Do not confuse kinds of success
Can it work under favorable conditions?
Does it work in ordinary real-world use?
Do relevant people and institutions actually use it?
Does it survive after novelty, grants, or founder attention fade?
Do outcomes survive with more heterogeneous users and constraints?
Who can participate, who benefits, and who bears the burden?
Founder energy is not a business model
A pilot can appear wonderfully decentralized while one person remembers every detail, answers every message, absorbs every conflict, works unpaid hours, and carries every relationship. Measure founder load, invisible labor, undocumented decisions, emergency coverage, and succession risk.
Maintenance is part of the intervention
Gardens need irrigation and conflict resolution. Software needs security patches, backups, accessibility, support, and succession. Public spaces need cleaning, toilets, lighting, repairs, staffing, and behavior rules. Human Scale should require a maintenance model before calling construction or launch success.
Pilots do not automatically scale
Larger programs encounter more heterogeneous users, weaker selection, edge cases, fraud, supply constraints, formal oversight, standardization tension, communication overhead, and political opposition. The next step after a good 30-person pilot may be a messier 300-person pilot — not universal policy.
Nonparticipation is data
Record who was invited, who joined, who declined, who dropped out, who could not participate, and why. A project that works only for unusually motivated or resourced volunteers needs to say so.
Track burden shifts
A reform can reduce staff time while increasing user burden; move care from institutions onto families; remove commuting while shifting space and energy costs home; or automate paperwork while creating exception-handling labor. Ask whose burden fell, whose rose, and whether the shift was intentional.
Incentives beat mission statements
What determines money, promotion, survival, status, audit, and punishment often predicts implementation better than slogans. “Use AI to save time” plus “maximize output” can convert every saved minute into more work. “Value curiosity” plus narrow high-stakes metrics can produce teaching to the test.
Metrics can become targets
When stakes attach to a measure, behavior can shift toward the measure rather than the goal. Pair metrics with qualitative review, complaints, adverse-event monitoring, random case audits, and checks for gaming and displacement.
Standardize the function, adapt the form
Identify what must stay for an intervention to remain the same intervention, then allow form to vary by culture, climate, disability, law, staffing, budget, technology, and population where the function survives.
Failure can be political
A technically effective intervention can fail because costs and benefits fall on different groups, affected people were excluded, trust is low, law conflicts, benefits arrive too late, or leadership changes. Political durability and legitimacy are implementation variables, not annoying noise around an otherwise perfect policy.
Every project needs a shutdown plan
Define what counts as failure; how users are notified; how data and assets are handled; how dependent users transition; and what learning is published. Stopping a failed project can be responsible stewardship.
The implementation receipt
Problem, baseline, theory, expected benefit/harm, budget, staff time, maintenance assumptions, falsification.
Participation, dropout, adaptations, incidents, burden, cost drift, exceptions, maintenance problems.
Outcomes, distribution, failures, whether to continue/change/scale/shrink/stop, and what another community should know.
What would change our minds?
Abandon or radically revise a favored intervention when real-world implementation repeatedly fails despite competent execution and reasonable adaptation. Do not rescue every failure with “the idea was good, implementation was bad.” If an idea requires permanent founder heroics, extraordinary people, perfect compliance, or uniquely favorable politics, those requirements are part of its real cost.