CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file. Format follows Keep a Changelog. This project adheres to Semantic Versioning.

[Unreleased]

Added

• Membrane Immunity Gate exploration — new exploration document (whitePaper/explorations/MEMBRANE_IMMUNITY_GATE.md) proposing AKD1500 as a hardware-level HTTP traffic classifier for the ecoPrimals membrane defense. BingoCube evolutionary reservoirs classify fleet vs human traffic at ~54 µs per request, feeding into MazeCube's pressure dimension. First neuromorphic IDS architecture with on-chip adaptive evolution (Nautilus shell). Connects to biomeGate brain architecture (hotSpring) and toadStool deiisle engine. Reactivation of rustChip as the NPU substrate for membrane defense alongside its existing physics workload role.

• glowplug absorption — VFIO device lifecycle management (bind, warm boot, tear down) absorbed from coralReef's ember/glowplug. rustChip now manages its own hardware lifecycle without external orchestrator or kernel module. Code is derivative of the ecoPrimals ecosystem and retains the full scyBorg license under the lineage principle.

• HW/SW backend separation — VfioBackend [HW] and SoftwareBackend [SW] are explicitly labeled and never conflated. BackendSelection enum and select_backend() function provide the composition entry point.

• Narrative explorations — four new whitePaper documents grounding NPU science in hardware evidence:

  • WHY_NPU.md — the foundational neuromorphic argument
  • SPRINGS_ON_SILICON.md — 5 NPU patterns × 3 science domains
  • NPU_FRONTIERS.md — 10 creative frontiers for neuromorphic hardware
  • NPU_ON_GPU_DIE.md — NPU as a GPU functional unit (area/power analysis)

• 5 standalone science demos — self-contained binaries reproducing peer-reviewed NPU science claims without external data:

  • science_lattice_esn — Hybrid ESN for Lattice QCD steering
  • science_bloom_sentinel — Streaming Sentinel for harmful algal bloom detection
  • science_spectral_triage — Microsecond Gatekeeper for LC-MS spectral triage
  • science_crop_classifier — Online Adaptation via (1+1)-ES for seasonal crop stress
  • science_precision_ladder — Precision Discipline: f64 → f32 → int8 → int4

• warm boot binary — warm_boot demonstrates sovereign device lifecycle using the absorbed glowplug module.

• Experiment 006 — BAR0 Register Probe: true layout discovery on cold-boot VFIO-bound AKD1000 (80 NPs, 10 MB SRAM confirmed via pure userspace probing).

Changed

• scyBorg licensing update — AGPL-3.0-or-later (code) + ORC (game mechanics)
  • CC-BY-SA 4.0 (creative/docs). Lineage principle: code absorbed from ecoPrimals retains full scyBorg license even within the BrainChip exception boundary. Akida-specific code exempt; systems endowed from ecoPrimals are not.
• Test count: 353 → 367 (new science demos, glowplug module).
• Zoo model count corrected: 29 → 28 (actual ZooModel enum variants).

[0.1.0] — 2026-04-30

First versioned release. 353 tests passing, 29-model zoo, pure Rust conversion pipeline, guideStone validation, CI, full documentation suite.

