Perci

What it is

Perci is described as an experimental cognitive OS for a single machine, incorporating a local Bitwork 200 MiB sparse pack, exact tools (e.g., math and geometry that compute), a typed world model, and a governance layer with append-only memory. It presents itself as a Rust-native, neuro-symbolic stack that routes input through reflex, Bitwork field, operators, and a SoftCascade thought arc, with human-authorization for weight promotion.

How it works

The README outlines a layered architecture where input passes through:

• reflex
• Bitwork field (α, residual, VSA, etc.)
• branches to exact tools, operators, or SoftCascade
• a thought arc for synthesis and governance The system includes append-only memory (Cortex), a governor with a local, typed world model, and exact tools (e.g., math and geometry). It emphasizes separating routing, memory, tools, and model to avoid opaque monolithic behavior. There are live mentions of margins and overlap_z for classification within the field, and a workflow where tickets, transfer, repair, and governance govern state changes.

Getting started

• Requirements include Windows, macOS, or Linux; Rust + Cargo; local pack file models/perci-cognitive-v0.3.pwgt (not in the clone).
• Steps for Clone & Launch (Windows):

  git clone https://github.com/jacksonjp0311-gif/Perci.git
  cd .\Perci
  # place PERCIW03 under models\  (or $env:PERCI_WEIGHTS = "...")
  Set-ExecutionPolicy -Scope Process Bypass -Force
  .\Launch-Perci.ps1
  

• Cargo commands to run various chat/utility tasks:

  cargo run --release -- chat
  cargo run --release -- ask "why does trust fail in distributed systems?"
  cargo run --release -- classify "invent a constrained metaphor for sparse cognition"
  cargo run --release -- fabric status
  cargo run --release -- fabric handoff "improve transfer on novel entities"
  python scripts/release_gates.py
  

• Additional protocol references: Cortex + fabric handoff details and AI evolve protocol, with commands to activate cortex and run handoffs.

Getting started (continued)

• Dark-blood CLI commands are available for cognition-related actions: /think, /concise, /deep, /balanced, /trace, /intel, /learning, /quit, etc., with explanations for each command. These are for inspection, planning, and auditing of the operator chain.

Recent releases

• The latest release entry in the provided data shows 0 releases, with the note "latest 0: - none".

Traction

• Stars: 29 (from the repository metadata). There is no separate 7-day or 1-day traction data provided.

Behind the repo

• The README references a local, binary, and model-centric architecture, including PERCIW03 pack format and a perci-cognitive-v0.3.pwgt file path outside the clone, suggesting a workflow that relies on local artifacts and a governance model for weight promotion.

Caveats

• License: none listed in the provided data.
• Age: created 2026-07-15; last_push 2026-07-15.
• Language: Python is listed as the language in the metadata, but much of the stack is described as Rust-native with Rust/BOM references; the README shows multiple languages in usage (PowerShell, Python scripts, cargo commands).