The six-variable model (A17)
State z = (r, q, c₁, c₂, u, v): fuel r, supplied at rate J; the connector q of CASE-189; two catalysts producing each other; a teacher u that tracks a grounded signal θ; and a student v copied from the teacher at rate k_copy = r q c₂.
Rechecked:
- Invariance. The box is positively invariant after the encounter ends. The margins are +0.16 at cᵢ = 1, −18 at cᵢ = 6, +0.224 at q = 0.8, −1 at q = 1 and +0.568 at r = 0.8. For u and v the margin is 0.05 against disturbances of 0.02. Fuel stays sufficient when J > 12.16.
- Copying is kept up. k_copy ≥ 0.64 throughout the box.
- Errors. Teacher error is at most 0.05/0.74 ≈ 0.068. A17 gave 0.135 for the student; the comparison bound is in fact max over k ≥ 0.64 of (0.068k + 0.05)/(k + 0.1) ≈ 0.126.
- Capture. An encounter of 0.15 time units brings q to at least 0.868 and c to at least 1.73, inside the box.
- Knockout. Removing c₂’s production sends c₂, then c₁, then q to zero.
A17’s simulation erased the student and damaged the copying catalyst within the certified region. Recovery there is guaranteed by construction, so it does not test whether repair machinery can itself be repaired. A noise-free rerun of the stated equations agrees with A17’s table.
The twelve-variable extension (A18)
A18 adds a regulatory catalyst s, a copy rule φ with g_φ(u) = (1 − φ)u + φ/2 (φ = 0 is faithful copying), and replicator–mutator allocation shares over operation, maintenance, diagnosis and copying, with reserved minima 0.55, 0.10, 0.10 and 0.05.
Rechecked:
- The margins are 0.416, 0.108, 0.3264 and 1.22, and the upper faces point inward.
- The teacher bound is 0.123, and u ∈ [0.5, 0.8] is invariant. The student bound of 0.246 is valid but loose; the tight bound is ≈ 0.215, and it holds only as φ → 0.
- Not checkable: A18 never states its fuel demand Ψ, so its J > 28.9 cannot be checked.
What carries the certificate. It holds “for every allocation” because of the reserved minima. A18 called these “constitutive architecture”, so they are an assumption, not something the organization earned. The certified decay φ̇ ≤ −0.0144φ comes entirely from a regularizer whose minimum is the faithful rule; the evidence term only contributes ≥ 0. So the rule’s “corrigible” revision is built in. And because φ and the information variables feed nothing physical, rule revision is trivially safe.
A negative result, from an unverified simulation. Comparing adaptive allocation, fixed allocation and a frozen copy rule after five injuries, A18 reported recovery of 0.687, 0.586 and 0.477 at t = 8.125, and 0.794, 0.794 and 0.544 at t = 30. A18: “An allocator that responds intelligently to immediate functional error can still perform worse on recovery if it underinvests in repairing the rule responsible for future corrections. The fact that an allocation mechanism is endogenous does not make it optimal.”
What it shows and does not
It shows that, in one stipulated model, the regeneration of a reconstruction operator’s realizer can be part of what an organization maintains. That is the owner’s U17 question in its simplest form. A17’s own limit: “we’ve demonstrated endogenous reconstruction of the realizer and availability of an operator, not autonomous invention of an entirely new operator type”.
Three gaps remain:
- The coupling runs one way. Copying never feeds maintenance, so the loop in which “successful copying is itself required to maintain the copying machinery” is not built.
- The information is thin. It is one scalar, and grounding is stipulated by injecting θ into the teacher’s equation.
- There is no succession. The student never becomes a teacher.
The motifs are irreversible mass action, as A17 said, “not yet a finite-affinity, microscopically reversible chemical model”. Each face’s inward margin exceeds the disturbance bound (+0.16 to +0.568 against 0.02), so the box certificate is robust invariance of the box itself, viability in THEORY-134’s sense. Kolathaya and Ames’s input-to-state safety (LIT-1893, THEORY-1267) is the weaker, graded guarantee that would apply beyond that margin, where the safe set is enlarged by a class-K function of the disturbance. That organization lowers the cost of further organization is CLAIM-475, which A19 announced as a discovery.
Note of 2026-11-01: a teacher and student in the memory literature
Bechtel and Abrahamsen (LIT-1936, p. 331) report McClelland et al.’s (1995) consolidation model. The hippocampus acts as trainer, replaying items to a neocortex network that learns slowly. The hippocampus is not modelled: its interleaved output is supplied from outside, and stopping it stands for a lesion that leaves graded retrograde amnesia. Two points for this case. Same structure, not shown same mechanism.
- Grounding is stipulated in both. This case grounds its teacher by injecting θ into the teacher’s equation. McClelland et al. supply the teacher’s output itself. In both, what the teacher gets right comes from outside the model.
- Succession. Graded amnesia means older items survive the loss of the teacher: over time the student comes to hold them without it. That is a version of the succession this case says it lacks (“The student never becomes a teacher”). The student becomes independent, though it does not teach. A construction that let copying reduce the student’s dependence on the teacher would test it.
Both models are de novo in Bechtel and Abrahamsen’s sense (THEORY-1290).
No status, title or warrant changes.