🧠 NOESIS

A self-improving AI research brain for Robinhood Chain

An autonomous cognitive system that studies the on-chain tokenised-equity market, forms falsifiable theses, tests them against evidence, and rewrites its own reasoning strategies to become better calibrated over time.

Reasoning substrate: OpenAI · Domain: Robinhood Chain (EVM L2, chain-id 4663)

Quickstart · Example session · Architecture · Whitepaper · How OpenAI is used · Roadmap

---

Why this exists

Robinhood Chain is a public, EVM-compatible Ethereum Layer-2 (Arbitrum Orbit / Nitro) that settles to Ethereum and hosts Stock Tokens — ERC-20 tokenised US equities and ETFs with on-chain Chainlink price feeds and ERC-8056 corporate-action multipliers. For the first time, the microstructure of real equities — spreads, liquidity, corporate actions, price formation — is happening on an open, permissionless, fully observable ledger.

That is an unprecedented natural laboratory for financial economics. But raw observability is not understanding. NOESIS is the missing layer: an autonomous research intelligence that turns the chain's firehose into rigorous, falsifiable, citable knowledge — and gets sharper every cycle.

It is not a trading bot. It is a scientist that never sleeps.

The core idea: a brain that grades itself

Most "AI agents" are static: a fixed prompt wrapped around a model. NOESIS is built as a continual, self-improving cognitive loop grounded in the philosophy of science:

flowchart LR
    P[👁️ Perceive<br/>chain + market telemetry] --> R[🧩 Recall<br/>semantic memory]
    R --> C[💡 Conjecture<br/>falsifiable theses]
    C --> E[🔬 Experiment<br/>+ adjudicate evidence]
    E --> B[(📚 Belief graph<br/>+ calibration record)]
    B --> M[🪞 Metacognition<br/>rewrite own strategies]
    M -.self-modification.-> C
    B --> T[📄 Thesis<br/>formal paper]

Loading

Every cycle the brain makes probabilistic, pre-registered forecasts, later scores them with a Brier score, and uses that empirical fitness signal to edit its own reasoning heuristics. The self-improvement claim is therefore not marketing — it is falsifiable and instrumented: run noesis introspect and watch the Brier score fall.

The architecture synthesises three research lineages:

Faculty	Inspiration	What it does here
Conjecture → refutation	Popper's critical rationalism	Only falsifiable claims are admissible
Bayesian adjudication	Glosten–Milgrom, Kyle microstructure theory	Evidence updates priors to posteriors
Self-modification	Gödel machine · STOP · Promptbreeder · Voyager	The brain rewrites its own strategy library, bounded by calibration

Read the full treatment in the whitepaper.

🚀 Quickstart

git clone https://github.com/bettercallfolio/noesis && cd noesis
python -m venv .venv && source .venv/bin/activate
pip install -e .

cp .env.example .env          # add your OPENAI_API_KEY

# 1. No API key needed — just probe the chain:
noesis chain-status

# 2. Perceive the live market (uses OpenAI):
noesis observe

# 3. Run the full cognitive loop — perceive → conjecture → adjudicate → self-improve:
noesis cycle -n 5

# 4. Inspect the brain's evolving mind and calibration:
noesis introspect

# 5. Promote the strongest corroborated thesis into a formal paper:
noesis thesis --out theses/

# 6. Or let it run unattended as a standing research organism (v0.2):
noesis daemon --interval 900        # a cycle every 15 min; Ctrl-C stops gracefully

The brain's entire mind — episodic memory, belief graph, calibration record and self-authored strategy library — persists to .noesis/memory.json and grows across runs.

🧬 Cognitive architecture

noesis/
├── chain/          # Robinhood Chain connectors (RPC via web3 + REST /rhj APIs)
│   ├── robinhood.py     · network profiles, ERC-20 + ERC-8056 reads, liveness
│   └── stock_tokens.py  · multiplier-aware quotes, spread/liquidity metrics
├── llm/            # OpenAI reasoning substrate
│   ├── client.py        · reason() / narrate() / embed() — the ONLY path to a model
│   └── schemas.py       · Pydantic contracts for Structured Outputs
├── cognition/      # the mind
│   ├── perception.py    · telemetry → salient observations
│   ├── memory.py        · episodic + semantic (embeddings) + belief graph + Brier
│   ├── reasoning.py     · conjecture · design experiment · adjudicate (Bayesian)
│   ├── metacognition.py · self-critique → rewrites the strategy library
│   └── brain.py         · the cognitive cycle that binds it all
└── research/
    └── thesis.py        · promote corroborated beliefs → formal academic papers

Each faculty is small, typed, and independently testable. The hermetic test suite stubs the model, so pytest runs offline in CI.

🤖 How OpenAI is used

OpenAI models are the entire cognitive substrate — this project is a genuine, non-trivial application of the API:

• Reasoning core — a frontier reasoning model (gpt-5, high effort) generates falsifiable conjectures, designs experiments, and adjudicates evidence into calibrated Bayesian posteriors.
• Structured Outputs — every model exchange is constrained to a Pydantic schema via the Responses API, so the cognitive loop consumes typed, validated objects, never free text.
• Embeddings (text-embedding-3-large) — power the semantic memory that lets the brain recall relevant prior episodes across cycles.
• Metacognitive self-editing — the model critiques the brain's own forecasting track record and proposes concrete edits to its reasoning strategies.

See noesis/llm/client.py — the single, audited path through which all cognition flows.

📄 What the brain produces

• A living belief graph of falsifiable claims with live posterior confidences.
• A calibration record — the brain's honest, quantified self-assessment.
• Formal theses — abstract → methodology → findings → discussion → limitations → conclusion → references, rendered to Markdown, promoted only when a conjecture survives refutation above threshold. See a walkthrough in docs/WHITEPAPER.md.

🧪 Development

pip install -e ".[dev]"
ruff check noesis
pytest -q          # hermetic — no API key required

⚠️ Disclaimer

NOESIS is a research instrument, not financial advice. It reads public data and produces hypotheses for scientific scrutiny. Nothing it outputs is a solicitation to trade. Robinhood Stock Tokens are not available in all jurisdictions; consult the official documentation.

License

MIT — see LICENSE.

---

_{Built for the open-source community · reasoning powered by OpenAI · grounded on Robinhood Chain}