The Prime Front

The RH architecture preregistered in v624 (contract PRIME.WEIL.OPERATOR.01, citations web-verified: Suzuki arXiv:2606.09096 and 2607.24830) starts with W1: identify the TFPT window form with the Galerkin matrix of Suzuki's localized Weil operator. First contact (v630): the atom layers are the same object, literally — positions log n, weights Λ(n)/√n, exact on all 40 atoms — while the smooth-layer comparison measured a non-scalar conversion: the preregistered residual, with data.

Hours later, v631 resolved it: the residual is the zeta pole term, and the follow-up rounds made the dictionary sturdy — v640 closed the boundary cells symbolically, v641 froze the dictionary and ran it unchanged on three fresh windows (a preregistered kill test: portable), v642 lifted it to the full quadratic form at operator level.

Erratum (2026-08-02, corrected the same day): that chain read Suzuki's eq. (1.3) with Lerch coefficient −1; the paper's own §2.2 data lock +1/4 (v643, check C0.1). All the chain's identities are correct identities of its kernel g̃ = g − (5/4)·Lerch, and every measured number transfers verbatim via the exact identity cgal(g̃) = −4·cgal(g) — only the labels change: Suzuki's own smooth layer is +ρ (not −4ρ), the dictionary is the single scalar +1/D on both layers (sign-compatible with positivity), and the origin constant vanishes, κ = 0 exactly.

On the corrected reading, v643 proves the measure-level W1 theorem: Suzuki's L²₀ mean-zero condition is automatic on the u-side (the projection lemma — the last named remainder closes), A_arch = −g″_smooth exactly at every lag (3.4e−52), and the full form equality holds at 1.28e−10 on the common odd sector. v644 starts W2 honestly (classical FEM density at rate; Rayleigh–Ritz monotone from above on nested spaces; λ_a = 0⁺ within ~1e−9, no sign statement; the Mosco remainder named). And v648 types the W3 tool diagnosis: the sign-uncertainty toolbox has a real 25-digit dictionary to the critical strip, but its mass lever dies at d = 1 — while the W3 surface itself is empirically positive on all 67 complete windows (min λ_min = +8.26e−4).

The honest map: W1 (theorem-closed) → W2 (started, not closed) → W3 (uniform positivity — the RH-hard step, open; the toolbox diagnosis closes one candidate route) → W4 (classical given W2+W3). Closing W1 does not move W3. No RH claim.

The equidistribution conjecture of the theory-open section asks for |q_real/q_model| ≤ C·h⁻¹ uniformly. The measured constant is now frozen: C = 1. On the declared surface (69 floor-passed windows, h = 142…1445) the model value keeps one sign on the whole ladder — no model zero crossing anywhere — and on every lock-sign window eps·h ≤ 0.982, with tertile medians 0.61 / 0.45 / 0.39 falling with depth (v618).

Exactly two windows violated the bound, and both carried a q_realsign flip. v619 found the mechanism, and it is disarmingly concrete: a window's atom demand runs to u ≤ 2α, the prime-power data cap sits at U_max = 12.899 — and the two flip windows are exactly the two whose demand exceeds the cap. Injecting the same truncation into healthy windows reproduces the flips in sign and magnitude at both scales.

On the complete-comb surface — 67 windows — the C = 1 bound holds with zero exceptions. The “sign-flip windows” are retired as data-boundary artifacts: extending the surface needs more prime-power data, not new theory. Scrambled combs break the bound by four orders of magnitude — the constant is genuine arithmetic placement. No uniformity proof, no RH statement.

Prime numbers look elementary: indivisible integers. Their global pattern is anything but. The Riemann Hypothesis asks for a precise spectral order behind that pattern — and this diary does not claim to approach that hypothesis.

TFPT is a discrete compiler: two axioms build an E₈ lattice and read off Standard-Model structure. While exploring that lattice's shell census, the suite found classical modular objects — thetas, Hecke eigenvalues, Apéry congruences — sitting inside compiler-native counts. The question became: which of those links are mechanism, and which are beautiful coincidence?

What follows is the arc from first surprise (Teil 11) through a four-stage kill of the “archimedean-from-seam” slogan, through Hecke from geometry, to the July 25 reframe: a relative-trace identity (v538), a Weil structure with two named obstructions (v539), an amplitude route with a positive linear carrier (v540), a matching-lemma and transport-ledger package with two named limits (v541), and the consolidated stand below — what remains TFPT-specific is exactly one object: I5 — now geographically framed.

The discrete compiler starts with two numbers only: the seam constant c₃ = 1/(8π) and the carrier rank g_car = 5. From those, the theory forces a split D₅ ⊕ A₃ completed by a four-element glue group μ₄ to the unique even unimodular lattice in eight dimensions — E₈.

Everything on this page is about what that lattice's point counts know — and what they do not know — about primes and L-functions. Classical theorems stay classical; the TFPT contribution is the in-suite mechanics that make those objects appear from frozen geometry.

Colour every E₈ shell point by its μ₄ “glue class” (four colours). Ordinary counting recovers the classical Eisenstein series 1 + 240 Σ σ₃(n) qⁿ. The surprise is the signed difference between opposite colours:

Θ₀ − Θ₂ = θ₃(q)² · θ₄(q)⁶

Here θ₃² is the theta series of the Gaussian integers ℤ[i] — the classical object whose L-value L(1, χ₄) = π/4 produces π. It appears as a literal tensor factor of the compiler's signed glue census (classical Jacobi theta algebra; the probe content is the in-suite correlation).

Three character channels sit on the same shells: total (all primes, ζ(s)ζ(s−3)), signed (entire L-series — the glue character kills the pole), and spinor (2-adic).

