cert-machine · the splitting atlas · γ = 0.01 · rebuilt from certificates at every build

The crowd splits — mean-field games beyond the uniqueness wall

A congestion-averse crowd in a single-well cost landscape can settle into TWO peaks — or THREE. The mechanism is a band-pass response with an exact, discount-free crossover at σ* = 1/(8π²); the phenomenon is proved — two computer-assisted theorems and a bracket table of certified instances whose enclosure balls fix the exact peak count of the exact solution, bracketing every predicted threshold. The crowd re-weights a harmonic the potential already contains. Floats draw the map below; the diamonds are the territory.

exact crossover
1/(8π²)
σ* = 0.012665147955292… — independent of the discount γ, decided in exact rationals (Machin bracket, width 1.3e-44)
theorems
2 + table
certified peak counts both sides of both thresholds — incl. THREE peaks from one well, and the amplitude threshold pinned inside [0.13, 0.14]
tightest ball
2.97e-14
ℓ¹_ν enclosure radius — density positivity and branch certified over every ball
the window
1/16 < r < 1/4
A₂/A₁ where a one-well potential splits the crowd (σ, γ → 0 limit); third harmonic: (1/27, 1/3)

Published, not peer-reviewed, not independently rerun. Two priority claims, each fenced: to our knowledge these are the first certified equilibria of a mean-field game whose exact peak count strictly EXCEEDS the potential's well count — the nearest published phenomena are spontaneous instability (arXiv:2605.20213) and peaks that mirror the potential (arXiv:1705.10741), and we found no validated-numerics equilibrium enclosures for MFG anywhere in the literature; and the base instance of the program (the companion page, mfg-congest.html, released together with this atlas) is to our knowledge the first validated-numerics enclosure of an MFG equilibrium at all. Refutations of either claim are invited. Every claim on this page re-runs from this repository: the enclosures come from two independent implementations of the radii-polynomial argument that agree on the certified radius to the last digit, the peak counts derive only from certified region signs, and every instrument carries red controls that fire. Grid dots are float candidates and are labeled as such; nothing on this page is a forecast.

the phase map

Peak count over (σ, A₂/A₁)

Dots: 224 float equilibrium solves (N = 32, fresh seed each, pinned by sha256) · dashed line: the closed-form linear-response boundary r_c(σ) · diamonds: the certified theorems — filled where EVERY density in the enclosure ball has exactly two maxima, open where exactly one. The T7/T8 pair straddles the boundary.

0 0.05 0.1 0.15 0.2 0.25 0.3 1e-3 2e-3 5e-3 1e-2 2e-2 4e-2 σ (log) A₂/A₁ σ* = 1/(8π²) r = 1/4 r_c(σ) ceiling holds (float) SPLIT: peaks > wells (float) boundary r_c(σ) — predicted theorem: 2 peaks (filled) / 1 peak (open)
The wedge between the boundary curve and r = 1/4 (where the potential itself goes two-well) is where gain-weighted well-counting beats flat well-counting — the crowd re-weights the second harmonic the potential already contains, across the ¼ critical-point threshold. The predicted boundary meets r = 1/4 exactly at σ* (dashed vertical). Certified: two-peak theorems (filled diamonds) at σ = 0.002 inside the wedge; one-peak theorems (open) below the threshold (r = 0.12, 0.13) and past the crossover (σ = 0.02). The threshold is pinned inside [0.13, 0.14] by T7/T8, and the exact-rational prediction r_c = 0.132725… lands inside.
the main specimen

T1 — one well, two peaks

The potential (dim, dashed, right scale — a positional comparison, never a magnitude one) has a single well at x = ½. The certified equilibrium density peaks at x ≈ 0.363 and 0.637 with a saddle between — the exact solution provably carries two strict maxima, with the certification margin six orders above the enclosure pads.

0.895 0.934 0.974 1.014 1.053 0 0.25 0.5 0.75 1 x m -0.002 -0.001 0.001 0.002 0.004 V m*(x) candidate V(x) — the one-well potential
σ = 0.002 · γ = 0.01 · A₂/A₁ = 0.2 · certified ball radius 2.52e-13 · min m over ball ≥ 0.8944. The curve is the candidate (floats are the map); the theorem is that EVERY density in the ball has exactly 2 strict maxima — dots mark the certified locations — while V has exactly 1 well.
the three-peak specimen

T6 — one well, three peaks

The mechanism does not stop at one extra peak. With the third harmonic inside its predicted window — the Chebyshev U₂ law gives (1/27, 1/3) at k = 3 — the crowd splits three ways at the bottom of the same single well.

