A deficiency kept deliberately because it pays, and removable on demand because a compensation channel takes it out of the loop in the rare situation where it would be fatal.
The genus is broader than defects: a feature whose value depends on context — harmful in one situation, beneficial in another. The best you can do with such a feature is control it — enable it where it helps, disable it where it hurts — and if that were free, it would simply dominate. It is not free, and this project is the study of why. Control needs two things: a switch (cheap, usually available) and a detector that tells you which situation you are in (the binding constraint). A flawless switch you cannot time is worthless.
A removable defect is the sharp, asymmetric corner of that genus: the feature is beneficial in the common case and fatal in a rare one, so it is kept enabled by default and switched off — routed to compensation — only on detection. Concretely, it is a deficiency that is:
- Beneficial by default. In most situations it is non-fatal, and it is what the edge is made of: it reduces noise, concentrates capability, and raises output in the chosen area. It is kept on purpose.
- Removable on demand. In the rare fatal situation, an owned, bounded compensation channel — buy the missing work, defer to a principal, rent judgment — takes the defect out of the loop. Removal is a mechanism, not luck.
Two things distinguish this from the familiar praise of specialization:
-
The economic framing. The defect is kept because it pays, and removal is priced per event. The specialist pays for its tail per event; the generalist pays always, carrying rarely-used capability. The advantage condition:
specialization gain per period > fatal-event frequency × price of compensation
When fatal events are rare — the premise — the deliberate deficiency is a computable economic position, not a character trait.
-
The detector-placement constraint. The defect must never include the detector. A flaw that prevents recognizing the fatal situation is not removable — the compensation channel is unreachable exactly when it is needed. So the asymmetry: action-competence may be arbitrarily narrow; risk perception must stay general, cheap, and always-on. The detector never has to solve the out-of-frame situation, only notice it and route — and detection is orders of magnitude cheaper than competence, which is what makes the architecture affordable. The theorem-shaped version, now proved (see
PROOFS.md): a defect is profitably removable exactly to the extent that the detector's distinction survives outside it — and in the worst corner, where the feature and its harm share a substrate, the two rates coincide and no switch, however perfect, can separate benefit from harm.
The pattern is not specific to any one kind of entity. It reads on humans (specialization and the attention it buys), organisms (immune tolerance, costly signals), organizations (focus strategies with insurance and outsourcing as the removal channel), and ML systems (mixture-of-experts is narrow experts plus a router; selective prediction and learning-to-defer are the detector constraint under another name).
The primary deliverable is a short paper, PAPER.md (draft v2): the general
principle, the advantage condition, and the detector-placement iff as the central
result, grounded in the value theory (below), positioned against the prior
literature. It assembles the finished results in the companions below, which are
its technical appendix.
-
PAPER.md— the paper itself: abstract, model, advantage condition, the coupling lemma and the two-sided removability theorem, the observation/capacity split, the Agent World worked example, related work, honest limitations. -
FORMALIZATION.md— the mathematical model: the growth equation, the advantage condition and optimal-detector results (propositions), and the detector-placement theorems. -
PROOFS.md— both directions of the iff, proved. Converse: inside an observation defect the gain-capture rate and the fatal-miss rate are the same number (coupling lemma), so the specialization premium collapses — unconditionally for observation defects, in the inductive regime for capacity defects. Achievability: a detector outside the defect earns an explicit positive premium under the advantage condition, and the two halves meet exactly at the large-loss boundary. Removability is a coefficient (residual separability$\sigma_0$ ), not a yes/no. -
LITERATURE.md— the literature check: per-claim verdicts, what was killed or repaired, and the defensible residual. -
QUESTIONS.md— the live research backlog (formalizing the advantage condition under heavy tails, detector economics, adversarial removal, the human case, fleet composition, whether "removable" is one concept or three). Questions graduate from the backlog into the paper's arguments or into their own investigations. - A design-pattern document (planned, after the paper stabilizes) — runtime-agnostic engineering guidance: when to ship a deliberately narrow system plus tripwire plus escalation path. A translation of the paper's conclusions, citable from engineering projects.
- A Mathematical Theory of Value (Cheng Qian, public preprint) supplies the foundations: diversity is worth paying for while redundancy adds zero (why the defect pays); overconfident error produces negative realized growth — dissipation is the ex-post punishment of the unremovable defect, and this architecture is the ex-ante fix; the is/ought asymmetry (beliefs must keep updating even when the goal frame does not) is the ancestor of the detector constraint.
- Agent World carries the first engineering instantiation: a design-rationale
section and a normative spec field (
onOutOfScope) by which an agent declares its escalation posture. That work lives there; this project is the concept in its own right, and agent-world is one application of it.
Related literature to position against (to be verified before citing): division of labor (Smith), bounded rationality (Simon), hedgehog and fox (Berlin), optimal foraging, ensemble diversity, mixture-of-experts routing, selective prediction and learning to defer, defense-in-depth, graceful degradation. The ground to defend is the combination: kept because it pays, removal priced per event, and the detector placed outside the defect.
Carried into everything downstream:
- Fatal-event frequency is estimated from history, and tail events are precisely what history underestimates. The mitigation is structural — the detector constraint plus bounded blast radius — not statistical.
- The framing is about competence, not values. An entity whose narrow persona serves a different goal frame than its declared one is not exhibiting this pattern; that is undeclared misalignment.
- Declare what is proven vs. conjectured; publish what flunks.
- Docs before code; plain prose over hype.
HANDOFF.mdis the founding document from the originating session and stays authoritative on the concept's history and naming.