A portfolio of experiments with a death criterion declared in advance

🌱 The problem with ideas that never die

Every team has a list of ideas in progress. Most of them are still there months later, because no one wrote down what would count as a failure. Without that, effort goes to the loudest idea and nothing is ever stopped.

📜 The laws

Our registry treats initiatives as a population that moves through states: hypothesis, trial, then selected or culled. Four laws apply.

1. Success metric and kill criterion are written before the start, not after the result.
2. At most eight trials run at the same time. A ninth needs a written exception from the operator, with a name and a date.
3. A trial older than its maximum age is overdue and must be decided: selected or culled. "Let it hang" is not an outcome.
4. A culled initiative must leave an autopsy. Nothing is deleted from the registry.

The decision to cull always belongs to a person. The registry produces the list, and a weekly report names what is overdue.

📊 What it looks like today

On 3 November 2026, 21:53 UTC the registry held 142 initiatives: 122 hypotheses, 9 trials, 4 selected and 7 culled. Among those that have reached a decision, 4 of 11 were selected, a survival rate of 36.4 percent. All 142 entries have a fitness metric and 136 have a cull criterion.

The numbers also show where the rules strain. There are 9 trials, one above the cap of eight. That is the single named exception recorded by the operator on 13 September 2026, and the validator accepts it. And the weekly report lists 7 trials as overdue, waiting for a decision. A rule that says "decide or cull" produces a list of undecided cases when nobody decides, and showing that list is the point.

🛠️ Practice: start a registry

1. Create one file with one record per initiative: key, title, birth date, state, spec link.
2. Add three required fields: fitness metric, fitness target and cull criteria. Refuse records without them.
3. Add a maximum trial age in days, with a default.
4. Write a validator that fails on a missing field, an unknown state, more than the allowed number of trials, and a deleted record.
5. Produce a weekly report of overdue trials and put it on the agenda of a standing meeting.
6. When something is culled, write a short autopsy: what we expected, what happened, what we learned.

✅ Check yourself

• Add a record with no kill criterion. The validator fails.
• Move a ninth initiative to trial. The validator fails unless an exception is named.
• The weekly report lists overdue trials by name and age.
• You can show three culled initiatives with an autopsy each.

🚧 When this does NOT work

• Metrics chosen after the fact. If the metric is edited once results arrive, the registry is theatre. Freeze it at the start.
• Overdue lists nobody reads. The report must be discussed in a meeting, or it is a graveyard.
• Population of hypotheses. Most entries sit at the first state. Many hypotheses and few decisions means the intake is too easy.
• Vague kill criteria. "If it does not work" is not a criterion. Name a number and a date.
• Small teams. For five people a registry may be too heavy. A shared page with the same three fields is enough.

📊 Evidence

➡️ Next

Write the spec for an idea with Spec-driven development with existing task packets, and see how we judge results in AI judging.

Sources

• Spec-driven development with existing task packets: specs.
• AI judging: why “it says it is done” is not acceptance: judging results.
• Costs, budgets and limits: budgets.
• An AI usage policy and register a small team can enforce: registers.