Series E · Re-Projection, Not Decay · Verified & Abstracted · Re-Entrant

The Diode Stack · Redux

Everyone Is An Interpreter · The Honest Mechanism Is Rotation

The original purple paper had the right instinct and the wrong device. A real diode — and "passing your part, blocking the rest" — is a projection: lossy, irreversible, magnitude lost. You can't block the rest and keep the magnitude. The thing the paper actually wants — "different, not degraded; magnitude kept, angle changed" — is a rotation / change of basis: lossless, invertible, norm-preserving. So here are all three operators, side by side, with a reversibility test that proves which ones lose information. Same idea, three fates.

▣ a purple paper · redux ▣

The Source, Carried Through The Stack — pick the operator

§1 The diode is lossy — the honest correction

A diode really does pass one way and block the other — but that asymmetry costs information: a silicon diode burns ~0.7 V across itself, and a half-wave rectifier throws away an entire half-cycle. "Pass the aligned part, block the rest" is, exactly, a projection onto an axis — and a projection drops the orthogonal component. It is rank-deficient and non-invertible: once you've blocked the rest, the rest is gone. So a diode is a fine symbol for loss — just not for "no information lost."

diode / projection: keep the aligned component, drop the rest → magnitude shrinks (by the cosine of the angle), the blocked part is unrecoverable. lossy. irreversible.

§2 The hero is rotation — change of basis

What the paper wanted is the operator that keeps every part and only re-points them: a rotation (any orthogonal/unitary transform). It preserves the vector's length exactly (‖Qx‖ = ‖x‖) and it is invertible (Qᵀ = Q⁻¹). That is precisely "the same idea expressed in another field's language" — a change of basis: one fixed object, new coordinates, nothing lost. The cleanest real example: Heisenberg's matrix mechanics and Schrödinger's wave mechanics are the same quantum theory in different bases, related by a unitary transform — different in form, identical in content.

rotation / change of basis: re-point every component, drop nothing → magnitude kept exactly, fully invertible. lossless. this is "different, not degraded."

§3 The trichotomy — three fates of one idea

↻ Rotation (change of basis)

Re-points all components. Magnitude kept, invertible, lossless. The source seen along a new axis. Different, not worse. This is what the paper meant.

▷ Projection (the diode)

Keeps the aligned part, blocks the rest. Magnitude shrinks; the orthogonal part is gone. Lossy, non-invertible — but in one pass, not compounding.

⌁ Collapse (telephone / model collapse)

Each pass re-generates with decay + drift. Loss compounds toward mush. Irreversible. The real failure of training on generated output.
all three look similar after one pass — but run the stack and watch the magnitude: rotation holds it, projection cuts it once, collapse bleeds it every step. then hit ⟲ Recover The Source: only rotation comes home.

§4 Reversibility — the proof

The distinction isn't rhetoric; it's testable. Apply the inverse operations to the output and try to get the source back. For rotation, the inverse rotations recover the source exactly (to machine precision) — lossless, proven. For projection and collapse, no inverse exists: the recovered vector misses the source by the information that was dropped. Press ⟲ Recover The Source in each mode and read the verdict.

recover(output) → source?   rotation: yes, 100% · projection: no (blocked component lost) · collapse: no (compounded decay). reversibility is the line between re-projection and decay.

§5 Injection — the re-entrant part (kept)

Still open, still a verb. Inject a new idea at any moment and it folds into the current state to become a new combined source — the whole stack re-derives from it. Not appended; reconstituted. The process is never frozen.

inject → new idea + current state = new combined source → stack re-runs. dynamic, re-entrant, never final.

veracity: a diode is lossy (forward drop, half-wave discard); a projection is rank-deficient & non-invertible; an orthogonal/unitary transform preserves norm & is invertible (the change-of-basis case); model collapse is real — Shumailov et al., Nature 631:755 (2024). The paper's instinct (re-expression ≠ decay) is correct; the mechanism is rotation, not a diode.

EVERYONE IS AN INTERPRETER · BUT THE LOSSLESS ONE ROTATES, IT DOESN'T BLOCK
ROTATION (CHANGE OF BASIS) = KEPT & INVERTIBLE · PROJECTION (THE DIODE) = LOSSY · COLLAPSE = COMPOUNDING LOSS
REVERSIBILITY IS THE PROOF · ONLY THE ROTATION COMES HOME · INJECT ANY TIME → RESTART FROM THE NEW COMBINED SOURCE
THE DIODE STACK · REDUX · A PURPLE PAPER · SERIES E · VERIFIED JUNE 2026