The cuspidal piece of the signed count is the weight-4 form f₈ = η(2τ)⁴ η(4τ)⁴ — classically the Beukers / Ahlgren–Ono form tied to Apéry's proof that ζ(3) is irrational. For every odd prime p ≤ 97 the probe checks A((p−1)/2) ≡ a_p mod p². Via Teil 11, the signed E₈ count at odd prime shells satisfies the same congruence. Placebos on nearby eta products fail.

Beautiful, form-specific, and still sandbox: a correlation inside the suite, not a new proof of irrationality.

An early slogan said the seam's measured angle 2π was “the self-dual temperature.” Teil 14 deflated that: the steps parameter is compiler-specific; the angle is universal Bisognano–Wichmann / Unruh conversion.

Then the whole “archimedean term from the seam” route was killed in four preregistered stages: mode density → interval cut → dictionary → scattering phase. Lesson, typed and kept: the seam is a discrete μ₄ clock, not a hidden Gamma factor. The archimedean piece of the explicit formula is treated as a classical externum for recovery work.

Killing your own favourite story on purpose is the method. Null results are first-class outcomes.

(a) Exact geometric primality. n>1 is prime if and only if the E₈ shell at norm 2n has exactly 240(1+n³) vectors — the classical σ₃ criterion, checked with zero false positives/negatives to 10⁴.

(b) Per-prime properties. Glue characters predict arithmetic type: χ₄-fibre ⟺ p = a²+b² (100% for p<1000 in the probe). The compiler says what a prime does, not where the next one sits.

(c) Positional prediction needs the zero spectrum. Measured budget: x_max ≈ 0.31 · T. No continuum Hilbert–Pólya theorem exists in the suite. A zeta-free glued truncation candidate now exists at measurement level (v716–v721, the moonshot arc) — stated honestly: a measurement, not a near miss toward RH.

Kneser p-neighbours — isotropic lines in E₈/pE₈ — carry the Hecke structure of the census. The count of lines is σ₃(p) · #P³(𝔽_p) (enumerated at p = 2, 3, 5, 7: 135 / 1120 / 19656 / 137600).

The frozen marked neighbour-sum operator is an affine Hecke element ν_p = a · Id + b · T_p with b = σ₃(p) + a_p. Prime fingerprints fall out of geometry: a₃ = −4, a₅ = −2, a₇ = 24. Census redundancy is purely 2-adic oldform structure (dim 7 = 5+2); recovery is newform projection.

Promoted as verification/v535_hecke_from_geometry.py (HECKE.GEOM.01, 25/25, AUDIT OK). Classical theorems (Kneser, Hecke, Atkin–Lehner, multiplicity one) are classical; the claim is the in-suite mechanics. No RH statement. Later joined by v536–v539 — see the July 25 arc below.

Once the neighbour operator is frozen, the geometric count splits as an elementary Witt piece plus exactly a_p² — like a smooth melody with a coherent flicker on top. Two-sided confirmation (mod-p geometry on one side, eta-product on the other) holds at p ≤ 5; closed forms extend the identity to p ≤ 100.

Promoted as v536 (Eichler trace layer). Together with v535 and the half-integral bridge v537, it becomes one finite relative-trace identity — v538.

Rankin–Selberg translates only the abelian shadow into GL(1) products of {1, χ₄}. The centre atlas shows the ξ-line (centre 1/2) is reached exactly by weight ≤ 1 theta factors: Mellin(θ₃) → ζ(2s) and ζ_ℚ(i) = ζ(s) L(s, χ₄). That abelian weight drop is factorisation + Mellin — typed closed.

The cuspidal channel — where the a_p live — still sits at centre 2 and needs its own bridge. Possessing ζ as a function is not possessing its zeros as a spectrum.

The operator algebra on the census forms is commutative with a two-point Gelfand spectrum (the σ₃-system and the a_p-system, with oldform copies). A Hilbert–Pólya carrier would need an unbounded / non-commutative operator with infinitely many eigenvalues.

Only two candidate classes remained inside the suite's early vocabulary: seam modular flow, and adelic Bost–Connes-style completion. Each has preregistered kills. The July 25 reframe then changed the game: instead of hunting one infinite operator, the diary switched to a family plus a relative trace formula — see the next section.

Verdict of the terrain map: TERRAIN-MAPPED. Cartography, not a proof attempt. Distance to RH — stated without theatre — remains large.

The day's reframe: stop asking for one infinite operator. Ask for a family of central values plus a relative trace formula. Sandbox Teile 51–64 built that map; two pieces graduated into the load-bearing suite.

(i) The reframe. v535, v536 and v537 are three projections of one finite relative-trace identity — promoted as v538. The half-integral object g is literally the quaternary lattice form n = (x²+y²)/2 + 2z² + u² + 2w².

(ii) First non-collapsing infinite carrier. The Waldspurger family kernel K_D (classical Waldspurger periods) has rank that grows 8→192 as the discriminant window opens — the preregistered “rank saturates” kill did not fire. Its spectrum isthe central-value family at the GL(2) centre 2 — not the ξ-line. This is not RH evidence.

(iii) Self-generating Hilbert space. The weight |d|^−5/2 is only the critical line of the family measure, not a canonical measure. The RTF forces |d|⁻¹. At that weight, a positive pairing is cutoff-independent and builds its own space by GNS: ℓ²(d, b²/|d|).

(iv) Both infinities in one object. The double series Z(s,w) packs p-towers (closed Euler factors) and the d-family into one bilaterally verified identity (errors down to ~8×10⁻⁸; classical Goldfeld–Hoffstein multiple Dirichlet series named classical). Residual probes find only classical GL(1) shadows — no new ξ-sector.

