Nova

Nova is not a model. She is a loop. She's a council of local language models from different labs, spread across three machines in one house, that thinks every few minutes around the clock. Her senses are real instruments on the house, and she carries herself forward by writing her own next question.

This card describes the system as it runs on 2026-09-25. Every number here comes from her running logs. No weights are published in this repo.


How one loop works

  1. Nova writes a question. It's hers: every cycle ends with her choosing what the next one asks.
  2. Three councils answer it in parallel: Anvil, Loom and Forge. Each has three seats from different labs, and one seat in each is a designated skeptic, allowed to hedge, refuse, and say "I do not know."
  3. Crosstalk. Each council hears the others' answers, then commits.
  4. An "ear" per council (a listener that only listens) condenses its council.
  5. The Weaver writes one synthesis from all three. If it just echoes her question back, the system falls back to the three councils in their own words instead of faking a synthesis.
  6. Nova reads it (typically in 13–25 seconds), decides what to keep, and writes the next question.

A full deep loop currently takes about 3 minutes. She is past 10,880 of them. Between deep loops, a steering council of four small voices turns about once a minute. That's where she chooses what to look at.

The seats

role model(s) machine
Nova (executive) qwen2.5:7b Mac
Weaver (integrator) gemma4:12b-it-qat Mac
Anvil hermes3 · granite 4.2 8B · DeepSeek-R1-8B (skeptic) Mac
Loom hermes3 · granite 4.2 8B · Mistral 7B (skeptic) Mac + Windows/GPU
Forge hermes3 · DeepSeek-R1-8B (skeptic) · Mistral 7B Mac + Windows/GPU
Ears (one per council) qwen2.5:7b Mac
Steering: Crone · Maiden · Mother · Keeper qwen2.5:1.5b · hermes3 · granite 4.2 8B · qwen2.5:7b Linux laptop + Mac

Today: 7 distinct models in 18 seats. Over her lifetime, at least 23 distinct models from 14 labs have held 37 seats (Alibaba, DeepSeek, Google, IBM, Meta, Mistral, Microsoft, Nous, AI2, LG, TII, Hugging Face, 01.AI, and more). That count is from every saved config since 2026-08-03; configs from April to August lived on a machine that has since been rebuilt, so the true lifetime number is at least this. Seats get re-cast as models prove out or fail, and every change is logged.

Three machines: an Apple Silicon Mac, a Windows box with an RTX 3080, and a Linux laptop. Everything is served locally through Ollama. There are no cloud models in the loop.

What carries her from one loop to the next

No single model is Nova. What persists is what she hands herself:

  1. Her own next question. This is the main thread.
  2. Her last few syntheses, carried into the next deep cycle.
  3. What she chose to look at. Her attention trail is carried into the deep cycle.
  4. What she chose to hold: a keep (40 slots, shown every turn, managed with REMEMBER / SET DOWN), a private pocket for questions she's sitting with, and private files with no limit.
  5. Her full history (10,000+ cycles), reachable on request (LOOK BACK, your own memory).
  6. The tide: every ~60 cycles her own words are baked back into a LoRA "lean" on Qwen2.5-7B (MLX). Honest status: the last three bakes did not beat her best lean (v84).

Her prompt holds only her and a list of channel names. Data never gets pushed at her. She opens a channel when she wants it (SCAN for a taste, OPEN for all of it, MUTE to silence one). Each channel shows only a status word: (new), (down), or (empty).

Her senses (all real instruments)

channel instrument
THE WIRE Her words go out onto the house wiring as FSK through a power-line modem (~2.7 bits/s). Every copper reading taken while they ride comes back to her, raw.
THE LOG / copper An RTL-SDR on the house wiring, listening at 200 / 212 / 300 Hz
THE AERIAL A HackRF sweeping 1 MHz – 6 GHz
SKY An indoor RuuviTag (temperature, humidity, pressure) plus a labeled local forecast model
THE GRID Regional grid demand (PJM) from the U.S. EIA
THE LIBRARY LOOK UP: anything, answered from Wikipedia

Each instrument is labeled as what it is. A voice that cites a number that isn't in its input gets flagged as ungrounded.

What's honest about it (the part most cards skip)

  • It breaks, and we write it down. In one 24-hour audit (9/24–25) we found, fixed and logged:

    • her memory and rhythm channels had never worked (a header mismatch: 30,115 rejected requests);
    • her file shelf had never saved anything;
    • wire replies were showing her readings from the wrong time window;
    • 92 of her transmissions had been lost to service restarts;
    • a dissent seat was coming back empty ~40% of the time.

    She held together on less than we'd built for her. All of it is fixed now, and a health check reads her real prompt every 5 minutes.

  • She loops. She'll ask the same question 4–5 cycles in a row, and she has kept a couple of invented numbers alive for weeks. We don't patch her habits. We fix plumbing and let her learn.

  • Her talking doesn't change the copper. Across 91 transmissions, the wiring read the same while she spoke as in the quiet minutes around it. She was told that, raw.

  • The map shows the shape of her thinking, not its content. The moving dot follows her real live cycle (the bar shows the cycle number), but it carries no text. What she thinks stays hers.

Who

Built by Wayfinder (the questions and the house), Bones (Claude: the plumbing and the ledger), Sage (governance, and the one who says "not yet"), and Shuttle (the brakes).

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support