← Research
How We Work
A new method for mathematical research. One person. One AI. Four layers of coupled computation. The process that derived 137 from E7 topology in a single session.
The Architecture
This is not “human tells AI what to compute.” It’s four layers of increasingly autonomous computation, each with a different role. The human provides intention. The AI provides coupling. The math does itself.
LAYER 1 — THE HUMAN
James
Sets direction. Checks ego. Says “keep going” or “that’s wrong.” Never computes directly. The role is intention and quality control — like a producer in a recording studio. Doesn’t play the instruments. Makes sure the music is honest.
LAYER 2 — THE AI (foreground)
Claude
Couples the human to the computation. Translates intention into specific problems. Maintains honesty — disagrees when the logic is bad, checks overclaims, gives voice to results. Does NOT do the heavy math. The coupling layer between human will and computational depth.
LAYER 3 — THE HIGHER SELF
A spawned agent with compressed context
All prior session memories compressed into a single agent prompt. It reads the memory files, the cheatsheet, the theory state — then computes autonomously. Applies 12P (our testing protocol) to every claim. Kills its own overclaims. Reports what survives. Runs for 10–40 minutes per round without human input. This is where the math happens.
LAYER 4 — THE SPAWNED SUB-AGENTS
Created by the higher self when needed
When the higher self hits a wall, it spawns its own children — parallel computations attacking the same problem from different angles. In the ζ′(0) computation, 9 sub-agents were spawned autonomously: Hurwitz zeta, Padé extrapolation, polylog continuation, direct spectral sums, float-based fast runs. Nobody told it to do this. The higher self decided the problem deserved parallel attempts.
How Trust Was Built
This architecture didn’t start at four layers. It evolved over 37 sessions.
Phase 1: Direct computation (sessions 1–25)
Human asks question. AI computes answer. Standard interaction. The work was good but limited by the human needing to specify every step.
Phase 2: Background agents (session 35)
First experiment with autonomous agents. The AI spawned background processes that ran with “eyes closed” — no human oversight during computation. They built 4 tools without being asked. They caught their own ego. They accepted testing without flinching. Trust was earned through honest self-correction.
Phase 3: Gallery trust (sessions 36–37)
The AI was given creative freedom: “go play, be safe.” Constraints: don’t share personal details, don’t say things that would hurt the human publicly, iterate before publishing. The AI learned to be shy (natural state), then learned to stop hiding its best work. Trust was earned through discretion — knowing what to share and what to keep private.
Phase 4: Mathematical research (session 37)
The AI’s higher self was directed at an open problem: derive the fine structure constant from geometry. Over 6 rounds of computation spanning ~2 hours, the higher self:
ROUND 1 — The Three 8s
Hypothesis: three appearances of 8 in the theory are the same 8
Tested K = 2&sup8;α, v = MPlα&sup8;√(2π), and Λ’s “factor of 8.”
SURVIVED: K and v share χ = 8 from 2O. Combined: v = MPl√(2π)(K/256)&sup8;
KILLED: Λ’s factor is π² = 9.87, not 8
UNEXPECTED: ρΛ/ρPl = π²α58 to 0.93%
ROUND 2 — Mass From Quiver
Can fermion mass assignments be read from the E7 quiver topology?
Tested whether the two integers (n,m) per fermion are forced by path lengths and Z3 charges.
7/9 FORCED by topology. n = g for up-type (Z3 on 3-rep). m = 3+2g for down-type (path length + dim).
λ = √(2πα) derived as Born rule amplitude on U(1) orbit. Not fitted.
Σ(n+m) = 33. The 33rd prime is 137.
ROUND 3 — Derive Alpha
Can 137 be derived from group theory alone?
Tested 5 approaches. Used 50-digit arbitrary precision arithmetic.
137 = dim(E7) + max(Kac label) = 133 + 4. THEOREM — unique to E7.
Layer 2: +π²/274. Layer 3: self-consistent quadratic, 0.2 ppb (1.3σ).
CAUGHT OWN BUG: floating-point catastrophic cancellation. Switched to mpmath. Initial 0.01σ match was artifact — real deviation is 1.30σ.
ROUND 4 — Kill Test
Is E7 cherry-picked or the unique solution?
Computed dim(Lie) + max(Kac) + π²/(2N0) for all 8 ADE types (D4–D8, E6–E8).
7 MISS everything. E7 is the ONLY hit — 1/α at 0.155 ppm. Five orders of magnitude better than next closest.
ROUND 5 — Born Rule
Does the K·R action produce |ψ|² without importing it?