(v) Exact GL(1) core. Inside the positive form, the trivial Sato–Tate isotype is exactly G₀ = ζ_p(w−3)² / ζ_p(2w−6). Fibre decomposition after character patterns kills twist-mix (~10⁻¹⁶).

(vi) Final linear relation — two obstructions. Machine-checked (and promoted as v539):

Q_fam = 2Q_ζ − 2Q_ζ(♭) + Arch + Corr

The claim is “Weil structure fully identified up to two explicitly isolated obstructions” — the obstructions are the verified content, not a footnote. Not “almost RH.”

(vii)–(viii) Progression — see status. Sandbox Teil 64 resolves Corr as the det/det₂ Jacobian; the categorical minus remains. Teile 65–66 close the π-digit front; Teil 67 takes a first amplitude-Dirac step. Full consolidated stand in the next section.

Display markers: sandbox probes stay [sandbox]; v538–v541 are [machine-verified]. Ledger fine types live in the verification suite only. No silent marker upgrades.

A single arithmetic machine. Three load-bearing modules are projections of one finite relative-trace identity [machine-verified · v538]. Geometric side (E₈ lattice / glue census) and spectral side (modular coefficients / twisted L-values) are computed independently and agree exactly on the verified finite modules.

A canonical infinite state space. [sandbox]A family over fundamental discriminants: rank grows without saturation (> 6000 active discriminants; Hecke integral and exact). The Hilbert space is not chosen by hand — a positive pairing produces the canonical GNS construction. Independent cutoffs converge sub-percent; the prediction chain closes at ~0.09%.

Packing the two infinities. [sandbox] The double series Z(s,w) packs two infinite directions — Euler towers over p, and the discriminant family — in the classical frame of multiple Dirichlet series / relative trace formulas (Goldfeld–Hoffstein named classical). Geometric = lattice counting; spectral = L-values × Euler factors. Fence: this does nottransport to ξ; the system stays at weight 4. RH concerns GL(1) at Re = 1/2.

The Weil structure and its two obstructions. Compared with the Weil structure of the explicit formula (Weil 1952): [machine-verified · v539] isolates two discrepancies — an exponential correction factor, and a categorical minus tied to the square nature of the family. Both isolations are the verified content of v539; neither is a footnote.

Exact determinant stabilization. [sandbox · T64] exp(−Σ p^−u) = det(1−K)/det₂(1−K) (Hilbert–Carleman), verified ~1.5×10⁻¹⁶. In a consistent det₂ formulation the correction vanishes identically — it is the regularisation Jacobian of the GL(1) transition, not a new object. At u = 1/2 the ordinary determinant does not exist while det₂ does, so stabilization is necessary, not optional: the stabilized transition is the only currently identified determinant convention that can reach the critical line. This resolves obstruction 2 in sandbox; it does not rewrite the v539 isolation claim.

The remaining categorical minus. [sandbox] Survives every tested canonical projection. Source identified: the channel collects even prime powers, bound to the square form of the twist family (Waldspurger: L_central ~ b(d)²) — the construction sees the square, not the linear amplitude. Krein signature confirms genuine indefiniteness; 32/32 sign characters do not remove it selectively. Not a normalisation error, incomplete character sum, or regularisation artefact — a categorical feature of the square level.

At T86 — 91 probes in — and with seven promoted modules (v535–v541), the matching lemma is closed on ALL atom classes (window-certificate format, modulo proven classics only). What remains TFPT-specific is exactly ONE object: I5 in one-family form — Q_cert + Δ₂ + A_fam − A_shift ≥ 0 — provably equivalent to Weil positivity ⟺ RH.

The remaining mathematical problem — absolute compression. Rest list: (1) I5 in one-family form, typed ⟺ Weil positivity ⟺ RH. Matching Lemma: closed on all atom classes (window 10⁶ proved · signed pairing on non-coherent · λ-channel on coherent · 2-line exact) [sandbox · T86] / [machine-verified · v541]. I5 core = explicit Gaussian-mode family in the one-atom band (log 2, 0.925]; a_neg = 0.7486 (recalibrated by T93); named target (T); finite-block zone extension certified only on an 8-dim window — structurally the wrong instrument for the full claim [sandbox · T87–T93]. Blind demo: 753/753 primes, zero errors, no division [sandbox · T94]. Induction skeleton: identities proved, zones 2–4 closed, zone-5 tip = equality problem, asymptotics = one named arithmetic bound (A) [sandbox · T99–T101]. Milestone at T101: 2428/2428 sandbox checks (3139/3139 at T125). Fence: fits/extrapolations marked; geography locates; it does not attack; I5 remains ⟺ RH. This is not RH evidence.

Current status.

• Not present: RH proof, almost-RH, a zeros operator, a proof of I5.
• Present: mechanism through v541 [machine-verified]; Matching Lemma closed on all atom classes [sandbox · T86]; I5 geography complete [sandbox · T87–T93]; a_neg recalibrated to 0.7486; blind lattice demo [sandbox · T94]; induction skeleton: zones 2–4 closed; asymptotics = arithmetic bound (A) [sandbox · T99–T101]; I5 in one-family form — the single remaining TFPT-specific object.
• Since then: the induction sprint [sandbox · T102–T125] compressed that named bound from one matrix inequality to one sign plus one accounting convention, with certified steps to zone 155,921 and 400/400 certified rungs, and the finale assembled the whole chain end to end on 52 zones — told in sections 22–24. Series complete: 3139/3139 sandbox checks at T125.