0.883 0.941 1.000 1.058 1.117 0 0.25 0.5 0.75 1 x m -0.004 -0.002 0 0.002 0.004 V m*(x) candidate V(x) — the one-well potential
σ = 0.002 · γ = 0.01 · A₃/A₁ = 0.25 (third harmonic, A₂ = 0) · certified ball radius 2.73e-13 · min m over ball ≥ 0.8823. The curve is the candidate (floats are the map); the theorem is that EVERY density in the ball has exactly 3 strict maxima — dots mark the certified locations — while V has exactly 1 well.
the theorems

Two theorems, 6 bracket rows — every bound shown

Honest counting: T1 and T6 are the theorems; the rest are rows of a bracket table under them — negatives below the amplitude threshold and past the crossover, replications, and the threshold pin. Each row is an enclosure with full-ball local uniqueness (even AND odd blocks — evenness is a corollary, not an assumption), and every bound is displayed: the radius, the contraction bound Z₁, the closure margin, and the certified density floor. Records in certs/.

instanceσA₂/A₁ (or A₃/A₁)Nνball radiusZ₁marginmin mpeaks / wellsrecord
T50.0010.151761.0155.31e-130.92602.11e-60.96492 / 1terra-recert-t5.json
T6 (3-peak)0.002A₃: 0.25961.022.73e-130.90463.95e-60.88233 / 1terra-recert-t6.json
T20.0020.12961.022.25e-130.88895.54e-60.90591 / 1terra-recert-t2.json
T70.0020.13961.022.30e-130.88995.43e-60.90441 / 1terra-recert-t7.json
T80.0020.14961.022.32e-130.89095.32e-60.90302 / 1terra-recert-t8.json
T40.0020.15961.022.37e-130.89195.21e-60.90162 / 1terra-recert-t4.json
T1 (main)0.0020.2961.022.52e-130.89704.69e-60.89442 / 1terra-recert-t1.json
T30.020.2641.052.97e-140.52354.99e-40.96211 / 1terra-recert-t3.json
the companion result

Certified multiplicity where uniqueness theory is silent

Ergodic quadratic MFG, V ≡ 0, σ = ½. Past the pitchfork c* = −σ²(2π)² the constant state, the symmetry-broken branch and its half-shift mirror are enclosed in PAIRWISE DISJOINT ℓ¹_ν uniqueness balls with certified positive density, at each of six couplings c = −11 … −24: AT LEAST THREE distinct exact solutions at every listed coupling, exactly where Lasry–Lions monotonicity makes no claim. The radii grow and the density digs toward vacuum as the coupling deepens — min m down to 4.45e-4 at c = −24 — and the half-shift symmetry is thereby proved to produce a genuinely different solution, not a relabeling. At c = −9.5, inside the monotone regime, the branch collapses onto the constant and no claim is made — the recorded boundary. (certs/mfg-cap-multiplicity.json)

the census

EXACTLY three, for the truncated system

Krawczyk exhaustion over an explicit printed box B: every subbox of an adaptive partition is eliminated by an interval-residual or Krawczyk exclusion, and each surviving box is isolated by Moore–Krawczyk K(X) ⊂ int(X) — existence AND uniqueness per box. Box-bounded, truncation-level; the PDE-level count is the stated open problem.

Galerkin Ndimensionsboxes processedsecondssolutions in Bmatching
N = 252,0770.1EXACTLY 3constant · branch · mirror, one-to-one
N = 3721,6420.8EXACTLY 3constant · branch · mirror, one-to-one
N = 49335,86217.7EXACTLY 3constant · branch · mirror, one-to-one
N = 5116,954,073586.7EXACTLY 3constant · branch · mirror, one-to-one
the regime map

Multiplicity decided over whole rectangles

Coupling plane (s, d): s = symmetric cross-interaction, d = attack–defense asymmetry. Every rectangle of coupling matrices is decided UNIFORMLY over its whole cell — 11,628 cells certified MULTIPLE (two exact solutions for EVERY parameter in the cell, disjoint balls, positive densities), 24 UNIQUE, and the undecided region drawn as itself — hatched, not rounded away. The exact area identity (decided in rationals) confirms the three regions tile the rectangle. (certs/mfg2p-regime-map.json)

