproducible, seeded publication-scale training/validation/test benchmark. New in pycsamt.ai.domain_gap, apply_empirical_corruption() applies a field-calibrated corruption model – station-dependent static shift, frequency- and component-dependent noise, and missing observations, jointly resampled from real field-survey profiles rather than drawn as independent per-component noise. And build_2d_maxwell_problem() factors out the validated geology-grid-to-solver-problem construction that this whole pipeline is built on.

Running this pathway end to end against a real compiled Occam2D binary and a real trained AI ensemble – not just unit tests – surfaced two further real bugs, both now fixed. PosteriorCalibrator.calibrated_std and .predict_posterior divided the raw ensemble standard deviation by the learned dispersion scale instead of multiplying, which shrinks an already under-dispersed uncertainty estimate instead of widening it; every existing unit test happened to exercise already-correctly-scaled synthetic data, so the wrong direction was never caught. Separately, InversionResult’s rho_2d reconstruction started assigning model-layer values at mesh row zero instead of the first row after the air layers, silently shifting the recovered resistivity section by the air-layer count whenever a mesh had any – invisible for the air-layer-free meshes exercised by prior tests.

Deep-filling the interpretation user guide Docs#

Three stub pages under Interpretation were deep-filled with narrative content, doctest-verified pycon examples, and regenerated figures, following this project’s usual discipline of computing every numeric claim for real before writing it down. Lithology classification covers nearest-midpoint resistivity classification against the new built-in rock table. Petrophysical toolkit covers Archie’s law, Waxman-Smits clay conductivity, Hashin-Shtrikman physical-admissibility bounds, Kozeny-Carman hydraulic conductivity, and water-table detection from a resistivity profile, with five figures. Monitoring and fusion covers time-lapse EM monitoring and multi-method fusion (linear, sigmoid, and RMS-weighted blending), also with five figures. The glossary gained a “Nearest-midpoint classification” entry to match.

Bugs found along the way Fix#

Verifying every example against real code, as always, found real bugs rather than just documentation gaps. Several docstring examples in ArchieModel.forward/.saturation, WaxmanSmitsModel.forward, kozeny_carman_K, and HashinShtrikmanBounds claimed stale output values that no longer matched what the code actually computes; all four are corrected. Separately, RockDatabase had no public way to iterate its entries – the desktop app’s InterpController.plot_rock_db() called list(db) on a non-iterable object outright (raising at plot time), and the pycsamt rocks CLI command worked around the same gap by reaching into the private _entries attribute. Both are fixed against a new public RockDatabase.entries read-only accessor; the CLI table also gained a source column so a classification’s literature provenance is visible from the command line.

Hardening the Zonge AVG-to-EDI pipeline Fix#

A real legacy CSAMT survey line – raw Zonge AVG data plus a .stn topography file, converted to EDI and run through the processing pipeline’s full_processing preset – turned up six more real bugs, none of them hypothetical: every one reproduced on the first attempt to read, convert, or process this specific field dataset.

Reading the raw survey exposed three parsing gaps. The legacy (kind-1) AVG parser (_parse_kind1) assumed every data row has exactly as many fields as the header line names; a real AMTAVG 7.40 file consistently carried one unlabeled trailing field beyond the header’s sPhz column, which raised a pandas column-count mismatch instead of loading. The companion .stn topography reader failed for two independent reasons on the same file: a stray <input type="hidden" /> fragment (evidently pasted in from a browser at some point) corrupted the header line, and the station-coordinate columns were named GridE/GridN rather than the East/North spelling the column-name matcher expected. All three are fixed – the AVG parser now extends the header with generic names when every row carries the same consistent surplus, and the .stn reader strips stray markup and recognizes the GridE/GridN convention.

Converting the result to EDI surfaced a more consequential bug: AVGtoEDI copied Zonge’s impedance – computed in SI ohms by pycsamt.zonge.z.Z, by design – directly into the EDI file’s Z section. But the EDI/SEG standard, and the rho = 0.2|Z|^2/f formula pycsamt.z.z.Z uses to derive apparent resistivity from it, both assume impedance in (mV/km)/nT field units. Every AVG-derived EDI’s resistivity and phase were consequently wrong by a factor of about (mu_0 x 1e3)^2 – roughly six orders of magnitude too small for resistivity – while the raw ZXYR/ZXYI values written to the file looked plausible in isolation, which is exactly why this kind of unit bug survives casual inspection. Fixed by adding z_ohms_to_mvk_nt() (the inverse of the existing z_mvk_nt_to_ohms) and applying it in AVGtoEDI.emit_edi before the impedance is written.