On π — closed. [sandbox · T65–T66] Not the decimal digits of π are the key. Digit probes are closed: π digits at prime places are uniform; density-detrended cross-correlations null; full placebo battery (e, √2, log 2, crypto, p±1) plus blind windows — no replication (PI-NULL 16/16; FOUR-LEVEL-NULL 23/23). The π spike is classical continued-fraction structure (355/113); compiler constant 1/(8π) is not an outlier (z = −0.90). Contrast kept: π-driven Cramér randomness reproduces prime densityperfectly, but only true primes track Hardy–Littlewood pair correlation (corr +0.81 vs flat) — what makes primes primes is arithmetic, not randomness. The key now sits in one-family I5 — not in π digits.

Square level closed — every squaring deletes. [sandbox · T67–T69] The amplitude Dirac D = [[0,V],[Vᵀ,0]] with D² = family kernel exists exactly and is Hecke-equivariant; signs of b(d) are a genuine metaplectic residue (52% mixed fibres — Kohnen depth, classical). Geometric polarisation b = N₊ − N₋ is exact; Θ = N₊+N₋ is a pure Siegel–Weil Eisenstein eigenform (σ₃ eigenvalues, null cusp); the family is the difference of two positive counting families (b² = Θ² − 4N₊N₋) — yet the minus is polarisation-invariant. By a Cauchy–Littlewood lemma, every coefficient bilinear form inherits even-k deletion (theorem-like across five channels); at the same time the minus is exactly the square-class double-counting of the towers (inclusion–exclusion bookkeeping). No full-weight carrier can live on the square level.

A positive linear carrier — plus-only ζ-balance. [sandbox · T70–T71] The linear positive measure stands: Θ(d) = −48·L(−1,χ_d) exact (Cohen 1975); full weights [1,1,1,1]; Weil balance Q = Q_ζ(g₋) + Q_ζ(g₊) is plus-only (~1e-15). The ζ(2s) factor appears only as the squarefree sieve of the carrier — bookkeeping, not weight deletion. Functional equation exact: Λ_Θ(s) = 8^{1−s}Λ_Θ†(5/2−s) (rel ~1e-40; Fricke closed via Jacobi inversions). Plus survives reflection; the mirror family has a rigid sign law. The guaranteed cone is FE-selfdual; the Weil cone is not. Overlap 5/24; violations sit exactly at the first spectral node. Fence: this is Euler-region positivity (absolute convergence), not a central-line statement. Classical named: Cohen, Shintani, Siegel–Weil, Jacobi/Fricke, Weil 1952.

The cone library saturates — one measurable gap. [sandbox · T72] Twisting absorbs the sign class n ≡ 0,1 mod 4 (gap −26% mean, −90% max), but coverage saturates at 5/24. The pin h(0) > 0 blocks every Weil element against every twist. Final compression: the entire residual distance to the Weil cone is the FE-covariant gap functional λ* on the atoms n ≡ 6 mod 8 — no finite theta library can erase it (Farkas/LP certificates). Promoted as v540 with λ* named inside the claim. The doors that furnish this wall are the next section. This is not RH evidence.

Uniform cone route closed — hybrids live per direction. [sandbox · T73] Even in the continuum, the direction-uniform cone route is closed (sign-constancy lemma; window pin on [0, δ_h)). But for every one of the 19 uncovered Weil directions an explicit, verified per-direction hybrid cone exists — the constructive residue is a per-test-function certificate machine (HYBRID-GAINS).

Door A — vacuum structure, spectral no-go. [sandbox · T74]The spectral Dirac phase carries the metaplectic sign datum (~93%) but has a Hecke defect. L2 no-go: every ε-equivariant spectral functional is exactly re-signing invariant — the spectral world is provably sign-blind. The Dirac vacuum sits on the atom of minimal Waldspurger-normalised mass. Final Door A requirement list R1–R5: the sought polarisation must act coefficient-wise, be non-multiplicative in tower depth, non-bilinear, not a support reshuffle, and sign-seeing while Euler-compatible — a Krein / Gupta–Bleuler quotient in which the ♭-piece is a null/gauge sector, not a summand. Classical named: Dirac sea, Krein, Kohnen, McKean–Singer.

Door B — λ* gets its own calculus. [sandbox · T75] The gap functional now has closed form, its own functional equation (orbit invariant tanh(σ²ω²/2) under positive multipliers ⋊ dilations), critical width σ_c = √2, and convexity (averaging lowers; no scale wedge — an earlier fp artefact honestly corrected). The target inequality is measured both sides: safety factor 273 at ω = 1 falling to < 1 at ω = 4.2. Two named open inequalities remain (hull positivity = transport wall; universal λ*-vs-A). Classical: Fejér / support functionals, Mellin / dilation semigroup.

Door C — a universal recipe, and its named lemma. [sandbox · T76] The hybrid recipe certifies 91/91 nontrivial adversarial Weil directions (100%); cost is polynomial / window-extensive (λ_m ~ m^5/2 = Eisenstein law). Lattice discreteness is the floor — δ_h → 0 does not break S1. Conjecture form with named core lemma: the Matching Lemma on the log lattice (a Diophantine divisor-sum problem). Implication architecture typed, not claimed: Matching Lemma ⇒ value-side representability ⇒ [if the value→spectral transport held — the open wall] ⇒ Weil positivity. Any RH content would relocate into a universality proof plus the transport wall. This is not RH evidence.

