v0.1.3

Added

• CLI: decepticons --version flag.
• CLI: decepticons info subcommand — prints version, install path, and
  public surface size. Useful for bug reports and environment diagnosis.
• CLI: decepticons score subcommand — fits a model and prints bits/byte
  without sampling.
• CLI: decepticons fit --seed N for reproducible sampling.
• python -m decepticons works as an alias for the decepticons script.

Changed

• CLI: fit subcommand now uses the canonical ByteLatentPredictiveCoder
  class name internally instead of the legacy OpenPredictiveCoder alias.
• CLI help text for every subcommand now describes what it does.
• __all__ is now sorted case-insensitively (with __version__ pinned
  first), so additions land in a predictable place.
• Removed stale section-header comments in __init__.py that no longer
  matched the import order they sat over.

v0.1.2

Docs

• Install section now leads with explicit venv setup (Linux/Windows commands)
  and links PEP 668. Previously the venv step was implicit, which tripped up
  users on modern Debian/Ubuntu where pip refuses to write into the system
  Python without a venv.

v0.1.1

Added

• decepticons.__version__ derived from package metadata.
• scripts/release.sh — one-command version bump + tag + push.
• Release workflow now creates a GitHub Release with notes from CHANGELOG.md
  and attaches the wheel + sdist as release assets.
• Release workflow smoke-installs the built wheel and verifies the installed
  __version__ matches the pushed tag before publishing.

Changed

• Aligned canonical phrasing across README, docs, site, and package metadata:
  unified five-pillar list (substrates, memory, gating, routing, readouts),
  unified ecosystem labels (kernel · runtime · forensics), unified promotion
  rule wording, unified causality-test pitch.
• pyproject.toml description now matches the README/site tagline voice.
• decepticons.__init__ module docstring rewritten to match the canonical
  one-line description.
• decepticons --help description now describes what the CLI does instead of
  the generic "Decepticons CLI" placeholder.

Packaging

• README install section now leads with pip install decepticons for end users;
  source-install instructions moved under a "for development" heading.
• README image and inter-doc links converted to absolute GitHub URLs so the
  README renders correctly on PyPI (relative paths break outside GitHub).
• Added a PyPI version badge to the README header.
• Added MANIFEST.in so the sdist ships CHANGELOG.md, CONTRIBUTING.md,
  and .github/workflows/ for downstream packagers.

v0.1.0

First public release. Alpha-stage research kernel.

Kernel

• Reusable substrate primitives: EchoStateSubstrate, DelayLineSubstrate,
  LinearMemorySubstrate, OscillatoryMemorySubstrate, MixedMemorySubstrate,
  HierarchicalSubstrate, plus the factories config-driven dispatch.
• Memory primitives: ExactContextMemory, NgramMemory,
  StatisticalBackoffMemory, OnlineCausalMemory, plus ExactContextCache
  and StatisticalBackoffCache view layers.
• Control surfaces: ControllerSummary, PredictiveController,
  LearnedSegmenter, PathwayGateController, SummaryRouter,
  HormoneModulator, PredictiveSurpriseController.
• Feature views: ByteLatentFeatureView, HierarchicalFeatureView,
  LinearMemoryFeatureView, SampledMultiscaleReadout,
  ProbabilityDiagnostics, bridge_feature_arrays,
  BidirectionalContextProbe.
• Readouts and experts: RidgeReadout, FrozenReadoutExpert.
• Adapters: CausalPredictiveAdapter, NoncausalReconstructiveAdapter,
  OracleAnalysisAdapter, BridgeExportAdapter, TeacherExportAdapter,
  ByteLatentPredictiveCoder.
• Causal-bank family: CausalBankConfig, build_linear_bank,
  validate_config, scale_config, learnable_substrate_keys. Substrate
  modes frozen, learnable_decays, learnable_mixing, learned_recurrence,
  gated_retention. Optional memory attachment, stacked blocks, selective
  scan, banded readouts, byte-to-patch encoding, fast/slow hemispheres,
  polynomial expansion, training noise, adaptive regularization.
• Runtime: CausalTrace, FitReport, evaluate_rollout, score_next_step,
  evaluate_dataset, evaluate_rollout_curve, evaluate_transfer_probe,
  TrainModeConfig, ArtifactAccounting, artifact audits.

Backends

• decepticons.models.causal_bank_torch.CausalBankModel — PyTorch
  implementation with frozen substrate, selective scan augment, banded readout.
  Optional installs: pip install -e ".[torch]".
• decepticons.models.causal_bank_mlx.CausalBankModel — MLX equivalent.
  Optional installs: pip install -e ".[metal]".

CLI

• decepticons fit — single-command fit + sample over a UTF-8 corpus.

Verification

• tests/test_causality.py verifies every substrate mode is causal under
  perturbation. CI fails if any future-leak is detected.
• tests/test_dependency_firewall.py AST-scans the kernel to ensure
  decepticons never imports its descendants.