Skip to content

nanoIM v0.1.5

Choose a tag to compare

View all tags
@jlov7 jlov7 released this 01 Jun 14:06
v0.1.5
4645849

nanoIM v0.1.5

First public release after the Codex adversarial-review hardening pass.

Headline

Seven modern open-weights LLMs (Phi-4 14B, Llama 3.2 3B, Qwen3 4B / 14B MLX / 30B-A3B MLX, Gemma 4 26B, GLM-4.7 Flash) were given only the flattened transcript and asked for the next action. 420 alias pairs, 420/420 with paired separation = 0.000. Zero parse failures across 840 LLM calls. Every stream-aware method (rules, lookup table, GRU, Transformer) clears the same suite at 1.000 / 1.000.

The bottleneck is the representation, not the model.

What's verified

  • 253 / 253 tests pass
  • pyright clean on the nanoim/ package (0 errors, 0 warnings)
  • verify_release: pass
  • security_audit (with --include-history --expected-public-tag v0.1.5): pass. Zero forbidden paths (.codex, .claude, .cursor, .specstory, .env*, drafts) anywhere in reachable git history. Only v0.1.5 exists in the v* tag namespace.
  • package_audit: pass. Wheel install smoke uses uv venv --seed + the venv's own pip install --no-deps (uv 0.11 deprecated uv pip install).
  • hf_validate: pass. Local HF bundle contract check on the offline export.
  • Reproducibility manifest hashes 155 canonical artifacts with zero missing.
  • LLM baseline parser was fixed during the review pass: it now falls back to the reasoning field (Ollama and GLM emit it separately) and excludes parse failures from metrics rather than silently defaulting to WAIT.

What this does not claim

  • Not natural audio, video, or speech understanding.
  • Not a chatbot, ASR, TTS, or production assistant.
  • Not a frontier-model-parity claim. The point is a small from-scratch existence proof.
  • Not a generalization claim for the learned model across families — the LOFO numbers in reports/lofo_sweep.json say so explicitly.
  • Not yet validated on consented natural human-event data.

Reproduce

git clone https://github.com/jlov7/nanoIM.git && cd nanoIM
uv sync --dev
uv run pytest
uv run python -m nanoim.verify_release --out /tmp/verify.json
uv run python -m nanoim.security_audit --include-history --expected-public-tag v0.1.5 --out /tmp/security.json

Attached assets

  • nanoim-0.1.5-py3-none-any.whl, nanoim-0.1.5.tar.gz — the Python distributions
  • nanoim-0.1.5-release-bundle.tar.gz — the full release bundle with SBOM and SHA256SUMS
  • reproducibility.json — SHA-256 over every canonical artifact
  • release_verification.json — the scientific gate result
  • security_audit.json — the public-history audit
  • package_audit.json — the wheel install smoke result
  • release_bundle_manifest.json — the bundle manifest
  • huggingface_offline_validation.json — the HF bundle contract check
  • llm_baseline_summary.json — aggregate of the seven LLM transcript-only sweeps

Hardware used for the recorded run: Apple M4 Max MacBook Pro, 48 GB unified memory. Python 3.12.10, PyTorch 2.12.0.

Assets 12
Loading