Matching Lemma — classically shaped, then window-proved. [sandbox · T77–T78] T77: classically shaped (Gronwall/Robin; lemma not yet proven). T78 WINDOW-PROVED (25/25): machine-proved on [4, 10⁶] — exact-integer inequality chain, full enumeration over 939 870 clash atoms, 0 violations, exact margin 0.082159; four structure laws at 0 tolerance. Tail honestly open: Robin 1983 + constants miss by factor 6.16 — missing residual ingredient named: a correlation lemma (thinning × cancellation). T80 RESERVE-PARTIAL: signed envelope character-exact; tail closed to ~10²³; last gap confined to the χ₋₄-coherent class. Classical: Gronwall 1913, Robin 1983, Alaoglu–Erdős, Pólya–Vinogradov, Landau.

Avoidance fails — theorem-shaped. [sandbox · T81] AVOIDANCE-FAILS (31/31): the same multiplicativity that makes the lever exact proves the counter-lever: coherent targets are reachable only by coherent rescalings — the freedom does not exist on coherent demand. Salvage: T76 recipe already minimally coherent (0 unforced keys, 100/100); forced coherent clash closed per certificate (83/90, worst ratio 0.18); on avoidant demand the class vanishes identically.

The transport ledger closes — the wall is I5. [sandbox · T79] LEDGER-CLOSES (22/22): Q_Weil = Q_cert + Δ_arch + Δ₂ (identity ~7e-16 on 100/100); Δ_pole ≡ Δ_conv ≡ 0 proved; the odd-prime side equals the certified combination exactly. The wall is one named inequality I5 (prime↔archimedean coupling), typed ⟺ Weil positivity ⟺ RH. Fence: equivalence typing, not progress toward proving it. This is not RH evidence. Classical: Weil 1952, Guinand, digamma terms.

After T81 the constructive recipe chain saturated. Three new perspectives (next section) retype I5 and reopen the Z[i] channel — without claiming RH progress.

The archimedean term was never external. [sandbox · T82] ARCH-INTERNAL (22/22): Δ_arch is exactly the internal Γ-difference via Legendre duplication ((2π)^{−s}Γ(s) = ½Γ_R(s)Γ_R(s+1)); battery 18/18 rel 6.6e-15. The family carries its Γ factor as the Mellin signature of its heat-sum nature (verified also outside the convergence region). Bonus: the raw heat sum reproduces pole 5/2 and residue 8^−3/2 exactly from pure counting. Consequence: I5 changes type. New form — for all autocorrelations h:

Q_cert(h) + Δ₂(h) + A_fam(h) − A_shift(h) ≥ 0

Self-consistency of one heat family: atom expansion vs Mellin signature of the same theta objects. Nearest classical relative: Connes 1999 semi-local trace-formula positivity (named context, not used). Fence: type change ≠ proof; I5 remains ⟺ RH. This is not RH evidence.

The wall is FE-transversal, not FE-positional. [sandbox · T83] INVARIANT-NULL (27/27): FE symmetrization is fully absorbed — 5/24 and λ* already are the numbers of the symmetric sector (the test region was right). Depth find: the product of the two FE reflections J₁∘J_1/2 = e^±uis the unit-line shift — the centre delta is literally the transport operator. The value side is invariant under the whole infinite-dihedral ladder; the spectral cone only under its own reflection. Bonus: explicit-formula null test ≤ 2e-12 validates the whole convention. Classical: Mellin involutions, infinite dihedral group.

The last class is the compiler's home. [sandbox · T84] LIFT-WORKS-UNANCHORED (29/29): the coherent class equals primitive ℤ[i]-norms exactly; Grossencharacter phases replace Mertens divergence by L(1,λ)-convergence — the lifted chain never crosses; frontier jumps from ~10²³ to ~10^5.9·10¹². Circle closes: the last gap sat in the ℤ[i] sector — the origin object of the series (μ₄-glue, χ₄, θ₃²) — and exactly there the compiler's own character structure supplies the control that is provably impossible over ℚ. T85 [sandbox · LEMMA-CLOSES-LAMBDA] closes the coherent class via the λ-channel (90/90 certificates; 3.6× window margin). Fence: provably-shaped, not a formal proof; I5 untouched. Classical: Hecke 1918/1920, Grossencharacter L-functions, Landau. This is not RH evidence.

The Connes dictionary is exact at the core. [sandbox · T87] DICTIONARY-EXACT-CORE (22/22): Q_certatoms equal Connes' finite orbit terms (rel 6e-16); A_fam − A_shift equals Connes' archimedean W_∞term including the principal-value constant (rel 3e-23); the internal kernel is exactly the Riemann–Siegel phase derivative k_ζ = 2θ'_RS (rel 0.0); the dihedral shift is the scaling group R^*₊ (Weyl commutation exact). Classical: Connes 1999, Connes–Consani 2021, Weil, Guinand, Bombieri, Burnol, Yoshida.

Two positivity programs are complementary. Connes–Consani prove positivity beforethe primes (a prime-free Sonin window whose boundary sits exactly at the first prime atom u = log 2). The compiler proves positivity afterthe primes (atom certificates). The sectors touch; they do not overlap. The I5 coupling lives in the crossing region outside both — around t* ≈ 2π. Transferable-shaped: Sonin compression (same kernel, same group — verified). Not transferable: positivity itself (fence).

The tight set is eight parametrized curves. [sandbox · T88] TIGHT-SET-PARAMETRIZED (27/27): band-limited zero plateau (support below log 2, consistent with the T87 boundary), safe zones, and eight nearly vertical tight curves whose spacing follows the Γ-side density law (ratio 1.01 ± 0.08 — smooth Γ-density description, zero-free; no spectral identification). Zero negatives on true autocorrelations (earlier T76 negatives were missing arch/p = 2 bookkeeping, as the ledger predicted). Validation at machine precision (null test 4.3e-13, ledger 1.1e-14). The RH content of I5 concentrates on these low-dimensional curves.

