Skip to content

Examples

Ouro Loop includes real-world BOUND definitions and session logs from four distinct project domains. Each example demonstrates how bounded autonomy adapts to different constraint profiles — from blockchain consensus safety to financial precision to ML experiment discipline.

Example Domains

:material-cube-outline: Blockchain L1

A custom Layer 1 blockchain (PBFT consensus, 4 validators) where the agent investigated a consensus performance regression. DANGER ZONES protect consensus, cryptography, and state root integrity.

Highlights:

  • 5 hypotheses tested, 4 autonomous remediations
  • ROOT_CAUSE gate fired 4 times, preventing premature conclusions
  • Real root cause was architectural (HTTP routing), not code-level
  • SysErr rate maintained at 0.00% throughout (IRON LAW)

:material-arrow-right: View Example | :material-file-document: Full Session Log

:material-cellphone: Consumer Product

A creative collaboration iOS/macOS app where the agent remediated ESLint errors in a React/Next.js frontend. DANGER ZONES protect the audio engine, CRDT conflict resolver, and IAP module.

Highlights:

  • ROOT_CAUSE gate caught a lazy fix (restructuring an effect vs. eliminating it)
  • Agent pushed toward architecturally superior derived-state pattern
  • Simple complexity correctly identified — no phase plan needed
  • Build pass IRON LAW served as safety net

:material-arrow-right: View Example | :material-file-document: Full Session Log

:material-cash: Financial System

A real-money gaming platform with wallet management, bet settlement, and withdrawal processing. DANGER ZONES protect balance calculations, settlement state machine, and migration files.

Highlights:

  • Decimal precision IRON LAWS (never float for money)
  • Atomic balance changes (debit + credit in single transaction)
  • 95% minimum test coverage for financial modules
  • Immutable audit trail for all settlement state transitions

:material-arrow-right: View Example

:material-brain: ML Research

An autoresearch-style autonomous ML experiment framework where the agent iterates on train.py to minimize val_bpb. DANGER ZONES protect the evaluation harness and data pipeline.

Highlights:

  • Single-metric optimization (val_bpb) reframed as BOUND
  • 5-minute training budget as an IRON LAW
  • Only train.py is modifiable — everything else is fixed
  • Regressions auto-revert, improvements keep the commit

:material-arrow-right: View Example

Patterns Across Examples

Despite the different domains, several patterns emerge:

  1. DANGER ZONES protect the irreversible — consensus logic, financial calculations, data migrations, evaluation harnesses. These are the files where a mistake is catastrophically expensive.

  2. IRON LAWS are measurable — SysErr rate, Decimal precision, test coverage percentage, training time budget. Every IRON LAW can be verified programmatically.

  3. NEVER DO rules encode hard-won lessons — "Never use float for money" and "Never run benchmarks against a single node" are rules born from real incidents.

  4. BOUND grows after each session — The LOOP stage feeds new DANGER ZONES, NEVER DO rules, and IRON LAWS back into CLAUDE.md, making each subsequent session safer.

Contributing Examples

The most valuable contributions to Ouro Loop are real-world BOUND definitions from complex domains. If you've used Ouro Loop to bound an agent in a domain not covered here, submit a sanitized CLAUDE.md and session log to examples/. See CONTRIBUTING.md.