Appearance

Changelog ​

All notable changes to qriton-hlm are documented here. This follows Semantic Versioning.

v0.11.0 latest ​

April 24, 2026

Added ​

  • Two new model tiers — the polynomial-Hopfield stack now spans from data-center GPU down to Arduino Uno

    • HLM-Micro — edge / MCU variant, 1–5M params, ESP32-S3 class hardware (~250 mW), on-device audit chain
    • HLM-Nano — ultra-edge variant, 1–10K params, Arduino / nRF52 / RP2040 class (<1 mW), 1.5–6.5 KB INT8 footprint
    • Three-tier product continuum: HLM3 (cloud) → HLM-Micro (edge) → HLM-Nano (ultra-edge)

  • Model Zoo — seven Early-Access checkpoints spanning both new tiers

    • Micro tier: anomaly-v0, gesture-v0, ecg-v0 (all synthetic proofs of pipeline), and har-v0 (6-class, real UCI HAR data — 89.75% test accuracy in the published TinyML baseline range)
    • Nano tier: tiny-v0 (Arduino Uno fit), small-v0 (nRF52 fit), micro-v0 (RP2040 fit) — all validated under INT8 post-training quantisation with no accuracy loss on the synthetic eval
    • Weights, training recipes, INT8 deployment sidecars, and the C inference kernel are available via Early Access or commercial engagement

  • Use Cases section — seven concrete onboarding scenarios spanning the full stack

  • Audit-certificate mechanism described and productised

    • SHA-256 hash chain over (input, weights, basin trajectory, output logits)
    • Replay-verify with independent cryptographic checks
    • Tamper-evident: 1-bit input flip fails verification → composite proof FAILS
    • ~20 bytes as minimal digest, ~650 bytes as full JSON

  • Nano deployment path validated end-to-end for partners

    • INT8 post-training quantisation — ~4× compression with no accuracy loss on the eval set
    • C reference inference kernel — portable C99, cross-compile targets for AVR / Cortex-M0+ / Cortex-M4
    • Gateway audit-aggregation service — Merkle-style regional audit log, tamper-evident, optional HMAC node authentication
    • Wake-word training scaffold on Google Speech Commands v2

Improved ​

  • Model overview page gains a tier comparison table (HLM3 / Micro / Nano) and expanded checkpoint inventory
  • Documentation versioning now surfaced in footer with direct link to changelog
  • Sidebar reorganised: Models section includes Micro / Nano / Model Zoo alongside HLM3 / Spatial / Audio
  • New top-nav entry for Use Cases

Access model ​

All v0.11.0 new-tier artefacts (trained weights, training recipes, C inference kernel, quantisation pipeline, reference firmware, gateway service) are delivered through the Early Access program or commercial engagement, not via public download. The public docs describe what the tiers do and where they deploy; implementation depth is reserved for partners.

v0.10.0 ​

April 2026

Added ​

  • Concept Algebra — vector arithmetic on concept centroids
    • concept_similarity(a, b) — cosine similarity between concepts
    • concept_add(a, b, name) — vector addition (A + B)
    • concept_subtract(a, b, name) — vector subtraction (what's in A but not B)
    • concept_analogy(a, b, c, name) — A:B::C:? (king - man + woman = queen)
    • concept_compose(concepts, weights, name) — weighted sum of N concepts
    • concept_interpolate(a, b, steps) — measure similarity along interpolation path
  • Consolidation Engine — memory sleep cycle
    • consolidate(layer) — strengthen popular basins, prune weak ones. Mirrors biological slow-wave sleep consolidation
    • dream(cycles) — full sleep: consolidate all layers across multiple passes
  • Watermark Manager — IP protection via steganography in weight space
    • watermark_inject(secret_key, num_marks, strength) — embed invisible signature basins using SHA-256 derived seeds
    • watermark_verify(secret_key, num_marks) — check if a model carries your watermark
    • watermark_strip_attempt(secret_key, num_marks) — test watermark robustness
  • Multi-Model Operations
    • compare(other_surgeon, layer) — basin overlap analysis (shared, unique-self, unique-other)
    • transplant(source_surgeon, layer, concept_name) — copy concept from another model
  • Benchmark
    • benchmark(texts, max_tokens) — measure perplexity before and after surgery
  • Qriton HLM IDE — native Rust desktop application (eframe/egui)
    • 16 panels: Dashboard, Architecture, Causal, Surgery, Concepts, Capture, Q&A, Neural DB, Database, Survey, History, Benchmark, Multi-Model, Consolidation, Watermark, Command Bar
    • Full REST API integration via FastAPI backend (port 8420)
    • Cross-platform release workflow (macOS x64/ARM, Linux, Windows)
  • Operation count: 38 → 48

