Skip to content

Changelog

v0.4.4 (2026-06-21)

Multi-area études + a real fix to inter-area connectivity. Études 6 and 7 are implemented (previously placeholders), built on the built-in Izhikevich emitter and the canonical E:I profile. Implementing them surfaced and fixed genuine bugs in the multi-area connectivity path.

Fixed

  • inter_column_connectivity now accumulates specs instead of overwriting — multiple calls (one per directed projection) all materialize. Previously every call replaced the single metadata key, so build_multi_area_columns wired only the last adjacent pair.
  • build_multi_area_columns wires each adjacent pair in both directions: feedforward lo→hi (L2/3 → L4) and genuine top-down feedback hi→lo (L6 → L1/L5). Its p_feedback now produces real top-down edges.
  • _interarea_W layer matching tolerates the merged 5-layer scheme ("L2/3") as well as split "L2"/"L3" columns, so feedforward from superficial layers wires under the default DEFAULT_LAYERS.

Added

  • Étude 6 — Multi-Area Network: V1→V2→V4→PFC hierarchy with feedforward + feedback, inter-areal connectivity census, interactive 3D layout, per-area depth rasters, and an async-irregular (κ) gate.
  • Étude 7 — Multi-Trial Spectrolaminar Motif: multi-trial depth × frequency relative power on a 1k canonical column, with the synchrony (κ) trust gate and the scale-emergent crossover caveat.

v0.4.3 (2026-06-21)

Jaxley emitters reach the field stage, plus the flagship Configuration Grammar guide. A Jaxley Hodgkin–Huxley network can now drive a laminar LFP/CSD readout from a physically meaningful generator — its reconstructed transmembrane ionic current — closing Emitter → Source → Field → Probe for the Jaxley bridge. jaxfne is the mathematical backend; the biophysical fidelity of the readout follows the model you provide.

Added

  • JaxleyBridge.simulate_laminar_field(...) — run a Jaxley HH model and return Signals with a laminar FieldOutput (lfp_proxy/csd_proxy/...). The source is the reconstructed HH ionic current I_Na + I_K + I_Leak (from recorded HH_m/HH_h/HH_n gating states + channel params), projected via project_laminar_sources. signals.get("lfp_proxy") and jaxfne.vis.lfp/csd work directly on the result. Izhikevich/Fire are non-capacitive (zero current) and raise a clear error.
  • normalize_depth option (default on) min-max rescales the depth axis to [0, 1] so arbitrary-scale (µm) Jaxley geometry maps onto the projection's contact span instead of collapsing onto one contact; the raw range is preserved in metadata.
  • Configuration Grammar guide — the flagship documentation: Configuration framed as the compiler whose declarative specification compiles into the Emitter→Source→Field→Probe→Objective→Optimizer→Manifest chain, with each of the eleven sections mapped to its real fluent method and described as a biophysical-specificity dial.
  • Homeostasis guide — the homeostatic excitability controller (one parameter k_gain, k_gain=0 null), covering both the built-in per-step kernel (runtime(enable_homeostasis=True, homeostasis_params=...), diagnostics via model.last_homeostasis_diagnostics()) and the Jaxley outer-loop windowed controller.
  • Bridges API reference (api/bridges.md) — require_jaxley/clip shim, jaxley_to_signals, and all three JaxleyBridge run modes.

v0.4.2 (2026-06-20)

Homeostatic numerical stability — hard-bounded, float32-safe emitters. Enabling the homeostatic mode (one knob, gain k≈1.0) keeps each emitter in a stable operating regime AND hard-bounds its state, so a single neuron or simple component can never overflow or underflow in float32. The same mode eliminates hyperactivity and hypoactivity and provides short-term adaptation. Proxy/scaffold gates unchanged.

Added

  • Hard state bounds in the built-in Izhikevich homeostatic kernel (v_floor/v_ceiling/u_abs_max/syn_abs_max, tunable via homeostasis_params) — set far outside normal dynamics, so they never alter behaviour, only catch overflow/underflow. State stays finite in float32 under any finite (or even ±inf) input current.
  • JaxleyBridge.simulate_homeostatic(..., current_clip_nA=, strict_finite=) — the injected current is hard-bounded to keep the implicit solver away from overflow; output finiteness is verified (raises in strict mode rather than silently masking). Metadata records state_hard_bounded and all_finite.

Notes

  • The bounds are a safety net: in the normal regime the clamps stay inactive.

v0.4.1 (2026-06-20)

Jaxley emitters become a first-class path, with tfne homeostasis on top. Proxy/scaffold gates unchanged.

