Repository navigation
Correctness fixes, component tiers, and revised plan - #3
Merged
Merged
Conversation
The report's time-dependent results were computed on frames loaded out of time order, and its Lyapunov values came from the removed estimator. Replace the report with a withdrawal notice, remove it from the site navigation, and correct the README and docs index claims about chaos gating and the Wolf algorithm. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…ader The sort key matched the '2' in 'Vmem2D' for every file, so runs of ten or more frames loaded in directory order (0, 1, 10, 100, ...). Anchor the frame index to the end of the file stem and raise on unparseable names. Also add load_betse_cells() for analysis without interpolation, default to linear interpolation (cubic overshoots the data range), and report an inside-hull mask and frame indices in the metadata. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- 'sublevel' negated the field before handing it to GUDHI, which already computes sublevel persistence, so the two options were swapped. - Cells were flattened in C order; GUDHI expects the first axis fastest, which scrambled non-square fields. - Replace the threshold-counting fallback with an exact union-find H0 computation that matches GUDHI, and return empty diagrams for the dimensions it does not compute. - Rewrite the Wasserstein and bottleneck fallbacks (p-th power and root, own-diagonal matching only, essential classes ignored). The Wasserstein fallback previously crashed on infinite bars. - All fallbacks now emit RuntimeWarning. NaN fields raise instead of yielding empty diagrams. compute_cycles raises NotImplementedError instead of returning empty arrays. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…te test - Detectors in mneme.core.attractors assigned fixed-point, limit-cycle and strange labels from variance and spread thresholds, bypassing the surrogate gate. A sine wave and white noise were both labelled strange. They now report UNDETERMINED. - classify_attractor returned FIXED_POINT or LIMIT_CYCLE for any near-zero estimate, including white noise. Those labels now require the caller to assert whether the signal oscillates. - surrogate_test raises when the surrogate count cannot reach alpha (the scripts used n=30, whose smallest p-value is 0.065), and warns on multi-dimensional input and on series shorter than 4000 points. - Recurrence detector counted each time index once per recurrent pixel, giving basin sizes above 1. - Vectorise the divergence loop in largest_lyapunov. Results are unchanged to 1e-13 on ten reference signals; runtime drops about tenfold. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The default 'IFT' / 'sparse GP' reconstructor fits an exact GP to a random
subset of the observations and discards the rest. Rename it
SubsetGPReconstructor (method 'gp_subset'). The dense reconstructor is a
Wiener filter; rename it WienerFilterReconstructor ('wiener_filter').
Old class names, method names and n_inducing keep working and emit
DeprecationWarning.
Fixes found while adding ground-truth tests:
- Wiener filter response rows were not normalised, scaling the
reconstruction down by the number of grid points in a correlation length.
- Wiener filter correlation_length was in pixels but compared with
unit-square distances.
- optimize_hyperparameters=False still optimised the kernel.
- NeuralFieldReconstructor.uncertainty() returned zeros; it now raises
NotImplementedError and fit_reconstruct reports uncertainty=None.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- A stage that raised was logged and the run returned success=True. Runs now return success=False with the stage in failed_stages and its message in errors; the partial analysis result is still returned. - 'mneme analyze' without --config passed an empty config, which disabled every component. The CLI now overlays the user's config on the chosen pipeline's defaults, prints each stage's status, and exits non-zero on failure. - Without sparse observations the pipeline returned the input array as a 'reconstruction'. The stage is now reported as skipped. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Narrow what the project claims to what has been checked. - Add mneme.ExperimentalWarning. Attractor detectors, symbolic regression, the VAE and neural-field reconstruction warn when constructed. Default pipelines run core stages only. - docs/SCOPE.md defines core, frozen and experimental tiers. - docs/LYAPUNOV_OPERATING_RANGE.md records measured accuracy and detection power; the Lyapunov modules are frozen. Probe scripts and their outputs are in review_artifacts/2026-09-26. - docs/MULTISTABILITY_PROTOCOL.md and project_plan.md v3 re-order the work around the hypothesis the project is meant to test. - Comparative summaries report None, not 0, for analyses that did not run. The quality checker reports 'unknown' when one of its checks fails. - Remove withdrawn PhysioNet result files; mark the data acquisition plan superseded; update README, CLAUDE.md, CHANGELOG and the API reference. - CI installs GUDHI and POT, and the coverage floor rises from 35% to 60%. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Fixes the defects found in the September 2026 review, narrows what the project claims to what has been checked, and re-orders the plan.
This branch is stacked on #2 (report withdrawal). Merge #2 first, or merge this one and close #2.
Correctness fixes
Each has a regression test.
sublevelandsuperlevelwere swappedstrangeUNDETERMINEDFIXED_POINTUNDETERMINEDunless the caller assertsoscillatorysuccess=Truesuccess=False, withfailed_stagesanderrorsmneme analyzewithout--configdisabled every stageNaming
The default reconstructor fits a GP to a random subset of the observations. It is now
SubsetGPReconstructor(gp_subset). The dense reconstructor isWienerFilterReconstructor(wiener_filter). Old names still work and emitDeprecationWarning.Scope
docs/SCOPE.mdsorts components into core, frozen and experimental.mneme.ExperimentalWarning.docs/LYAPUNOV_OPERATING_RANGE.mdrecords their measured accuracy and detection power.Breaking changes
classify_attractor(..., oscillatory=...)defaults toNone.PersistentHomology(compute_cycles=True)raisesNotImplementedError.NeuralFieldReconstructor.uncertainty()raisesNotImplementedError.PipelineResult.successisFalsewhen any stage fails.Testing
Local run on Windows, Python 3.13, GUDHI without POT: 372 passed, 1 skipped, coverage 70.4%.
CI now installs GUDHI and POT, so this is the first run of the GUDHI code paths in CI. The coverage floor rises from 35% to 60%.
The
largest_lyapunovvectorisation was checked against the original on ten reference signals: identical fit regions, largest difference 2e-14.Not done
🤖 Generated with Claude Code