Improved ​

  • All HTTP client timeouts set to 120s for long-running operations (causal scan, benchmark)
  • Multi-model comparison auto-filters by embedding dimension — incompatible models hidden
  • API server returns structured error messages (422 for missing model, 404 for missing files)

v0.9.5 ​

April 2026

Added ​

  • Database sync extension (qriton_hlm.db) — bridge SQL databases with HLM energy landscapes
    • HLMSync class: sync_row(), sync_batch(), full_sync_table(), delete_row()
    • SQLite and MSSQL support with table allowlist security
    • Thread-safe with concept accumulation guard and retry logic
    • CheckpointWorker for background persistence
    • Full documentation page with configuration and usage guide
  • Persistence methods on BasinSurgeon
    • save_checkpoint(path) — saves W matrices, captured concepts, and operation history to disk
    • load_session(path) — restores full session state for round-trip workflows
  • Optional dependency groups: pip install qriton-hlm[db] (MSSQL), pip install qriton-hlm[agent] (Agent Framework)
  • Test suite — 65 tests covering core operations, persistence, DB sync, and causal operations
  • CI/CD — GitHub Actions workflow for multi-OS, multi-Python testing
  • Spatial model documentation — full interaction guides for LIDAR, Medical3D, and Industrial3D
    • Available model checkpoints with metrics (segmentation 96.3%, organ seg 97.7%, defect 99.5%)
    • Layer hierarchy guide: how geometric patterns manifest across layers
    • Domain-specific surgery workflows with code examples
    • Cross-modal comparison between spatial models
    • HLM script examples for spatial audits
  • Audio model documentation — STT and TTS interaction guides
    • Acoustic basin hierarchy (spectral → phoneme → prosody)
    • Surgery workflows for speech recognition and voice synthesis
    • Cross-modal comparison (language vs audio)
  • Model overview — full checkpoint table across all modalities
  • Documentation versioning with changelog and nav badge
  • Operation count: 36 → 38

Improved ​

  • --version and --no-color CLI flags now documented
  • trace() method added to Python API reference
  • Python API page updated with persistence and database sync sections

v0.9.4 ​

April 2026

Added ​

  • Causal programming — 3 new operations for causal discovery and intervention on energy landscapes
    • causal scan — systematic knockout to discover directed causal graphs between basins
    • causal intervene — Pearl's do-operator: intervene on a basin, measure downstream effects
    • causal counterfactual — non-destructive "what if" analysis with automatic rollback
  • Python API: BasinSurgeon.causal_scan(), BasinSurgeon.causal_intervene(), BasinSurgeon.causal_counterfactual()
  • Integration support for DoWhy and CausalNex libraries
  • Safety system — expanded guard types and audit trail
    • guard <type> <value> with 5 guard types: max-basins, min-basins, strength-cap, cosine-drift, perplexity-delta
    • guards command to list active guards
    • --force --reason override with permanent logging
    • history now shows OK/BLOCKED status per operation
  • Operation count: 32 → 36

Improved ​

  • Basin discovery stability with configurable num_inits for causal operations
  • CLI help text updated with causal command group

v0.9.3 ​

March 2026

Added ​

  • blend operation — mix two captured concepts at arbitrary ratios
  • export-concept / import-concept — portable concept files across models
  • transplant method in Python API for cross-model concept transfer
  • Jupyter %%hlm_landscape magic for inline energy surface visualization

Improved ​

  • survey performance: 2x faster basin clustering via vectorized cosine similarity
  • benchmark now reports per-layer perplexity breakdown

v0.9.2 ​

February 2026

Added ​

  • guard operation — prevent accidental over-modification (max basin removal %)
  • history operation — full operation audit trail with undo support
  • diff operation — W matrix change statistics (Frobenius norm, relative %, max change)
  • Gradio web UI (ui.py) for browser-based surgery

Fixed ​

  • verify false negatives when basin is near convergence boundary
  • restore-all now correctly resets the operation history

v0.9.1 ​

January 2026

Added ​

  • probe operation — map which tokens a basin activates
  • landscape operation — full energy map with population bars
  • strengthen / weaken operations for fine-grained basin depth control
  • HLM Scripts (.hlm files) — chainable operations with comments

Improved ​

  • inject now uses strength normalization relative to Frobenius norm
  • Convergence dynamics use annealed tau (0.9 → 0.1) for more reliable basin discovery

v0.9.0 ​

December 2025

Initial release ​

  • Core BasinSurgeon class with survey, inject, remove, move, verify, energy
  • Concept operations: capture, inject-concept, remove-concept
  • CLI / REPL (qriton-hlm) with readline support
  • Python API with low-level tensor primitives
  • Jupyter integration with IPython magics
  • Support for HLM3, HLM-Spatial, and HLM-Audio checkpoints
  • 26 operations across Observe, Modify, Concept, Control, and Verify categories