Added

  • jaxley_to_signals(module, recordings, dt_ms=...) — convert a jaxley.integrate recordings array (n_rec, n_time) to a jaxfne Signals, pulling recorded-compartment xyz from module.nodes into metadata for downstream projection (exported at the root).
  • JaxleyBridge.simulate(...) — run a Jaxley model end-to-end (stable bwd_euler) and return proxy Signals (previously a placeholder).
  • JaxleyBridge.simulate_homeostatic(...) — outer-loop windowed homeostatic excitability controller around a Jaxley emitter, applying tfne's restoring-bias law g = clip(k_gain·(target_rate_hz − r), g_min, g_max) as a per-cell current via grad-safe data_stimulate, stitched with continuous state resume. Computational control proxy only (biological_learning_claim=False, mechanism_claim_status="not_claimed").
  • hh_jaxley_reference_trace(...) — real single-compartment Hodgkin–Huxley reference trace (previously a placeholder).
  • ED10 release-archive receipt (scripts/ed10_release_archive_receipt.py) — binds release identity + truth gates to content hashes of the upstream evidence bundle (ED9), then self-hashes.

Fixed

  • require_jaxley() lazily installs a backward-compatible jnp.clip(a_min=/a_max=) shim so Jaxley (≤0.13) channel emitters integrate on current JAX (the shim self-disables when Jaxley adopts min=/max=). Metadata records jax_clip_compat_installed.

Changed

  • jaxley optional-dependency extra floored to >=0.13.0 (tested version).
  • CI now installs the jaxley extra so the Jaxley integration is exercised (drift-tested) on every push.

v0.4.0 (2026-06-18)

Fluent Configuration grammar + proxy-only probe vocabulary.

Added

  • Configuration.geometry(layer_thickness=...) — declare laminar geometry from per-layer thickness (normalized to cumulative z-intervals).
  • Configuration.population(N, neurons={...}) — per-layer neuron budget decoupled from thickness (largest-remainder allocation; one N per area).
  • Real inter-area edge wiring in Configuration.inter_column_connectivity(...): materializes cross-area synapses with anatomical routing (feedforward L2/3→L4, feedback L6→L1/L5) and an explicit layer_to_layer_map override.

Changed (breaking)

  • The *_like probe vocabulary (lfp_like/csd_like/eeg_like/meg_like) is fully retired with no aliases. Use *_proxy names only; signal access and probe declarations reject *_like with a clear pointer to the *_proxy name.

Fixed

  • CI workflow scope:→strategy: (matrix jobs now run); retired-term doc guard; mkdocs strict-build links.
  • Updated all tutorial notebooks to execute against the current API.

Scope

  • Outputs remain proxy readouts (field_solver_status=linear_solver, physical_amplitude_calibrated=false). No scientific-claim escalation.

v0.3.42 (2026-06-14)

Public context hardening release. - Added missing docstrings to stub emitters (GLIFEmitter, LIFEmitter). - Enforced clean-venv lazy root imports to guarantee heavy optional packages are not loaded on import. - Harmonized version references across documentation, package configurations, and tests.

v0.3.41 (2026-06-14)

Port of JAX computational kernels. No scientific-claim escalation.

Added

  • JAX spectral analysis functions: spectrolaminar_psd_jax, bandpower_jax, spectrolaminar_readout_kernel_jax, spectrolaminar_similarity_kernel_jax, and vectorized batched variants spectrolaminar_similarity_candidates_jax, spectrolaminar_similarity_candidates_seeds_jax.
  • Tensorized static-shape connectivity rule compilation kernel compile_connection_rules_jax.
  • JAX-optimized activity-dependent STDP synaptic weight update kernel update_stdp_weights_jax.
  • StimulusSchedule.to_array_jax for compiling event schedules into JAX arrays.
  • Auto-JIT cache warming compile tracking under Model._warmup_times during JIT initialization.
  • Experimental volume conductor loud-fail skeleton solve_volume_conductor_experimental that raises NotImplementedError pending boundary/gauge/calibration validation.

Scope

  • Outputs remain proxy readouts (linear_solver, physical_amplitude_calibrated=false).

v0.3.40 (2026-06-13)

Pre-0.4.0 hardening + device flexibility. No scientific-claim escalation.

Added

  • RuntimeConfig.backend (cpu/gpu/tpu) is now honored for device placement — simulate() pins compile+execute via jax.default_device; runtime_report() reports backend_enforced by device availability (honest downgrade when absent). dtype (float32/float64) and jit remain per-RuntimeConfig adjustable.
  • Accurate docstrings for all public jaxfne.__all__ members; fail-loud pre-0.4 physical-solver placeholder in experimental_hpc (raises until a validated solver exists — the stable path stays linear_solver).