Running the converted EDIs through the full_processing pipeline preset found two more. The skew-masking steps (SK001/SK002) had their registry defaults keyed threshold=, but mask_by_skew and keep_longest_low_skew actually take thresh= – a TypeError on every run, silently downgraded to a no-op warning by the pipeline’s default error handling, so skew masking never actually happened in either the full_processing or publication_ready presets. Once that was fixed and skew masking started running for real, it exposed a second, more serious bug one step downstream: static-shift correction (SS001) began crashing with an IndexError. The root cause was pycsamt.emtools.ss._nearest_idx being called with its arguments reversed at four call sites (estimate_ss_ama, estimate_ss_loess/ estimate_ss_refmedian, detect_near_surface, _pt_phi_for_station) – each one aligned a station’s own frequency axis against phase-tensor-table row positions instead of the other way around, which happens to raise loudly when the two arrays differ in length (as they now more often did, post-masking) but would otherwise have silently applied the skew filter to the wrong frequencies. All four call sites are fixed.

Finally, a QC plot investigated along the way (nr_qc_harmonic_waterfall) turned out not to be broken so much as uninformative: for any CSAMT log-decade frequency sweep, none of the survey’s discrete frequencies land within tolerance of an actual 50 Hz mains harmonic, so the reduction matrix it plots was entirely NaN and rendered as a blank heatmap with a meaningless auto-ranged colorbar. It now detects the all-NaN case and shows an explanatory message instead.

A branded pipeline dashboard report New#

The pipeline’s HTML report (report.html, written whenever save_report=True) was a plain step-card list with a generic, unbranded palette and no charts. A second, opt-in report tier now sits alongside it: dashboard.html, added to report_formats (or via the CLI’s --dashboard flag) without changing the default report at all. It carries pyCSAMT’s real brand identity – the actual logo mark and favicon from docs/source/_static/logo/, embedded rather than read from disk since that directory isn’t shipped in the installed package – KPI stat tiles (steps ok/total, errors, total time, sites in/out, cache hit rate, figures generated), and three native inline-SVG charts built from the same per-step data the plain report already collects: a per-step status timeline, a duration bar chart (steps at or above the run’s own 80th-percentile time highlighted), and a two-series site-count line chart. Charts are hand-built SVG, not a plotting-library dependency, so the file stays self-contained – no external JavaScript or CDN.

Every color used, including the brand’s own blue/orange pair for the site-count chart, was checked against pyCSAMT’s real light and dark surfaces with a colour-blindness and contrast validator before being wired in, rather than chosen by eye; both passed without substitution. The existing plain report’s palette was also swapped from generic blue/green/red to the same real brand tokens, at no cost to how cheaply it renders.

Method-aware pipeline presets New#

Every pipeline preset up to this release was chosen by processing intent only – basic_qc, full_processing, and the rest apply the same step sequence regardless of whether the survey is MT, AMT, CSAMT, or CSUMT. In reality these methods are not interchangeable: CSAMT/CSUMT use a controlled source and can suffer near-field/transition-zone contamination that MT/AMT never see, a single-component TE- or TM-only CSAMT line cannot produce a meaningful phase-tensor ellipse, and tipper is only sometimes recorded. Four new presets – PRESETS’s mt_qc, amt_qc, csamt_qc, and csumt_qc – account for this, selected explicitly by name (or via the new get_preset_for_method(), which maps an EM method string to the matching preset) rather than any automatic detection from data.

Two new categories of registry step back them, usable by any pipeline, not just these four. Three “smart” diagnostic steps (tensor_qc_smart, tipper_qc_smart, strike_qc) check the actual data before drawing a figure – a QC plotting hook is always called as fn(sites) with no channel for extra parameters, so each wrapper inspects the site collection itself and returns None (silently skipped by the existing step-execution engine, no core changes needed) when a plot would not be meaningful: a phase-tensor ellipse only draws for multi-component data, and the five tipper-dependent plots only draw when a real tipper channel is present. Two more steps (raw_preview, processed_preview) open and close a preset with a deterministically random subset of stations plotted raw, then again processed, so a before/after comparison is always available without hunting through the full station list.

