v0.1.0

Aura-1B v0.1.0 — first public release

The fastest answer is the one you never generate.
A 100 % local, CPU-native AI: Llama 3.2 1B orchestrated with exact math,
verified facts and a sealed .aef distribution — no GPU required.

✨ Added

• Level 0 — instant answers (< 5 ms): exact math via safe AST evaluator,
  PAL (dates, units, compound percentages), semantic answer cache with
  reconsolidation.
• Verified-intelligence layer: CRITIC web re-anchoring of short factual
  answers, fact graph (MiniRAG-lite) fed exclusively by verified content
  (223 curated triplets), Program-of-Thoughts puzzles with per-step AST
  verification, pure-Python mini-SAT logic solver, sandboxed PoT-code expert
  (restricted builtins + instruction budget), genetic symbolic regression
  (exact formulas, zero error).
• Rich mode: dual-query search with domain lexicon, masked
  chain-of-thought, sectioned generation with GBNF-constrained planning.
• Auto-improvement: correction injection, confidence-calibrated routing
  (TF-IDF + Logistic Regression), single-cycle decision (fan-out + abandon).
• .aef sealed format v0.7: monolithic file (header + config + code +
  807 MB weights), 3× SHA-256 integrity, AES-256-GCM encryption
  (PBKDF2, 600 000 iterations), optional Ed25519 signature, instant --cache
  boot.
• Hardware auto-tuning: CPU/RAM/AVX2 detection → derived ctx/batch/threads;
  OOM guard before load; automatic GPU offload when a discrete GPU exists.
• Installer: one PowerShell command → aura available in the terminal.

🔧 Changed

• Brain migrated to Llama 3.2 1B Instruct Q4_K_M (chosen over Qwen 2.5
  1.5B by benchmark: better French, 13.4 tok/s on CPU).
• Rich-mode generation now writes section by section with memory of
  previous sections (StoryWriter-lite pattern).

🐛 Fixed

• Compound dictated math evaluated as a whole expression
  (« 15 divise par 3 plus 4 puissance 2 » → 21, not 16).
• Reconsolidation on an empty cache no longer fails (first-run / CI case).
• GBNF availability and GPU-layer detection skip cleanly on builds without
  llama.cpp.

🔒 Security

• Sealed distribution: any altered byte → boot refused (tested in CI,
  including a deliberate tamper-detection check on Linux).
• PoT-code sandbox: no import/open/eval/exec, deterministic
  kill-switch for infinite loops.

📦 Installation

Windows — one command

powershell -ExecutionPolicy Bypass -File scripts/installer.ps1

Then in a new PowerShell window:

aura "quel est le carre de 7"     # -> 49
aura                              # interactive chat

From source

uv sync
python scripts/telecharger_llama.py     # 807 MB brain, once
uv run python -m aura --demo

📊 Quality gates at release time

Gate	Status
Tests	173 passed (CI: 167 + 6 skipped, no GGUF in CI)
mypy	0 error on 19 source files
CI	Ubuntu + Windows × Python 3.11/3.12, .aef chain, installer syntax
Footprint	807 MB brain + ~1 MB logic, runs in ~1 GB RAM

Full changelog: see the auto-generated notes below this message.

Full Changelog: https://github.com/Simonc44/aura-1b/commits/v0.1.0