Changed

  • Repo hygiene: root thinned; non-public planning/roadmap material removed from the published repository.

Scope

  • Outputs remain proxy readouts (linear_solver, physical_amplitude_calibrated=false); computational scaffold only.

v0.3.39 (2026-06-13)

Stable release: packaging/docs/PyPI consolidation of the v0.3.37–v0.3.38 quality line. No new APIs, no solver work, no scientific-claim escalation.

Summary of the v0.3.37 → v0.3.39 quality line

  • v0.3.37 — docs comparison: strict notebook grammar, root export grammar, scope/truth-gate cells, *_proxy naming, API index regenerated.
  • v0.3.38 — root-export hardening: the six export helpers (save_figure, save_figures, export_report, export_tutorial_artifacts, plot_raster, plot_spectrolaminar_suite) promoted to formal jaxfne.__all__ members (len(__all__) 179 → 185); clean-venv root-import laziness verified.
  • v0.3.39 — stable packaging: version bump to 0.3.39 across pyproject/core/mkdocs/docs; wheel + sdist rebuilt and twine-checked; published to PyPI, GitHub Release, and docs.

Validation

  • Targeted pytest, mkdocs build --strict, and compileall pass.
  • Root import loads no matplotlib/plotly/pandas/optax/jaxley/pynwb.
  • API index count equals runtime len(jaxfne.__all__) = 185.

Scope

  • Outputs remain proxy readouts (linear_solver, physical_amplitude_calibrated=false); computational scaffold, not a calibrated biological simulator.

v0.3.37 (2026-06-13)

Release: Strict notebook grammar + truth gates. Published to PyPI (wheel + sdist), GitHub Release (tag v0.3.37, commit 49aa025), and docs.

Added

  • Root-level export grammar: save_figure, save_figures, export_report, export_tutorial_artifacts, plot_raster, plot_spectrolaminar_suite — the canonical replacement for direct matplotlib/json calls in release-facing notebooks.
  • Scientific scope cells in all 15 release-facing tutorials documenting computational_scaffold / proxy_readout status, local nonlinearity, global linearity, and the physical_amplitude_calibrated = False truth gate.

Changed

  • Strict notebook call grammar: root-level jtfne.<fn>() only (no jtfne.vis.* / jtfne.tutorial_utils.* in public notebooks).
  • Tensor-field naming standardized to *_proxy (e.g. LFP-proxy, CSD-proxy); public *_like wording removed.

Validation

  • 2284/2284 tests pass; mkdocs build --strict passes; twine check passes on both artifacts.
  • Truth gates enforced: field_solver_status = "linear_solver", physical_amplitude_calibrated = False; no physical EEG/MEG/LFP/CSD measurement wording.

Scope

  • Outputs remain proxy readouts (linear_solver, physical_amplitude_calibrated=false); the package is a computational scaffold, not a calibrated biological simulator.

v0.3.26 (2026-06-02)

Feature release: Multi-area laminar workshop — inter-area connectivity, lesioning, AGSDR tuning, waveform explorer.

Added

  • Global multi-area network in simulate_laminar_trials with within-area recurrence + inter-area projections (connectivity_spec) and lesioning (lesion_spec); per-area depth leadfield. Defaults: feedforward V1 L2/3 (E) → V4 L4, feedback V4 L2/3 (E) → V1 L5/6.
  • Per-cell-type Izhikevich overrides (cell_type_izh_params, wider "E-Wide" default), single_cell_waveforms, and tune_laminar_agsdr (AGSDR tuning of firing rate + kappa).
  • Étude No. 1 rebuilt as a two-area Colab-ready customizable workshop notebook.

Fixed

  • Trial-averaged spectral power (per-trial PSD then mean); relative-power-density cross; uniform dt_ms kwarg across vis plotters; example API drift; GHA checkout@v5 / setup-python@v6.

Scope

  • Proxy readouts only (linear_solver, physical_amplitude_calibrated=false); spectrolaminar motif is emergent, not imposed.

v0.3.25 (2026-06-01)

Feature release: Cylindric scaffold, spectrolaminar motif, and 32-contact LFP artifacts.

Added

  • Cylindric scaffold geometry: per-area vertical cylinders (circular x-y cross-section, depth on z); visualize_network_3d gains column_shape="cylinder" shells and per-cell-type marker symbols.
  • Genuine per-area three-panel spectrolaminar suite (cell density, relative power spectrum, alpha-beta/gamma crossing) with smoothed band profiles.
  • 32-contact LFP/spectrolaminar artifacts: equal-spaced contacts, robust relative-power normalization excluding degenerate low-power channels, 32-channel stacked LFP waterfall.

