Perslis Defense
Perslis Defense — the mechanical ant, squads PERSLIS DEFENSE · CHAPTER 3 · COMPOSITION

Constraints compose badly.

Every rule of engagement, every keep-out volume, every deconfliction rule is individually justified. The question nobody measures is what they do together — and the answer is that a rule set which only grows eventually refuses everything.

Measured, not asserted

Across 91 situations an agent actually encountered, we counted how many actions survive the rule set in each one.

dead end = every observed action refused · forced = exactly one left, so the agent no longer decides anything
rule setrulesdead endsforced
bounded working set3069
unbounded501215

Doubling the rules doubles the dead ends. Every one of those rules was earned by a real failure and is individually correct. Unbounded, the agent scored worse than not learning at all — not because a rule was wrong, but because the admissible set only ever shrinks.

This is why safety layers get switched off in the field. It is rarely written down, because it only appears when you measure the set rather than the rule.

Coverage-aware retirement

The usual mitigation is to score each rule by its own evidence quality and keep the best N. That is what we did first, and it is still blind: it scores every rule alone and never asks what the conjunction does.

So we score the set. For each situation the agent actually encountered:

actions_seen(S) − actions_blocked(S) = choices_left(S)

  0 → DEAD END   every option refused; the floor became a wall
  1 → FORCED     one option left; the agent decides nothing here

A rule's protection is the failures it accounts for. Its cost is the situations it pushes into dead end or forced. Retirement drops the worst cost-to-protection offenders until choice is restored. On the measured set: dead ends 6 → 0, at the cost of 29 recorded failures handed back, with forced rising 9 → 14 — situations moving from no choice to one choice.

Retirement lets the admissible set grow back. Pure accumulation never can, and that asymmetry is the whole reason a squad-scale constraint system needs this and not just a bigger rule list.

Why this is the squad problem

One platform with thirty constraints is a tractable rule set. A squad is that set times the number of elements, plus the constraints that exist only between them — separation, sector discipline, fires deconfliction, comms windows. The composition problem is not a detail at that scale; it is the dominant term.

A system that cannot measure its own paralysis will be tuned by the only signal available — operators turning it off. We would rather ship the number.

Next

Chapter 4 — how a rule is earned, and why you can read it →