BarunAction-35M candidate-v2 — usable research release

BarunAction-35M candidate-v2

BarunAction-35M is a usable public research release of a 35,072,768-parameter, local,
proposal-only personal-action compiler. It maps requests plus supplied context and tool schemas to
typed Action IR, or to ABSTAIN, CLARIFY, or CONFIRM. The runtime never contacts people,
calendars, devices, or external tools; execution and confirmation remain the caller's
responsibility.

Quick start

uv venv --python 3.11
uv pip install https://github.com/harrrshall/barunlm-35m/releases/download/barunaction-v1.1.0/barunlm-1.1.0-py3-none-any.whl
barunaction demo

The built-in demo is deterministic and weight-free. It loads no checkpoint, requires no network,
and applies the proposed action only to an in-memory simulator where the side effect remains
blocked.

Download the float or Darwin ARM64 int8 checkpoint from the immutable public W&B v0 artifacts
using the pinned retrieval instructions.

Audited result

Model	Parameters	Strict exact match
BarunAction-35M candidate-v2	35,072,768	602/756 (79.63%)
Qwen2.5-0.5B-Instruct, matched one-epoch adaptation	494,032,768	663/756 (87.70%)

BarunAction is 14.09 times smaller and retains 90.80% of Qwen's exact-match rate, but it
trails by 61 rows or 8.07 percentage points. This is useful parameter-efficiency evidence, not
a larger-model breakthrough. The 961-row official Mobile evaluation split was never read or
materialized.

Included

• barunlm 1.1.0 wheel and Python sdist;
• barunaction Python inference API and CLI;
• deterministic parsing, schema, policy, checkpoint, and quantization checks;
• weight-free, in-memory-only barunaction demo;
• pinned float and Darwin ARM64 int8 checkpoint retrieval;
• tagged GitHub source containing training/evaluation code, configs, model/data cards, benchmark
  evidence, append-only experiment ledger, and mistakes log.

The GitHub tag/source archive is the complete research and process bundle. The attached
barunlm-1.1.0.tar.gz is a Python package sdist and does not contain every repository document or
experiment record.

Verification

• immutable annotated tag targets commit ea26b83eba1a1819ab0088c611dad977265669f7;
• 1,499 repository tests passed; Ruff, formatting, JSON, provenance, staged-diff, and secret scans
  passed;
• wheel and sdist passed twine check;
• the exact attached wheel installed and passed the network-denied safe demo on CPython 3.10.20
  and 3.11.15;
• the wheel reproduced byte-for-byte across two independent clean clones;
• SHA256SUMS authenticates both attached package files.

The authoritative W&B v0 evaluation evidence contains non-secret workstation paths and protected
resource identifiers. A proposed redacted v1 was stopped before upload because W&B version
assignment could not satisfy the frozen atomic publication contract; no replacement artifact was
published. This limitation is documented rather than hidden.

The current next hypothesis, Generate-Correct SFT v1, has CPU-only replay and bounded-forge
evidence. It authorizes no full synthetic population, model load, CUDA, training, human-label
access, or JarvisLabs resource.