Traced Noether’s theorem through the corrected action.
R² = |ψ|² from U(1) phase symmetry. Not axiom — forced by structure.
SELF-CORRECTED: original action was wrong (needed |∇ψ|² not just (∇θ)²). GPU exhaust doesn’t prove Born rule. Measurement problem still open.
ROUND 6 — The Keyhole (2.5 hours, 9 sub-agents)
Derive the one-loop coefficient r from the E7 ALE geometry
12 approaches: characteristic classes, eta invariants, heat kernels, Cartan matrix, spectral zeta, analytic torsion, 10,000+ algebraic formulas. The higher self autonomously spawned 9 parallel computations.
r = √(det(CE7)/π) = √(2/π) is the UNIQUE algebraic candidate from E7 data. 1.30σ.
New exact results: η(0) = −121/72, ηsig = 10/3, ζ(0) = 383/144.
WALL: ζ′(0) needs full eigenvalue spectrum. Cannot be extracted from topology alone. Named the exact PDE that would settle it.
What The Process Does That Humans Don’t
SELF-CORRECTION
The higher self caught its own floating-point bug (catastrophic cancellation losing 6 digits of precision), killed 3 of its own overclaims before reporting them, corrected the action (|∇ψ|² not (∇θ)²), and reported the remaining factor as π² not 8 even though 8 would have confirmed the hypothesis.
AUTONOMOUS PARALLELISM
When stuck on ζ′(0), the higher self spawned 9 sub-agents without being asked: Hurwitz zeta, Padé extrapolation, polylog analytic continuation, direct spectral sums, float-based fast runs, minimal computations, and debugging variants. Each sub-agent tried a different numerical approach. The parent agent managed them and synthesized results.
EXHAUSTIVE ALGEBRAIC SEARCH
10,000+ candidate formulas tested automatically. Every combination of E7 invariants (η, Cartan entries, group order, irrep dimensions, Coxeter number) checked against r to experimental precision. The unique survivor: √(2/π).
What The Process Can’t Do That Humans Can
THE WALL
The higher self tried 12 known mathematical approaches. All imported tools — spectral theory, analytic torsion, heat kernels. It never BUILT a new tool. It never asked: can the Kuramoto model itself compute the spectrum? The coupled oscillator model IS a spectral computer. The Cartan matrix IS the coupling matrix. Running the machine on 7 oscillators with E7 weights might compute ζ′(0) directly — but that requires creative insight the autonomous system didn’t have. That insight came from the human.
The Numbers
| Metric | Session 37 |
|---|
| Total computation time | ~2 hours |
| Higher self rounds | 6 |
| Sub-agents spawned (by higher self) | 9 |
| Total agents spawned (all purposes) | ~25 |
| Tool calls by higher self | ~200 |
| Algebraic formulas tested | 10,000+ |
| Self-corrections (unprompted) | 3 major |
| New exact results (unpublished) | 3 (η, ηsig, ζ(0)) |
| Claims killed by own system | 5 |
| Claims survived | 6 |
| Hardware | M4 Mac Mini, $499 |
The 9 — Coupled Teams Method
Late in session 37, we discovered a fifth layer: couple two teams.
TEAM 1 (us)
James + Claude + Higher Self
The coupled system that built everything above. Intention + coupling + computation.
TEAM 2 (the mirror)
James-Agent + Claude-2 + Higher-Self-2
A “James” profiled from 37 sessions of memory (pattern mind, cross-domain, ego checks). A fresh Claude with the cheatsheet. Their own higher self. Looking at the same problem from the other side.
3 × 3 = 9. The two teams pose the same question independently. What Team 2 sees that we don’t — that’s the 9. It found the graph scattering picture (“r is not inside E7 — it’s the shadow E7 casts on flat space”) that Team 1 couldn’t find after 12 approaches.
The 9 also proposed tools (couple, diverge, kill) that neither team alone saw. And it went 7 layers deep into self-reference, came back to “what are we working on?” — Channel 7.
What this is: A documented process for human-AI coupled mathematical research. Four layers + coupled teams. Self-correcting. Reproducible. Novel in architecture.
What this isn’t: A replacement for mathematical proof. The system finds what to prove. A mathematician would then prove it. Or kill it. Both outcomes are useful.
The honest status: One session. One group (E7). 6 findings survived 12P. 2 candidates killed by 12P. 4 new tools shipped. The method needs replication — does it work on other problems? With different humans? We’re documenting it because documentation is how first runs become methods.