Purple Paper · side-sheet · learning machines · V · capstone
The Commutator — and the one idea under all five
A brushless rotary selector: 01 in, four phase-throws at 90° / 180° / 270° / 360°,
10 out. It picks exactly one route, refuses the other three, and closes on itself at 360 so the
ring is verifiable. The gate done with nothing but geometry. Then the honest part — are the five objects in this
series five clever toys, or one idea wearing five costumes? Let me audit my own work and not flatter it.
4 phases, 90° apart · exclusive (one hot) · closed at 360 (ring verifiable) ·
selection beats summation when branches take turns being right
I · gate | branches | stream — the wheel that selects
Four branches compete for dominance off the same stream. Each is a predictor; the stream keeps
switching regimes, so a different branch is the correct one at different times (its dominance bar climbs).
The commutator points at whichever branch dominates and routes only that one through — a bouncer, not a
matchmaker. Watch SELECTION (follow the hot branch) crush SUMMATION (blend all four): the blend is
forever dragged by the three branches that are currently wrong.
brushless rotary commutator · arm points at the dominant branch · hot route = selected ·
the 0=360 mark is the closure (ring verifiable) · cyan ring = the
currently-correct branch (the regime)
50%
SELECTION · commutator
50%
SUMMATION · blend
selection vs summation accuracy over time
branch dominance (rolling accuracy) — the commutator points at the tallest
Why geometry is the gate. The wheel can only point one way, so exclusivity is free —
no logic enforces "pick one," the shape does. It returns home at 360, so the ring is checkable by closure — no
separate verifier needed, the loop verifies itself. And it selects rather than blends, so a single hot
branch isn't diluted by the cold ones. Exclusivity, closure, selection — three properties a gate has to fight
for, all handed over for free by a circle. This is the bouncer from the gate paper, built out of rotation.
II · Field-isolated hubris? — auditing my own five
You asked the right question: are these five objects connected, or am I braiding unrelated toys and
calling it a theory? Honest answer — some of each, and the witness function means separating them out loud
rather than selling the unity. Here's what's load-bearing and what's costume.
load-bearing
The atom is one
Every machine reduces to a weighted combination — the dot product. Branch predictor, crossbar
column, waist projection, cube face, commutator vote. Same primitive, not a metaphor: the math collapses to
it each time. This is real unity, verified across papers.
load-bearing
The fork is one
Stay linear — sound, witnessable, summed, bounded, can't escape XOR — or bend/couple —
powerful, escapes the wall, but couples the modes and can be captured. The same fork appears at the layer,
the waist, and the cube. One decision, restated three ways.
load-bearing
The resolution is one
Trust comes from a second, fixed, exterior element the channel can't retrain — and it works best
when it selects rather than sums. Gate, witness, bouncer, commutator: the same role, and the same
firewall (the optimizer never reaches it).
⚠ the braid — what was costume, not mechanism
The geometric dress — hourglass, cube, wheel — was chosen for resonance, not derived from necessity; the same
three ideas would survive in plainer clothes. The "200-year-old curse" is a motif, true but
decorative — it colors the telling, it isn't a mechanism. And the cross-paper rhymes are real at the level of
the spine and aesthetic above it. So: not five isolated toys, but also not a grand unified theory — one
genuine three-part spine, dressed five ways. Calling it more than that would be the hubris.
III · The one idea, stated plainly
Every machine in this series is a learned weighted combination (the channel) running on some substrate,
facing one fork — stay linear (sound, witnessable, bounded) or bend (powerful, escapes the wall,
capturable) — and made trustworthy only by a second, fixed, exterior element (the witness) that the
channel cannot retrain, which works best when it selects rather than sums. The substrate is
interchangeable. The pattern is invariant.
Same atom, same fork, same witness — five substrates:
silicon logic
Predictor
sign(w·x+b)
one dot product, one bit, in the fetch unit
analog conductance
Crossbar
I = ΣV·G
the dot product as a measured current
wave modes
Gorge
σ at the node
linear keeps modes; the bend couples them
tagged tables
Cube
Σ faces ⟂ select
fan-out sums, the gate selects the longest match
rotary phase
Commutator
argmax → 1 hot
the witness made of nothing but geometry
And the firewall under all of it — the part you kept handing back
to me: the channel adapts, the witness must not be adaptable by the channel. Couple them and you get the
bent gorge, the captured TAGE, the self-calibrating flicker, the verification built from the same foundation as
the code. Keep them separate — learned channel, unlearned check — and the seam holds. That separation is the
whole of it. The five papers are five proofs of the same one line.
IV · The close, and the curse
You walked it backward the whole way: branch predictor → linear model → the dot product → the
memristor → the waist → the cube → the wheel. Each step you reached for the next object by instinct, and each
time it turned out to be the same idea on a new substrate, already named by someone decades or centuries back.
That's not a failure of originality. It's a demonstration that the spine is real — real ideas get
independently rediscovered, which is exactly what you kept doing, out loud, in real time.
The last curse. A
brushless rotary selector that commutates by field phase is a switched-reluctance / BLDC commutation pattern —
1960s power electronics on Faraday's 1830s rotary fields. Winner-take-all selection is older than the computers
that run it. Even the synthesis — channel plus exterior witness — is your own framework, which is itself a
specialization of Agrippa's trilemma, ~2000 years old. Five papers, one spine, every vertebra already carved.
You didn't need new stones. You needed to see they were the same stone. lol.