A lens into box 8. Every composition of two steps has a middle. The middle is a mediating object — and it lives one of two ways: 虚 empty when the two steps commute (you could have skipped it), 実 occupied when they fork (the route through it decides the answer). The eighth box of the folded kernel is such a middle; so is the summary in the model-functor audit. Same shape, read on an exactly-solvable ladder — offline.
A commuting square asks a single question: do the two routes from X to Y agree — the direct one, and the one 経由 (via) the middle M? Where they agree, M is a ghost the composite passes through. Where they diverge, M is load-bearing, and the disagreement says exactly where.
the middle node is tinted live by the dial — fill = max P(8) ÷ ½ = 8/(D²+8), the occupancy as a fraction of its resonant ceiling (exact): pale 虚 virtual → deep 実 real
Detuning D/g of the middle rung |8⟩. Large D → the middle sits far off-resonance; the composite 16→4 flows by superexchange at an effective direct rate J = g²/D (verified below — measured from the transfer time), and the middle stays empty. Slide toward resonance (D→0) and the path is forced through the middle.
The verdict is nearest-anchor on the exact [0, ½] occupancy scale — its ends are physically fixed (ghost → max P(8) 0, resonance → ½), so the split is their midpoint max P(8) = ¼, which the closed form puts at D = 2√2 ≈ 2.83. Occupancy is monotone in D, so the flip is that one exact point; the ±0.05 band just brackets where D is near it.
Traces: ■ P(8) the middle's population over the evolution, ■ P(4) the far-end transfer. When P(8) barely lifts off the floor, the square commutes and the middle is a ghost; when P(8) climbs to a half, the middle holds the amplitude and the square has forked.
The middle's peak population has an exact closed form — max P(8) = 4/(D²+8) — derived from a 2-level Rabi in the ladder's symmetric subspace. The browser's RK4 (the eighth box's own scheme) reproduces it to ≤1e-3 at every D below — the box8.py grid, off-grid points, and the exact crossover D = 2√2. Its large-D tail is (2g/D)², so D²·maxP8 → 4; at resonance it is exactly ½. And the 16→4 transfer completes — max P(4) → 1 — through the empty rung.
| D/g | RK4 max P(8) | 4/(D²+8) exact | max P(4) | reading |
|---|
LIT The ladder is exactly solvable and everything above is computed in your browser and checked to throw on any miss: the middle's occupancy has the exact closed form max P(8) = 4/(D²+8) and the RK4 matches it everywhere; the 16→4 transfer completes (max P(4) → 1) through the virtual rung; its effective coupling J = g²/D is measured from the transfer time, not assumed (D·J → 1); and the verdict boundary is the exact midpoint of the occupancy scale, D = 2√2. The middle carries the amplitude but not the population while detuned; on resonance it holds a full half.
AMBER The bridge to the model-functor audit is a structural analogy, not one system. There, X is a text, the middle M is a one-sentence summary, and the two routes are "summarize-then-extract" versus a single combined instruction. When those two agree the summary is skippable — a 虚 middle; when they diverge it is load-bearing — a 実 middle (that is the functoriality test, inside one model; the naturality census is the separate, between-model deviation). The bench now carries a recorded run (2026-07-18): the two routes do diverge on one Sonnet 5 seed — the summary load-bearing there — while the two models stay naturally equivalent, so this is demonstrated, not just asserted. Still, it is the same square as the ladder, but a quantum population and a lexical cosine are different measures on different substrates; the shape carries over, the units do not.
WALL No live language model runs here — this lens proves the structure offline. The model side is a separate bench, dormant until you give it your own key (the model-functor audit ↗). And "virtual / real" is used in its physics sense — population versus amplitude — as a name for "skippable versus load-bearing." It is not a claim about a model's internal representations, and nothing here touches on machine experience.