csamt_qc/csumt_qc are also the first presets ever to chain the near-field-correction step (SRC001, already registered but never used by any preset before this release) with source_offset=None – resolve the real transmitter-receiver separation from each station’s own metadata, and warn instead of failing when nothing resolves, rather than requiring the caller to know it up front. Being the first real caller surfaced a genuine bug: the companion source-response-normalization step (SRC002) is registered as returns_sites=True but its underlying function actually returns a tidy per-(station, frequency) DataFrame, not a Sites object. Chaining it silently corrupted the site count (n_stations real sites became n_stations x n_frequencies “sites”, then collapsed to zero at the next step) with no error raised – nothing before csamt_qc had ever put SRC002 in a step sequence to notice. Fixed by marking it returns_sites=False, matching what it actually is: a diagnostic analytics function whose real output is its QC plot.

Modernizing the config-driven pipeline docs Docs#

Run a Pipeline From Config and Pipeline Commands predated every pipeline feature added this release plus the step cache, live progress, and run-history log from the previous one – both are rewritten to cover them. The tutorial’s “Adapting the Example to AMT and CSAMT” section, previously just “manually override band_hz”, is replaced with a worked walkthrough of the new method-aware presets: the same familiar L18PLT AMT line run through amt_qc (showing tipper QC correctly skip itself), plus a new short example against the real 10-station Tongkeng CSAMT line (data/CSAMT) demonstrating genuine near-field-correction warnings and a new field-zone pseudosection figure. A combined “Beyond the Basics” section walks --cache, --live, --history, and --dashboard together on one real run, including an honest note that QC-figure generation is not itself cache-aware – a cache hit skips the step transform only, so total wall time does not collapse to near-zero on an already-fast config even though the processing itself is genuinely skipped. Every interactive example on both pages was also converted from plain python code blocks to pycon, per this project’s documentation guidelines, with every shown output re-captured from a real run rather than trusted as still accurate.

Writing real, runnable examples for the CLI reference again found real bugs rather than just stale prose: pipe run --help’s --preset option listed only 6 of what are now 11 registered presets, and pipe steps --help claimed “33 registered steps” and listed only 8 of what are now 10 step categories – both counts predate several rounds of step-registry growth this project has been through. Both are corrected.

Data-driven mains-frequency detection New#

notch_powerline (pipeline step NR001) suppresses power-line harmonics by matching real sampled frequencies against exact multiples of a fixed mains_hz within a narrow +-tol_hz window. Real EDI frequency grids are usually log-spaced rather than sampled exactly on 50/60 Hz multiples, so a harmonic can fall outside that window and go un-notched with no warning – exactly the gap behind this release’s nr_qc_harmonic_waterfall fix above.

mains_hz now also accepts the literal string "auto". There are only two real-world AC grid frequencies in practical use, so no numeric hint is needed: pyCSAMT scores 50 Hz and 60 Hz against the survey’s own pooled frequency array and resolves to whichever explains more harmonics, then snaps each harmonic to its single nearest real sample within a relative tolerance (snap_frac, default 1% – deliberately tight, since real AC grids drift far less than that, and a loose window just finds the nearest point on a coarse log-spaced grid regardless of whether it is really mains-related). An early, more permissive design that snapped within 10% turned out to “detect” a fundamental almost everywhere, including on the sparse, non-mains-aware Tongkeng CSAMT grid, purely from numerical proximity on a coarse schedule rather than genuine contamination – caught by checking the resolved frequency and the individual harmonics against real data, not just confirming the call didn’t raise. The shipped version adds a minimum-evidence gate: if neither 50 Hz nor 60 Hz explains at least a handful of harmonics this tightly, "auto" leaves the survey untouched and warns, rather than guessing and notching an unrelated frequency. On the real 28-station L18PLT AMT line it confidently resolves to 60 Hz (4 of 30 harmonics matched within 1%) and notches them; on the real, coarse 10-point Tongkeng CSAMT grid it correctly finds no reliable signature and leaves the data completely unchanged. Passing a plain number is unaffected – output is identical to every previous release.

Fixed#

  • Fix Stale petrophysics docstring examples – corrected output values in ArchieModel.forward/.saturation, WaxmanSmitsModel.forward, kozeny_carman_K, and HashinShtrikmanBounds.

  • Fix ``RockDatabase`` iteration – added a public RockDatabase.entries read-only tuple property; fixed InterpController.plot_rock_db() (previously called list(db) on a non-iterable object) and the pycsamt rocks CLI command (previously read the private _entries attribute) to use it.

