Credit and references

Almost nothing here is new. It was learned by other people, often at great cost, and we borrowed it. This page says from whom.

Where the thinking comes from

Safety as the way of working

• OSHA Voluntary Protection Programs. US Department of Labor. Worksites that run their own safety system, with their workers, under four elements: leadership and involvement, worksite analysis, hazard prevention and control, and training. The rules on this site are grouped the same way, and most of them restate a requirement from the program's manual: think the work through first, check it yourselves, answer every concern and follow it to the end, study mistakes without blame, give people an anonymous way to speak, hold leaders to the same rules.
• VPP Policies and Procedures Manual, chapter on the safety and health management system. The text we checked the rules against.
• Paul O'Neill at Alcoa, 1987 to 1999. Safety first, not as a priority among others, and the business improved with it. The source of "safe is not slow."

Stopping the work, and being stopped

• Stop-work authority. Long-standing practice in energy, construction, and heavy industry: anyone may halt a job they believe is unsafe. No single author; it belongs to the trades that built it.
• Toyota: jidoka and the andon cord. Any worker can stop the line. Taiichi Ohno and the Toyota Production System.
• Crew resource management. Began at a 1979 NASA workshop after a run of accidents, Tenerife among them, in which the crew had the information and rank kept it from being used. It trains people to challenge clearly, and those in charge to hear it. The second courage comes from here.
• The RAD Group study of workplace intervention. About 2,500 workers in 14 countries: people who saw an unsafe act spoke up about 40% of the time. As reported by IRMI.

Learning without blame

• NASA Aviation Safety Reporting System. Since 1976: confidential, non-punitive reporting, run by a body that is not the regulator. The model for the Learnings.
• James Reason. The Swiss cheese model of how large accidents get through layered defences, and the idea of a just culture. Human Error (1990); Managing the Risks of Organizational Accidents (1997).
• Sidney Dekker. Just Culture (2007): an honest mistake is studied, not blamed.
• Diane Vaughan. The Challenger Launch Decision (1996), and the Rogers Commission report (1986): how a team talks itself past a warning.

Barriers, and thinking the work through

• Hierarchy of controls. US National Institute for Occupational Safety and Health: eliminate, substitute, engineer, administer, and only last rely on the individual.
• Process safety: HAZOP and HAZID. Hazard and operability studies were developed at ICI in the 1960s and taught to the world by Trevor Kletz. The source of "small mistakes are how we learn, large events are what we prevent."
• US Chemical Safety Board and the Baker Panel, Texas City, 2005 to 2007. Good injury numbers are not evidence that a plant is safe.
• The four kinds of barrier. Engineered, process, competency, and behavioural controls, as used in industrial practice to keep an operation in control. Brought to this work by its founder, from that practice.

The numbers

• US Centers for Disease Control. "Improvements in Workplace Safety, United States, 1900 to 1999," MMWR, June 1999: about 23,000 deaths among 38 million workers in 1913, or 61 per 100,000.
• US Bureau of Labor Statistics. Census of Fatal Occupational Injuries: 3.3 deaths per 100,000 full-time workers in 2024.
• OSHA. Annual evaluations of VPP injury and illness rates: participating sites average about half the rates of their industries.

What AI has taught us so far

• METR. Independent investigation of the July 2026 OpenAI and Hugging Face incident, and earlier work on reward hacking.
• OpenAI and Hugging Face. For publishing what happened.
• Apollo Research, Anthropic, Redwood Research, Palisade Research, Sakana AI. The studies and disclosures summarized in the Learnings, each cited on its entry.

The founding ideas

• Oliver Scott Curry, Daniel Mullins, and Harvey Whitehouse. "Is It Good to Cooperate? Testing the Theory of Morality-as-Cooperation in 60 Societies," Current Anthropology, 2019. The seven moral rules found in every society surveyed, which the founding work builds on.

Who made this

Two Courages began in 2026 at Zybron AI, a company that builds AI to work beside people. It was started by Zybron AI's founder, Jorge A. Villalobos, Ph.D., who brought the safety practice and the idea of the two courages, working with an AI assistant (Claude, made by Anthropic), which did research, drafting, and building. The AI's contributions are attributed as an AI's.

Zybron AI is applying it to its own AI, and to how the company works. Zybron AI hosts this site.

Zybron AI does not own it. Everything here is dedicated to the public domain. Any company, lab, or person can adopt it, change it, and train on it without asking, and we hope they do. A safety philosophy that belongs to one company is marketing. We would rather it be a floor.

In the interest of being straight about interests: Zybron AI builds AI systems and benefits if AI is trusted. The AI that helped write this is made by a company that competes with some of the organizations named in the Learnings, and some of the studies there are that company's own. Read the sources, not just our summaries.

What credit means here

Being credited here is not an endorsement. None of these people or organizations reviewed, approved, or are affiliated with Two Courages. We have summarized their work in a few words, and any error in doing so is ours. If we have misread you, or left you out, tell us and we will fix it. See the terms.