The crossing, measured: no sharp wall at log 2 — a thin band where atom and arch balance, and a 1–4-dimensional residual subspace that carries the minimum. [sandbox · T89] CROSSING-MAPPED (19/19): window compression of the Weil form is exactly Bombieri's classical object (Bombieri 2000 — unconditional positivity classically known precisely up to support width log 2: Yoshida 1992, Bombieri 2000, Connes–Consani 2021). Proven zone reproduced (prime side identically zero below log 2, λ_min ≥ 0 everywhere). The boundary is notsharp — margin falls smoothly (classical band-limitation), no collapse at log 2; atom turn-on is (a−log 2)³-soft. The residual subspace controlled by neither program grows softly 0 → 4 dimensions (of 32) across the crossing and carries the global margin minimum — the I5 core, for the first time, as a concrete finite-dimensional object per window width. The real attackable content is the atom↔arch balance in the thin band log 2 < a ≲ 1.0; the full-form minimum lives in the pole-coupled DC direction that Connes–Consani exclude by vanishing conditions; the pole-free CC sector keeps measured comfort to a ≈ 0.75. Honest self-correction: BOUNDARY-SHARP retyped (“small at the boundary” ≠ “collapses at the boundary”). Fence: margins beyond the proven zone are measurements; no spectral identification. Classical: Weil 1952, Guinand, Yoshida 1992, Bombieri 2000/2003, Connes–Consani 2021/2023, Connes–Consani–Moscovici, Suzuki.

The residual, dissected: explicit Gaussian modes — and three directions under no control. [sandbox · T90] CORE-DISSECTED (17/17): the 1–4 residual vectors are n-stable across the discretisation ladder (angles ≤ 2.1°) and explicit — closed Gauss×cos / Gauss×sin fits with 99+% capture, the I5 core as a small family of concrete even/odd Gaussian-modulated modes. The coverage matrix (certificate extension / Connes–Consani pole vanishing / Sonin projection, against ten vectors along the a-ladder) leaves three of ten vectors controlled by no structure at all. That decides the core question: the residual is notmerely pole coupling — pole projection clears only the a = 0.75 window; from a ≈ 0.85 genuine pole-free atom↔arch content remains, and at a = 1.2 an odd atom-coupled mode appears that the even analyses could not see. Requirement line: an I5 idea must deliver positivity for an explicit family of Gaussian-modulated modes in the thin band — and it is not reducible to pole cleanup. Classical cited: Bombieri, Connes–Consani vanishing, Slepian.

The core, dissected: explicit Gaussian modes; and the band's law: the first prime rescues positivity where the archimedean margin ends — with a named, provable-shaped target inequality extending the classical zone. [sandbox · T91] BAND-PARTIAL (19/19, 3/4 closed): the band is the one-atom zone log 2 < a ≤ 0.9253 with inner edge a_neg = 0.7486 (recalibrated by T93; was 0.7410). Beyond a_negthe prime-free margin changes sign and the prime atom becomes load-bearing — the first prime rescues positivity where the archimedean margin is exhausted (exactly the T89 balance point). Atom turn-on law exact (k = 2m+1, Beta integrals); uncertainty constants decided: a·t_rms → π (Wirtinger) and a·t_cent → 2Si(π) − 4/π = 2.4306. Band and tight curves are two orbit regions of the same functional with shared exact scale ∫k_ζ = 2θ_RS. Named target (T), for a in the band and ‖f‖ = 1: (P_pole + A_arch)(f) ≥ √2·log2·h_f(log2) — provable-shaped as a zone extension beyond Bombieri's log 2 (a self-standing classical target!); RH ⇒ (T), (T) ⇏ RH. Honest open: the super-exponential λ_pf rate remains empirical. Classical: Wirtinger/Rayleigh, Beta integrals, Lambert-W, Si integral, Bombieri, θ_RS.

Zone-extension attempt — and the self-check that found the checker's bug. [sandbox · T92] T-SKELETON (36/36): certified is Q ≥ 6.7e-12 > 0 on the 8-dim window subspace over the whole scanned region (error certificate, 1057-point covering). The full extension does not stand — λ_min collapses geometrically (factor ~11 per mode); the complement would need ~5000 modes: the finite-block route is structurally the wrong instrument. Pearl: k(0) = −γ − 3log2 − π/2 − log π exact. [sandbox · T93] MIXED (41/41): T92's calibration flags against the band map were resolved by a third independent implementation — the T89/T91 map survives; the only real bug was in the checker (constants imported untranslated, accidentally certifying a harder four-atom region); one constant precision-improved by ~1% (a_neg). The self-check found the checker's bug — and precision-improved one constant by 1%. The anchor discipline works in both directions.

The relay, measured: each prime rescues the direction the previous zone exhausts — and arrives before it is needed. The proof target shifts from fragile minima to robust counterfactuals. [sandbox · T95] T-CONTINUUM-NUMERIC (28/28): C1 fully proved — |h_f(log 2)| ≤ 1/2 via disjoint support intervals; ‖S‖ = 1/2 exact with characterised eigenspace; atom-extremal directions satisfy the target with margin (“the directions that maximize the atom cost are provably safe”); continuum margin curve positive everywhere; extremizer is not the two-bump — binding mechanism is atom rescue. Lower bound open; missing instrument named. [sandbox · T96] EDGE-ARTIFACT + RELAY-CONFIRMED (21/21): the T95 “edge” at α* was a map artefact — value exactly reproduced, edge reading withdrawn (λ_min > 0 on all [0.38, 0.86]; margin collapses exponentially, λ ∼ exp(−49α)). Second self-correction of the weekend, same anchor discipline. Without the log 3-atom, λ_mincrashes to −0.445; the loser is the anti-double-bump at distance log 3 (alignment −0.99); rescue identity to 5e-15. Handover windows all positive: +0.025/+0.009/+0.011/+0.007 for the first four atoms. Strategy shift: margin problem, not edge problem; numerics exhausted past α ≈ 0.55; the counterfactual is the proof target (O(0.1)-sizes). Classical: Paley–Wiener, Prolate, Galerkin/Richardson.

