Aura

# Aura

DEMOS: [https://youtu.be/iTyxeugcZtI?si=B91No0Hjz3eKLMwz](https://www.youtube.com/@AuraDemos)

Aura is a local AI research project that runs entirely on your Mac. It explores
how to build an AI agent that can think continuously, remember things long-term,
govern its own behavior with auditable receipts, steer its own emotional tone
during text generation, and keep itself running reliably over time.

**Aura is not proof of consciousness, life, or personhood.** Nothing here settles
those questions. Some file and module names sound like they might — they're named
after the mechanisms they implement, not achievements they've proven.

The actual claim is narrower and testable: Aura's internal state measurably
affects what it generates, what it remembers, what tools it's allowed to use,
what tasks it picks up on its own, and how it repairs itself — through code
paths that leave audit trails you can inspect.

That's a smaller claim than the vocabulary suggests. It's also one you can verify.

[![License: All Rights Reserved (Read-Only)](https://img.shields.io/badge/License-All_Rights_Reserved_(Read--Only)-red.svg)](LICENSE)
![Python 3.12+](https://img.shields.io/badge/python-3.12+-blue.svg)
![Platform: macOS Apple Silicon](https://img.shields.io/badge/platform-macOS_Apple_Silicon-lightgrey.svg)

For the full technical deep dive, read [ARCHITECTURE.md](ARCHITECTURE.md). For
the same ideas explained without math, read [HOW_IT_WORKS.md](HOW_IT_WORKS.md).
For the standard of evidence we hold ourselves to on autonomy and novel output
claims, read [docs/BEHAVIORAL_PROOF_STANDARD.md](docs/BEHAVIORAL_PROOF_STANDARD.md).

**The main research program is
[docs/RECURSIVE_LATENT_CORTEX.md](docs/RECURSIVE_LATENT_CORTEX.md).** It asks:
can a frozen 32-billion-parameter model be made to reason deeper without changing
any of its stored weights? It builds the machinery to test that, runs a
pre-registered experiment, and reports that the hoped-for capability improvement
didn't appear. Summary in [Recursive Latent Cortex](#recursive-latent-cortex)
below.

**Evidence map:** Claims should point to runnable tests, proof bundles, receipts,
or replayable logs. Test counts change with the repo; use `pytest --collect-only`,
`make proof-bundle`, and [TESTING.md](TESTING.md) for current numbers rather
than trusting prose.

If you want to see it work, keep reading.

## Evidence boundary

This is a working AI research project. It is **not** proof of consciousness,
subjective experience, legal personhood, or moral standing.

The repo enforces that in code, not just in words. A boundary guard treats
loaded terms — "consciousness guarantee," "personhood proof" — as names for
test batteries rather than claims, unless independent evidence says otherwise.
A module named `qualia_synthesizer.py` is a name. Names are not evidence.

What is actually claimed, and what each claim costs:

- **Governance is a design goal, not a guaranteed fact.** Important actions
  must go through receipt-producing governance checks. A lint failure, a direct
  tool fallback, or a default-approved gateway is a bug — not acceptable
  behavior. A bug.
- **Autonomous self-improvement is not proven mature.** There's scaffolding
  here for autonomous repair, evaluation, patch tracking, and proof bundles.
  Scaffolding is not a result. Unverified self-modification stays outside the
  evidence scope until long real-world runs, hidden external tasks, and
  independent replication all succeed.
- **The cognitive layer hasn't been shown to earn its cost.** This is the
  biggest open gap in the project, and it belongs at the top of any honest
  reading. Two components have been measured against simpler alternatives
  under matched budgets — memory retrieval beats a basic context window when
  the needed fact falls outside the window (claim 31a), and the revision
  gate beats both "always keep the first answer" and "always keep the second"
  over the same two generations (claim 31b). Two subsystems on two task
  types is not the whole layer. Nothing here yet shows that the IIT metrics,
  qualia modules, neurochemical simulation, substrate differential equations,
  dream cycles, or theory arbitration make Aura *better at a task a user
  wants done*. A far simpler computer-use agent performs the demo tasks
  without any of them. Those layers cost latency, memory, tuning surface,
  and failure modes this project has repeatedly paid for. Complexity is
  justified by measured advantage; most of it hasn't been measured. See
  `CLAIMS_MATRIX.md` claim 31.
- **Production maturity is limited.** This is research software being
  hardened. The local-only architecture, the broad fallback surface, and
  the runtime fragility are all real, and none of them belong under the
  word "enterprise."
- **Emotional steering is causal, and that's testable.** Internal state
  reaches text generation through non-text channels (Contrastive Activation
  Addition — a technique that modifies the model's hidde

... (truncated)