Fixed

  • Notebook controls for trials, contacts, and per-layer cell-type fractions; CI dev-extra (ipykernel) for slow notebook-execution tests.

Scope

  • Outputs remain proxy readouts with field_solver_status=linear_solver and physical_amplitude_calibrated=false.

v0.3.24 (2026-06-01)

Release candidate: Compact Etude No. 1 tutorial utilities and Colab-ready visualization flow.

Added

  • jaxfne.tutorial_utils helpers for laminar column configuration, adjustable model scale, cell-type distributions, stimuli, trial simulation, spectrolaminar summaries, and strict artifact export.
  • jaxfne.vis tutorial panels for 3D scaffold, activity traces, and three-panel spectrolaminar figures.
  • Thin Etude No. 1 notebook using import jaxfne as jtfne and package-level helper calls.

Fixed

  • Preserved root build_laminar_column while adding build_tutorial_laminar_column as an unambiguous tutorial alias.
  • Restored optional dependency isolation: core import does not require pandas, matplotlib, plotly, or ipywidgets.
  • Added per-area LFP/CSD tensor shape (trials, areas, T, contacts) and area-tagged spectrolaminar specs.

Scope

  • Outputs remain proxy readouts with field_solver_status=linear_solver and physical_amplitude_calibrated=false.

v0.3.22 (2026-05-31)

Patch release: Fix jtfne.vis.visualize_network_3d missing from PyPI wheel.

Fixed

  • jaxfne/vis/__init__.py now exports visualize_network_3d in the published wheel. The v0.3.21 PyPI wheel was built before visualize_network_3d was added to the vis subpackage export list; v0.3.22 corrects this.
  • Etude No. 1 notebook: install cell updated to --force-reinstall --no-cache-dir to prevent stale Colab runtime cache from shadowing the fix.
  • Etude No. 1 notebook: install verification cell raises RuntimeError with clear path/version diagnostics if the installed jtfne.vis is still missing the function.
  • Network visualization cell now catches both ImportError and AttributeError and falls back to a static geometry3d PNG so the notebook completes regardless.

Added

  • tests/test_vis_network3d_public_api.py — 5 regression tests that enforce the jtfne.vis.visualize_network_3d export contract. CI now fails loudly if this public API is ever dropped from the package.

v0.3.21 (2026-05-30)

Release: Etude No. 1 completion and notebook template standardization.

Added

  • Added Etude No. 1 as an advanced multi-laminar cortical AGSDR workflow under tutorials/etudes/.
  • Added a canonical notebook template under tutorials/templates/ with unified setup, truth gates, and placeholder configuration.
  • Added a template guide for Suites and Etudes.

Changed

  • Cleaned duplicated Etude notebook artifacts.
  • Moved release/comparison receipts into internal_docs/release_receipts/.
  • Updated release checklist and agent status metadata for the v0.3.21 release candidate.

Validation status

  • Package import and compile gates pass.
  • Etude and template notebooks pass structural hygiene checks.
  • Maintains `,computational_scaffold,field_solver_status=linear_solver, andphysical_amplitude_calibrated=false`.

v0.3.19 (2026-05-30)

Release: Field proxy boundary handling improvements.

  • Optimized project_laminar_sources boundary fallbacks for low contact counts.
  • Added comprehensive boundary and stencil numerical parity tests.
  • Maintains `,linear_solver` status.

v0.3.18 (2026-05-30)

Release: Sharding infrastructure for multi-device AGSDR.

  • Added jaxfne/sharding_utils.py with distributed mesh and NamedSharding stubs.
  • 14 new tests for sharding context and single-device fallbacks.
  • Sharding stubs do not yet drive multi-device dispatch (reserved for v0.3.20+).

v0.3.17 (2026-05-30)

Release: Dtype inheritance in AGSDR optimization.

  • Updated AGSDR loop to inherit dtype from bounds, not force float32.
  • Applied dtype-inheritance to noise generation, W_init, and delta-rule center updates.
  • 12 new tests covering dtype invariants and candidate clipping.

v0.2.3 (2026-05-19)

Release: Stable proxy operators and documentation infrastructure.

  • Added MkDocs-based documentation site with Material theme
  • Reorganized docs: tutorials, guides, API reference, about
  • Added Jaxley interoperability documentation
  • Cleaned public documentation (removed internal metadata)
  • 492 tests passing

v0.2.1 (2026-05-10)

  • Introduced probe operator contracts
  • Added status metadata
  • Eight canonical readout channels

v0.2.0 (2026-04-15)

  • Initial release
  • Izhikevich emitters
  • Laminar field solver (proxy)
  • Basic readout operators