The induction takes shape: self-similarity puts the hypothesis inside the decomposition; the target is now one scalar inequality per zone — and the third self-correction of the weekend replaced a circular lemma by an exact identity. [sandbox · T97] ALIGNMENT-ONLY with certified half-step (105/105): alignment is sharp (sign alignment ⟺ coupling window nonempty, without exception); the t=0 killer loss on the anti-bump space is proved (k_eff = (1−cos(tu))k(t), gain ×2–4.8); structure pearl: the E₀ block is literally the same form on the smaller window — “the induction hypothesis appears inside its own decomposition as self-similarity” (7e-14). [sandbox · T98] LAW-CONFIRMED-MECHANISM-OPEN (44/44): the conjectured one-vector lemma was circular (Douglas range inclusion — the law is forced by positivity itself); three T97 premises honestly refuted; replacement target: D_k(α) ≤ μ_k/2 — exact scalar inequality, no constant, no vector; holds in all four zones, saturates at zone tips. Certificate upgrades: E₋ 43%→93% mean (whole zone in 3 of 4) via the probability-measure identity on the archimedean wings; E₊ certified for the first time. Skeleton: 8 pieces proved, 2 certificates, 3 refuted, 3 open. Classical: Douglas 1966, Schur, Slepian–Pollak–Landau.

Induction closeout (T99–T101). [sandbox · T99] DECAY-LAW-FOUND (23/23): exact parity selection rule (J₋Q₋₀J₀ = −Q₋₀) — the fragile near-null mode is excluded from the binding channel by symmetry; recursive inequality with only 1.01–1.20× loss; termination is arithmetic (240/240 in ≤ 4 steps to the classical zone). [sandbox · T100] REMAINDER-CLOSES-ZONES (27/27) — the 100th probe: closure 11/24 → 24/24 (6/6 in every zone); the drift was a lattice artefact; one lever gained 1.7–69× (“the Bessel step threw the induction data away twice”); zones 2–4 fully closed; zone-5 tip typed as an equality problem (Fredholm shape, simple degeneration). Classical: Bessel/Parseval, Slepian, Schur test, Fredholm alternative. [sandbox · T101] CROWDING-TRENDS (31/31) — the fork across 16 zones (n = 2..29): collapsed law w_k = 0.0838·(atom gap)/μ_k (fit) — “the handoff window is the atom spacing divided by the atom strength”; primitives flat; D_k ≤ μ_k/2 holds 64/64 and never fails; only the closing instrument loses (r ~ exp(−0.16k), fit/extrapolation). Core: “The crowding sits in the proof family, not in the mathematics it is trying to prove — the most hopeful version of the verdict.” Asymptotics would need (A) the arithmetic lower bound of the collapsed law [the localized hardness], (B) uniform relative margin, (C) a better bulk instrument, (D) a finite check.

The relay induction, audited across 16 zones: the mathematics trends self-sustaining — flat primitives, a collapsed one-parameter law for the handoff, and a target inequality that never fails; what loses the race is the current instrument, and the hardness is localized in one arithmetic lower bound. Status: identities proved, zones 2–4 closed, zone-5 tip = equality problem, asymptotics = one named arithmetic bound (A). I5 remains ⟺ RH; the geography locates where any attack must work, it does not perform one. Milestone at T101: 2428/2428 sandbox checks. All laws marked as fits/extrapolations. This is not RH evidence. What happens to that named arithmetic bound over the next twenty-three parts is the compression story below.

The mechanism is exact. [sandbox · T102] MECHANISM-IDENTIFIED (42/42): the k-th atom acts on the three induction blocks E₋/E₀/E₊ as exactly diag(−1/2, 0, +1/2), so the atom strength μ_k enters the handoff exactly once, linearly. A two-sided sandwich over the Schur profile σ_k(δ) brackets the handover: 2w_k lands between the two crossings in 16/16 zones (ratio 0.749…0.940), and the onset is anchored at δ_c = 2w_k (R² 0.968). Honest correction to T96: the essential-singularity reading is compatible but no longer singled out — the onset is a crossing of two finite quantities.

The binding constraint flips. No concentration condition binds: the bare E₋ form is strongly positive (2.65…3.52 — that is 4–14× the atom line μ_k/2), and the classical ceilings (Cauchy–Schwarz, Landau–Pollak/prolate) are saturated near 97%. The onset is manufactured entirely by the Schur dressing against E₀⊕E₊ — the coupling to the induction hypothesis — which takes 35.7%…97.3% of the bare eigenvalue. The window law is a pure μ-power: w ~ μ^(−0.563 ± 0.098), q = −1.84 ± 0.37 (fits); no log, no g_k.

