Define What Must Stay Fixed and What May Adapt

Scalable AI systems become reliable by protecting a small number of invariants while allowing legitimate variation everywhere else.

A format should organize the work, not decide what knowledge is allowed to survive.

Separate invariants from variables. An invariant is something whose loss would damage the purpose or integrity of the system. Examples include: * safety requirements, * accessibility, * essential user controls, * required data, * pedagogical completeness, * evaluation criteria, * legal or policy constraints. A variable is something that may legitimately change with context: * length, * number of examples, * presentation pattern, * density, * sequence, * visual arrangement, * amount of supporting explanation. Weak specifications often try to control both categories equally. They prescribe the exact shape of every output because sameness is easy to test. But sameness is not the same as consistency. A strong specification is strict about what matters and flexible about how those requirements are satisfied. This gives AI enough freedom to respond to different problems without losing the identity or integrity of the system.

**Example A — the template becomes the author.** A course uses a standard lesson card. Most concepts can be explained with one short paragraph and one example. A later concept requires a worked example containing five reasoning steps. AI forces the lesson into the existing card by deleting three steps and replacing them with a summary sentence. The layout remains perfectly consistent, but the learner can no longer see how the answer was reached. At that point the template is no longer presenting the content. It is authoring it. AI systems are particularly vulnerable to this failure because they are very good at satisfying visible structural constraints. If asked to fit something into a format, they may sacrifice nuance or completeness without explicitly saying what was lost. The correct response is not automatically to shorten the reasoning. The system should first ask whether the fixed card height is a genuine pedagogical requirement or merely a presentation convention. If the reasoning is essential, the format must adapt. **Example B — a variable mistaken for an invariant.** A learning platform defines every lesson as exactly one screen because the first lessons were short. As the curriculum grows, some concepts require comparisons, worked examples, or longer explanations. AI begins deleting material to preserve the one-screen pattern. The system has confused a variable with an invariant. The true invariants may be very different: learners should always know where they are, primary actions should remain accessible, required explanations must not be lost, and progress should remain visible. One lesson may occupy one screen and another three. They can still belong to the same system because the things that matter remain consistent.

More from the same record, pulled by shared topic — not written new for this page.

Scalable AI systems become reliable by protecting a small number of invariants while allowing legitimate variation everywhere else. A format should organize the work, not decide what knowledge is allowed to survive.