Akashaakasha-model

Authorize, abstain, or block — in one pass.

A small choice model for tool gates and typed menus. Score a proposed call before your host runs it; Path A needs no PyTorch.

Visual usage demo: tool-gate state, typed signals, and ready probability bars
Local visual demo — ready / abstain / blocked with live probability bars

Tool gate first

For Akasha OS the product surface is a tool gate: your System 2 proposes a call; Akasha returns ready, abstain, or blocked. The library never executes side effects.

Path A is deterministic policy over typed signals, catalog schema, and planner thresholds. Natural-language injection in arguments cannot grant capabilities or force ready.

ready → host may run abstain → ask / reformulate blocked → do not execute
Path AGate / host / outcomes — no torch
Path BOptional scorers & offline recipes
Path CTraining & typed-decision loops
Path DVision / games lab (shared option table)

Choice · Score · Noul

One forward pass over a changing list of typed questions. Each question returns its own probability distribution — not a token stream.

  • Choice

    Pick among discrete options (routes, placements, tool names). Softmax over the menu.

  • Score

    Rate or rank continuous-feeling signals when the host needs a graded preference.

  • Noul

    Two-way false / true for yes–no gates and confirmations.

Demos you run locally

The static site stays light. Clone the repo, install Path A, and open a browser page generated on your machine — screenshots below are author-owned captures of those demos.

Visual demo showing a blocked delete-without-confirm decision

Usage visual demo

Steps through state → typed signals → probability bars → ready / abstain / blocked rationale (tool gate, then ticket-router abstain).

uv venv && source .venv/bin/activate
uv pip install -e '.[dev]'
python examples/usage/visual_demo.py --open

Install

Stable pin for OS consumers. Path A installs without torch; add [torch] for scorers, training, and vision.

pip install akasha-model==0.3.0

Examples, checkpoints, and this site’s demos live in the GitHub checkout — the PyPI wheel ships the package and CLIs only.

Go deeper

Integrator docs and use cases beyond the gate.