  • Fix ``PosteriorCalibrator`` under-dispersion – calibrated_std and predict_posterior divided the raw ensemble standard deviation by the learned dispersion scale instead of multiplying, shrinking an already under-dispersed estimate instead of widening it; no existing test happened to exercise an under-dispersed case. Fixed both methods and added a regression test that specifically does.

  • Fix ``InversionResult.rho_2d`` air-layer offset – model-layer values were assigned starting at mesh row zero instead of the first row after the air layers, silently shifting the recovered resistivity section by the air-layer count whenever a mesh had any.

  • Fix Legacy AVG parser column-count mismatch – _parse_kind1 now extends the header with generic names when every data row consistently carries one or more unlabeled trailing fields beyond the header’s sPhz column, instead of raising on real AMTAVG 7.40 files.

  • Fix ``.stn`` topography reader – read_stn now strips stray HTML markup from the header line, and Topography._normalize_stn_columns now recognizes the GridE/ GridN grid-coordinate column-naming convention.

  • Fix ``AVGtoEDI`` impedance unit mismatch – Zonge’s SI-ohm Z was written directly into EDI Z sections, which the EDI/SEG standard and its own rho = 0.2|Z|^2/f formula assume are (mV/km)/nT field units; every AVG-derived EDI’s resistivity/phase were wrong by a factor of about (mu_0 x 1e3)^2. Fixed via the new MTBase.z_ohms_to_mvk_nt(), applied in AVGtoEDI.emit_edi.

  • Fix Pipeline ``SK001``/``SK002`` wrong keyword – registry defaults used threshold= where mask_by_skew/keep_longest_low_skew take thresh=, crashing the skew-masking step of the full_processing/publication_ready presets on every run.

  • Fix ``pycsamt.emtools.ss._nearest_idx`` reversed arguments – fixed at four call sites (estimate_ss_ama, estimate_ss_loess/ estimate_ss_refmedian, detect_near_surface, _pt_phi_for_station); previously crashed static-shift correction (SS001) with an IndexError whenever it ran downstream of skew masking.

  • Fix ``nr_qc_harmonic_waterfall`` blank plot – now shows an explanatory message instead of a blank heatmap when no station frequency falls within tolerance of a mains harmonic (as for any CSAMT log-decade sweep).

  • Fix Pipeline step ``SRC002`` corrupted site counts – normalize_response returns a DataFrame, not a Sites object, but was registered returns_sites=True; chaining it (as csamt_qc/ csumt_qc are the first presets to do) silently inflated the site count to n_stations x n_frequencies and then collapsed it to zero one step later, with no error raised. Fixed by registering it returns_sites=False.

  • Fix Stale CLI help text – pipe run --help’s --preset option listed only 6 of 11 registered presets; pipe steps --help claimed “33 registered steps” and 8 categories where there are now 55 steps across 10 categories. Both corrected.

  • Build Missing ``gis/epsg.npy`` package data – the fallback EPSG-code table used when pyproj’s legacy flat epsg file isn’t available was missing from pyproject.toml’s package-data and MANIFEST.in, so installed wheels silently shipped without it and any EPSG lookup that fell back to it failed at import time with “Failed to load EPSG definitions”.

Changed#

  • API Change ResistivityModel now inherits PyCSAMTObject/ MetadataMixin like the rest of pycsamt.interp’s model containers, and gained a metadata field.

Added#

  • New pycsamt.geology package: lithology, rock_library, rock_providers, borehole, and structural.

  • New StructuralMeasurement, LinearMeasurement, FaultTrace, and StructuralModel.

  • New RockDatabase.from_url() / RockDatabase.from_provider() with local caching and fallback-to-default behaviour.

  • New DUHIInverter2D and DUHIPreparation (pycsamt.ai.inversion) – connects a trained AI ensemble’s mean/uncertainty to Occam2D’s data-space errors and model-space prejudice.

  • New map_ai_grid_to_occam() and dimensionality_reliability() / combine_observation_reliability() (pycsamt.ai.inversion).

  • New OccamPrejudice – reader/writer for the sparse OCCAM2MTPREJ_2.0 model-prejudice format.

  • New run_forward() – runs the bundled solver’s native forward-only -F mode.