0.0 3.0 6.0 9.0 12.0 15.0 0.00 0.30 0.60 0.90 1.20 1.50 symmetric coupling s asymmetry d MULTIPLE — non-uniqueness certified (11,628) UNIQUE (24) UNDECIDED — honest (9,915)
Adaptive refinement concentrates cells where the answer changes. Re-run: node labs/mfg2p/regime2p.js.
the selection wing

The face-dimension law

When the cost is class-independent, the two-population equilibrium is a FACE — a positive-dimensional set the equilibrium conditions cannot pin to a point — and its exact tangent dimension obeys a purely combinatorial law: k = |shared| − cons + z, where z counts exit-free components of the shared subgraph. The natural shortcut (drop z) is correct exactly when every shared component touches an exit, and undercounts by exactly z otherwise.

The law is decided against the exact ℚ null space on two seeded 4,000-network ensembles — the shortcut fails on 572 and 571 instances respectively, precisely the z > 0 cases, and every failing instance is ENUMERATED in the record so any reader can re-run any one. A constructed exit-free-cycle family realizes every deficit. Exact over ℚ; combinatorics on the constraint matrices, not an enclosure. (certs/facelaw-theorem.json)

the attention wing

A decidable attention flow, and the bifurcations that weren’t

Rational-kernel token dynamics on the sphere — (1 + β⟨x_i,x_j⟩)^p, chosen so equilibrium and stability are decidable in exact ℚ; never a Transformer/softmax claim. Three theorems: the consensus spectrum {0, −1} is β- AND p-free (the kernel slope cancels identically — decided by exact dual-number expansion); the ⟨u,v⟩ = −1/β two-cluster family’s cross-weights vanish identically, to first order for every p ≥ 2 (and NOT at p = 1 — the model’s own honest candidate for a true bifurcation); and the reduced flow below. Plus a four-artifact catalogue of how finite float budgets manufacture bifurcations — one artifact re-demonstrated live at every battery run, with its exact refutation. (certs/attnflow-theorems.json)

0 0.5 1 1.50 2 -1 -0.5 0 0.5 1 c = ⟨cluster, cluster⟩ ċ β = 1.5 β = 2.5 β = 4
The reduced two-cluster flow ċ(c) = 2(1+βc)²(1−c²)/[(1+β)²+(1+βc)²] for three β. The curve TOUCHES zero at c* = −1/β (a double zero — multiplicity decided by exact division) and never crosses: one-sided semi-stability at every β > 1, so every pitchfork claim about this flow is refuted, exactly.
the mechanism

The mechanism, in four lines

Linearize the discounted congestion MFG about its flat equilibrium: the density response to a cost mode of frequency κ is

m̂ = −c(κ)·A with c(κ) = κ / [ (1+σκ)² + γκ ]

— a band-pass, peaked at κ = 1/σ. The second harmonic overtakes the fundamental exactly when σκ₁ = ½, and the γ-terms cancel IDENTICALLY (an exact polynomial fact, certs/terra-sigmastar.json), giving σ* = 1/(8π²), discount-free. The equilibrium counts wells by gain-weighted harmonic amplitudes; the potential counts them flat; the splitting window is exactly the gap between the two counts.

scope, stated plainly

Mechanism is linear response, derived in advance and confirmed — Turing-adjacent mode selection in a driven, monotone-regime system, distinguished from spontaneous instability (Karuturi, arXiv:2605.20213) and from peaks that mirror the potential (Cesaroni–Cirant, arXiv:1705.10741). The theorems certify existence, full-ball uniqueness (both parities), density positivity and the branch selection; the peak counts never trust a float sign.

reproduce

Reproduce

Every certificate re-runs from this repository:

python3 instruments/mfgcap/run_recert.py t1      # the T1 enclosure, nine falsifiers
node instruments/critcount/run.js t1              # EXACTLY 2 maxima over the whole ball
python3 instruments/mfgcap/sigmastar.py           # sigma* = 1/(8pi^2), exact rationals
python3 instruments/mfgcap/bracket_table.py       # the bracket table + threshold pin
node labs/mfg/census.js --N 5 --c -12             # EXACTLY 3 (6.95M boxes)
node labs/mfg/multiplicity.js                     # >= 3 solutions per coupling
make test                                         # every battery, every red control

The write-up: paper/terra-peaks.md · PDF: paper/terra-peaks.pdf — generated from the same certificates as this page. Refutations and independent re-runs are invited: carlos@carlostoledo.co.