What Actually Works

The model’s rescue result is the least intuitive thing in it: a cheap fixed floor beats an elaborate one — and the elaboration is what breaks it

Third in the mechanism series — the measured work, as distinct from the implications. The mechanism post sets out the collapse and how it hides; the risk post asks who is at risk. This one asks the question both leave open: what stops it, and what happens when you try to stop it cleverly?

Status. All numbers below are measured in the frozen model and traceable to the paper (doi:10.5281/zenodo.22943641). The translation into practice is an interpretation, as everywhere in this series.

1. The result, in one table

Set the system at the collapsed fixed point — G = 0.049, stuck, switch armed — and engage a floor: a fixed lower bound that stops the cannibalisation switch from seeing an error above its threshold. Then vary the floor’s value.

The critical value is 0.4795, and its meaning is exact:¹ it is essentially the collapsed state’s own error level (E* = 0.4969). The floor must sit at or above what the switch is reading. Below it, nothing. Above it, everything — and how far above it makes no difference at all.

That is already the counter-intuitive part. Most people expect a graded dose-response: press harder, get more. Here the response is a step, and once you are over the step, more is exactly as good as enough.

2. The part that is genuinely strange

Now add sophistication to the rescue — the things a well-resourced person or programme would reach for.

The elaborate rescue fails, and the cheap one succeeds.² Not because the elaborate version is badly built, but for the structural reason this whole series is about: monitoring is itself inward attention. Adding a monitoring channel to the rescue adds a second drain to the same system, and past c_mon_crit = 0.511 the monitoring sustains the collapse it was installed to detect.

And the sharper version of the same finding: a floor you keep checking cancels itself. Hold the floor at 0.7 while continuously re-checking it — ungated monitoring at c_mon = 0.1 / 0.2 / 0.3 / 0.5 — and the outcome degrades: G_end = 0.53 / 0.34 / 0.24 / 0.16. The knowing-floor critical is kc ≈ 0.2.³. Simply re-deriving your support, at low cost, destroys it.

This is the model’s version of a familiar human fact: the support you keep auditing is the support you undermine. The whole value of a floor is that it is held without observation — which is why a cheap, fixed, never-re-examined one is not a lazy version of a good one. It is the only kind that works.

3. Why the value doesn’t matter (above the threshold)

The mechanism is worth stating because it tells you what to look for. The switch arms on the error E = D − G: the reducer is demanding, the generator is not producing, and the gap crosses a threshold. A floor does not repair the generator. It changes what the switch can see — it holds the visible error below the arming line, which disarms c, which releases attention from the inward capture.

Once that has happened, the ordinary dynamics do the rest: with attention no longer held inward, the growth term in dG/dt is no longer starved, and G regrows on its own to the healthy equilibrium. The floor is a starter motor, not an engine — it defeats the latch and then gets out of the way.

Which is exactly why the value is flat above the threshold: the floor’s job is binary. Does it disarm the switch, yes or no? Once yes, it has nothing further to contribute, and pushing it higher adds no force. The only thing that would improve the rescue is having less of it to do — which is what early detection buys.

And the rescue costs nothing when you don’t need it. With no episode, a floor at 0.7 changes the system by exactly zero (max|ΔG| = 0.00e+00, switch never arms, monitoring cost 0.000). It does not suppress legitimate response — sub-collapse dips are actually shallower than in the unregulated system. Holding a protection you rarely use is free.⁴

And the free protection holds over the long run: across sixty consecutive episodes, the unregulated system collapses in the first and never recovers, while the regulated one never enters the stuck attractor at all — re-settling to the healthy equilibrium between episodes.⁵

4. What it implies

Stated as interpretation, and with the obvious caveat that this is a five-state ODE:

Cheap and fixed beats elaborate and maintained. The model gives a specific reason, and it is the same reason as the sensor problem in the mechanism post: the monitoring channel is not orthogonal to the failure. So the practical rule is inverted from intuition — do not improve your support; protect it from being improved. Keep it cheap, keep it fixed, and above all stop re-examining it, because the re-examination is the cost.

And the threshold has a humane reading. The floor works at 0.4795 because that is where the switch’s own reading sits — the protection does not have to be good, or true, or insightful; it has to be at or above what the failure is showing you, and then held without observation. That is a low bar, deliberately, and it is the model’s argument that the bar is low in exactly the direction people find hardest to accept: less force, less sophistication, no checking.

5. What this is not

• Not a treatment protocol. The floor is a single scalar clamp on a switch in a toy system. Nothing here says what a person should do; the translation is interpretation, and the mapping is ordinal and unvalidated.
• Not a claim that support is unnecessary. The result is that a support works where a monitored one fails. Complexity is the failure mode, not support.
• Scoped, on the corpus’s own instruction. The cheap-versus-elaborate contrast is a property of the post-collapse-settled assay, not of deployment — the paper flags this explicitly. It is a statement about rescuing an already-stuck system, and nothing more.
• Not about “knowing” in the human sense. The “knowing floor” is a cost term, not a person doubting themselves. The translation — you cannot maintain a floor you keep re-adjudicating — is the interpretation, and it is offered as one.

Notes

Marking: measured — every number in §1–§3 (floor thresholds, the cheap/elaborate contrast, the knowing-floor critical, zero healthy cost, the 60-episode margins). Interpretation — the practical translation in §4, and the mechanism gloss in §3 (the floor as a starter motor).