Added

• Pure Rust model pipeline — complete replacement for Python SDK dependency. Models can now be created, quantized, serialized, and parsed entirely in Rust:
  • akida-models::quantize — per-layer and per-channel int1/2/4/8 symmetric quantization with nibble packing and round-trip verification (14 tests). Algorithm patterns absorbed from neuralSpring and hotSpring.
  • akida-models::import — weight import from .npy, .safetensors, and raw f32 slices with hand-rolled .npy parser and f16/bf16 upcast (16 tests). safetensors pattern absorbed from neuralSpring.
  • akida-models::schema — reverse-engineered FlatBuffer serializer matching the structure observed across all 25 .fbz artifacts. Produces valid Snappy-compressed FlatBuffers that round-trip through the parser (5 tests).
  • akida-models::schema_parser — schema-aware FlatBuffer parser using vtable navigation instead of heuristic byte scanning. Extracts version, layer names, and properties via proper field offsets (6 tests).
  • akida convert CLI command — end-to-end conversion: import weights from .npy/.safetensors/zeros:N/random:N, quantize, serialize via FlatBuffer, compress, write .fbz, verify round-trip. No Python.
  • .fbs schema at crates/akida-models/schemas/akida_model.fbs documents the reverse-engineered format.
• New dependencies: flatbuffers = "25", safetensors = "0.5" (both pure Rust, no C/Python/vendor SDK).
• guideStone computational artifact — akida guidestone [dir] runs a self-leveling validation: parses 25 models (21 BrainChip zoo + 4 physics), computes SHA-256 digests, validates structure (version, layers, decompression ratio, NP budget, weight blocks, file size), benchmarks parse throughput (15.8 MB/s), and emits a graded report (225 checks, 0 failures). Implemented in crates/akida-models/src/guidestone.rs. Any subsequent work on this build can reference the guideStone run as the anchored baseline.
• guideStone documentation — specs/GUIDESTONE.md maps the ecoPrimals guideStone verification class (5 properties) to NPU compute. baseCamp/GUIDESTONE_CERTIFICATION.md tracks certification progress.
• baseCamp documentation suite — ZOO_GUIDE.md (comprehensive model zoo guide with all 25 models, Rust ecosystem integration, and what it unlocks), SCIENTIFIC_DEPLOYMENT.md (whitepaper-style NPU deployment guide for scientific/data workloads with production results and deployment patterns), RUST_NPU_ECOSYSTEM.md (what the Rust NPU ecosystem enables and how rustChip fits).
• Full BrainChip model zoo export — scripts/export_zoo.py exports all 21 pretrained models from the Akida SDK (v2.19.1) to .fbz with a ground-truth zoo_manifest.json. Python SDK is a one-time validation oracle, not a runtime dependency.
• Physics model export — scripts/export_physics.py generates .fbz files for the 4 ecoPrimals physics models (ESN readout, phase classifier, transport predictor, Anderson classifier) with random weights matching documented architectures.
• Zoo regression test suite — crates/akida-models/tests/zoo_regression.rs validates that all 25 .fbz files (21 zoo + 4 physics) parse successfully. Run with cargo test --test zoo_regression -- --ignored.
• CLI model commands — akida parse <file.fbz> inspects model structure; akida zoo-status [dir] shows cache status against the full ZooModel enum.
• Spring profiles catalog — baseCamp/spring-profiles/README.md documents every NPU workload across ecoPrimals springs with reproducible instructions.
• Model::input_shape() / output_shape() — shape accessors on the parsed model, populated via set_shapes() from manifest metadata.
• ZooModel enum expanded to 28 variants — covers all 21 BrainChip MetaTF models, 2 NeuroBench benchmarks, 4 ecoPrimals physics models, and 1 hand-built test. Filenames match actual export output.
• ZooModel::brainchip_zoo() — convenience accessor for the 21 MetaTF models.
• ModelTask::FaceAnalysis, Segmentation, EyeTracking — new task categories for the expanded model catalog.

Changed

• ModelZoo::load_metadata now parses files via Model::from_bytes() to determine validity, version, and layer count. The old magic-byte check ([0x80, 0x44, 0x04, 0x10]) rejected all real zoo .fbz files.
• ValidationTier::Functional assigned to all 21 MetaTF zoo models (exported + parsed successfully).

Fixed

• ModelZoo::load_metadata no longer marks real .fbz files as invalid.
• baseCamp/conversion/README.md no longer references non-existent quantize.rs.
• baseCamp/zoos/brainchip_metatf.md updated with actual export results.

Added (0.1.0 release hardening)

• QUICKSTART.md — 5-command onboarding: clone, build, convert, parse, test.
• LEVERAGE.md — standalone tool leverage guide following wateringHole pattern.
• Nature Preserve — baseCamp/preserve/ with 7 domain application patterns (physics, biology, audio, vision, environmental, genomic, industrial).
• 4 Rust-native zoo models — ESN multi-head, ESN 3-head transport, streaming sensor 12ch, adaptive sentinel. Generated via pure Rust akida convert.
• NPU_LEVERAGE.md added to hotSpring, wetSpring, neuralSpring.
• sporePrint science page 27 — Nature Preserve: Applied NPU Science.
• wateringHole — rustChip entry in LEVERAGE_GUIDES.md (v1.1.0), NPU onboarding section in GARDEN_COMPOSITION_ONRAMP.md.
• CONTRIBUTING.md — quality commands, layout, workflow.
• CITATION.cff — machine-readable citation metadata.
• CONTEXT.md — AI/indexer context following PUBLIC_SURFACE_STANDARD.
• CI workflows — .github/workflows/ci.yml with fmt, clippy, test, doc, deny.
• TOLERANCE_REGISTRY.md — named numerical tolerances across all subsystems.
• justfile — common workflow commands (just check, just test, just guidestone).
• .cursor/rules/rustchip.md — project-level AI guidance.
• 16 new tests — 8 in akida-cli (arch parser), 8 in akida-bench (PRNG, timer, probe).
• README header reformatted to match SPRING_PRIMAL_PRESENTATION_STANDARD (Date, License, MSRV, Status with metrics).