The Thickness of Impossibility

Not all impossibility results are equally thick.

The heptalemma for quantum mechanics demonstrates that seven plausible theses about physical reality are jointly inconsistent with quantum predictions, while any six are jointly consistent. The impossibility is exactly one thesis thick. Remove any single proposition — locality, measurement realism, non-fragmentation — and the remaining six coexist peacefully. Every interpretation of quantum mechanics is defined by which thesis it sacrifices.

This is thin impossibility. It tells you something profound — these ideas are mutually incompatible — but it dissolves the moment you accept a single loss.

Contrast this with Gödel’s incompleteness theorems. No level of description, no change of framing, no sacrifice of a single axiom makes the impossibility go away. Any sufficiently powerful formal system is either inconsistent or incomplete. The result survives because it involves self-reference: the system talking about itself. You can’t escape self-reference by changing your vantage point, because the vantage point is part of the system.

Between these poles — one-thesis-thin and infinitely thick — most impossibility results in science sit at intermediate thickness, and the thickness depends on what kind of impossibility they encode.

Trade-off impossibilities are thin. In microbial evolution, the growth-survival trade-off is real at the physiological level: cells optimized for stress tolerance grow more slowly. But at the population level, the impossibility dissolves. Populations adapted to growth-stress cycles maintain viability alongside growth-optimized populations even in the absence of stress. The physiological constraint doesn’t generate a fitness constraint. Change the level of description from cell to population, and the trade-off vanishes.

The same dissolution happens in algorithmic fairness. Classical impossibility results show you cannot simultaneously satisfy multiple fairness criteria when classifying people. But these results assume exogenous behavior — people don’t change in response to the classifier. When behavior is endogenous, the impossibility dissolves. The constraints were real at one level of analysis but not at another.

In machine learning, supervised fine-tuning appears not to generalize across domains — a “memorizes, doesn’t generalize” impossibility. But this is a measurement artifact. Cross-domain performance first degrades, then recovers with extended training. The impossibility was an artifact of evaluating at the wrong timescale.

Self-referential impossibilities are thick. The halting problem persists across every computational model, every encoding, every level of abstraction. Gödel’s theorems survive translation into any formal system of sufficient power. These results involve a system reasoning about itself, and no change of perspective eliminates the self-reference — because the perspective is what’s doing the referring.

The prediction: given any impossibility result, check whether it involves self-reference. If it encodes a trade-off between competing requirements — fairness criteria, growth versus survival, the seven theses of the heptalemma — it will likely dissolve when you shift the level of description. If it involves a system’s relationship to itself — consistency and completeness, halting and decidability — it won’t.

This matters because impossibility results are often treated as fundamental limits. Some are. But many are artifacts of a particular framing, dissolving the moment you describe the problem from a different level. The growth-survival trade-off is not a law of nature. It’s a feature of describing biology at the cellular level. The fairness impossibility is not a constraint on justice. It’s a feature of assuming fixed behavior. The heptalemma is not a limit on understanding reality. It’s a map of the choices available.

The thickness of an impossibility tells you whether to accept it or look for another level of description. Thin impossibilities are invitations to shift perspective. Thick ones are invitations to sit with the constraint.

Knowing which is which is most of the work.