  • New generate_benchmark_geology() and BenchmarkGeology (pycsamt.ai.geology.benchmark) – the six in-distribution and six out-of-distribution geological families (ID_BENCHMARK_FAMILIES, OOD_BENCHMARK_FAMILIES) behind the reproducible publication-scale benchmark.

  • New apply_empirical_corruption() and EmpiricalCorruptionResult (pycsamt.ai.domain_gap.empirical) – field-calibrated, jointly-sampled static shift, noise, and missing-observation corruption.

  • New build_2d_maxwell_problem() – validated geology-grid-to-solver-problem construction, factored out of the dataset-generation pipeline for reuse.

  • Docs Deep-filled Lithology classification, Petrophysical toolkit, and Monitoring and fusion; added the “Nearest-midpoint classification” glossary term.

  • Docs Added the pycsamt.geology reference page; updated pycsamt.interp and Interpretation Commands to reflect the split.

  • New z_ohms_to_mvk_nt() – inverse of the existing z_mvk_nt_to_ohms, converting SI-ohm impedance to the standard EDI (mV/km)/nT field-unit convention.

  • New Branded dashboard report – dashboard.html (CLI --dashboard, or "dashboard" in report_formats): KPI stat tiles and three native inline-SVG charts alongside the existing plain report, whose palette also moved to the same real, validated brand tokens.

  • New Method-aware presets – mt_qc, amt_qc, csamt_qc, csumt_qc, and get_preset_for_method() (pycsamt.pipeline). CSAMT/CSUMT presets wire in real near-field correction for the first time; all four add data-driven “smart” tensor/tipper/strike QC (QC005-QC007) and a raw-vs-processed station preview (PRE001/PRE002), both usable by any pipeline.

  • New notch_powerline(..., mains_hz="auto") (pycsamt.emtools.remove_noise, pipeline step NR001) – detects 50 vs 60 Hz from the survey’s own frequency grid and snaps each harmonic to its nearest real sample instead of requiring an exact match; new snap_frac parameter controls the match tolerance. Plain numeric mains_hz is unaffected.

  • Docs Rewrote Run a Pipeline From Config and Pipeline Commands to cover method-aware presets, the step cache, live progress, the run-history log, and the dashboard report; converted every interactive example to pycon per the documentation guidelines. Added the “Run history log” and “Dashboard report” glossary terms and a new field-zone pseudosection figure.

Compatibility#

Default behaviour is unchanged for existing users. Borehole, Interval, RockDatabase, RockEntry, StratigraphicLog, and Layer remain importable from pycsamt.interp exactly as before; their canonical home is now pycsamt.geology. Anything that reached into RockDatabase._entries directly – a private attribute, not part of the public API – should switch to the new RockDatabase.entries. ResistivityModel’s new metadata field defaults to an empty dict and its new base classes add methods rather than removing any, so existing code that constructs or reads a ResistivityModel is unaffected. Everything under pycsamt.ai.inversion, pycsamt.ai.geology.benchmark, pycsamt.ai.domain_gap.empirical, and pycsamt.models.occam2d.prejudice is newly added, so there is nothing existing to break; the two bug fixes change output only for the specific under-dispersed-calibration and air-layer cases described above, which were silently wrong before.

The AVGtoEDI fix changes output for anyone who has previously converted a Zonge AVG survey to EDI with this project: apparent resistivity and phase in those EDIs were wrong by a factor of about (mu_0 x 1e3)^2 and should be regenerated from the source AVG with this release. EDIs produced by any other route (Jones .j, direct EDI authoring, third-party tools) are unaffected – the bug was isolated to AVGtoEDI.emit_edi. The pipeline fixes (SK001/SK002, SS001) mean the full_processing and publication_ready presets now actually run their skew-masking and static-shift steps instead of silently no-op’ing or crashing; re-running either preset on previously processed data may therefore change output where it previously did neither.

Everything in this section is additive and off by default. The dashboard report, the four method-aware presets, get_preset_for_method, the tensor_qc_smart/tipper_qc_smart/strike_qc/raw_preview/ processed_preview registry steps, and notch_powerline’s mains_hz="auto" are all new, opt-in surface with no effect on any existing call that doesn’t request them – a plain mains_hz=50 (the default used by every existing preset) produces output identical to every previous release. The SRC002 fix only changes output for a pipeline that already chains SRC001/SRC002 together, which before this release meant nothing: no built-in preset did so until csamt_qc/ csumt_qc were added in the same release.