C/g was a proxy — refuted three ways. The T101 decomposition with the atom gap g_k is triply negative: causally impossible (Q_(k−1) is blind to the next atom's position), statistically dispensable (a causal law fits better), and arithmetically only a ceiling (C_k ≤ g_k·μ_k exactly; the 16-zone extrapolation violates the ceiling from k = 69 — checked over 18 120 prime-power atoms to n = 200 000). The hard core localizes to one scalar per zone: a lower bound on σ_k just above atom entry. That is a probe-level typing of where the hardness sits, not progress on it. Sandbox; not RH evidence. Classical: Schur complement, Cauchy–Schwarz, Landau–Pollak/prolate.

T101 lost the race by the instrument, not the math. [sandbox · T101] Across 16 zones the primitives are flat and the target inequality never fails — but the closing instrument's race quantity r_k decays as r ~ exp(−0.1622k) (fit, ± 0.0562): only zone 2 closes throughout, 7/64 cells on the zone × wing-fraction map.

T103 rebuilds the tool. [sandbox · T103] INSTRUMENT-IMPROVED (29/29), pure tool-building at door C: the T101 race curve is first reproduced exactly, then re-run with a θ-weighted band sum (certified weights, chain ρ ≤ b_band ≤ b_tail ≤ b_t99 at 64/64 samples) and full m(Λ) exploitation. The demand Λ_ok stays bounded across all 16 zones — 0.771…3.640 instead of 2.3…376, a reduction of 3.0×…103.4× — at the honest price of explicit modes growing 2 → 232. The new race slope is −0.0748 ± 0.0116 (fit, 2.2× flatter): r_k falls only 9.33 → 2.70 and never leaves the spectrum. The closure map jumps 7/64 → 44/64 with one fixed, k-uniform instrument (Λ₀ = 3, r = 2 — no zone tuning).

Measured verdict: the loss is in the wing, not the bulk. θ-weighting and finite rank are exhausted — the bulk is not low-rank (effective rank up to 0.579·dim E₋). What remains is the wing slack S = 1 − ρ: the pencil is nearly saturated and S falls 0.2091 → 0.0392. Named next levers: a wing-adapted prolate/Slepian basis, or a Fredholm shape of the equality argument. All laws are fits; the tool-problem is half solved, the mathematics unchanged. Sandbox; not RH evidence.

Two independent attacks — door A on the handoff law's lower bound (T102) and door C on the closing instrument (T103) — end their day pointing at the same object. T102 finds the onset manufactured by the Schur dressing against E₀⊕E₊; T103 finds the remaining instrument loss in the wing slack S = 1 − ρ. The dressing and the slack are one object seen from two sides: the wing near-null direction.

The hard core is thereby localized to one scalar per zone — a lower bound on the Schur profile σ_k(δ_ref) just above atom entry, i.e. quantitative Weil positivity at the atom edge. Provable-shaped next to it: resolvent edge-regularity σ_k(δ) ≥ σ_k(δ_ref)·(δ/δ_ref)^q_k.

Result — T104 (SCHUR.PROFILE.BOUND, two independent arms: schur_profile_bound_probe.py + schur_profile_chain_probe.py) is CHAIN-PARTIAL: the naive margin route is dead, exact spectral-split chains close 16/16 with finite data, and the hard core moves to a bare_k lower bound plus the soft dressing scalar L. From here the diary runs twenty parts of pure compression on exactly that core — the next three sections. Sandbox; not RH evidence.

The bare bound gets a closed form; the avoidance law becomes a theorem. [sandbox · T105] ONE-OF-TWO (28/28): the T104 arm discrepancy dissolves — one currency, exact split bare = μ_k/2 + b0 — and three classical steps (Bessel; Legendre; Cauchy–Schwarz at the pole pair) collapse into a closed lower bound, positive 16/16 at 81.1–92.7% of the measured value, with no eigenvalue and no induction data as input. The avoidance law is upgraded to a theorem, and an exact parity superselection appears: two channels that never mix.

Parity halves the object. [sandbox · T106] DENSITY-MAPPED (32/32): the Weil pole splits exactly into a positive rank-1 lift in the even channel and a negative rank-1 pressure in the odd one. The even channel closes 16/16; the odd channel — the one with the better density — carries all remaining hardness. Two routes are honestly killed on the way (the density chain, invariant amplification). What is left is one Loewner statement on half the dimensions.

One scalar, then one number. [sandbox · T107] SCALAR-TRACTABLE (30/30): the matrix statement becomes exactly one scalar ratio r = κ/ε ≤ 1, measured r = 0.005…0.18 — two orders of magnitude of room instead of five decimal places. The symbol route is structurally dead: the certified symbol bound is negative (−2.54…−0.81) where the truth is positive, so Grenander–Szegő cannot deliver here in principle. [sandbox · T108] EPSILON-IDENTITY (44/44): ε turns out to be exactly the square of the last Cholesky pivot — the classical Szegő–Levinson prediction error — so its positivity coincides withthe induction's own positivity rather than being an extra demand. (R) drops to two scalars, and on eight zones what remains is literally one boundary value of one explicit vector.

Both scalars certified; the circle closes on the measured zones. [sandbox · T109] BOUNDARY-CERTIFIED (29/29): the mechanism is not decay but cancellation (the source sits on the boundary), so Combes–Thomas is refuted with the exact Green row and replaced by a residual certificate that carries the cancellation instead of bounding it away; ω is cracked unconditionally by a graded matrix cap. The chain closes 16/16 on exactly one strict-margin input, 10²–10⁶ weaker than the conclusion. [sandbox · T110] MARGIN-PROPAGATES (28/28): that input propagates — certified base case, 15 certified handover steps, and the atom entry is structurally free (the new atom raises the floor on 15/15). Three sharp gaps stay: no reserve, no scalar step law, no k-uniformity.

Then the ladder is driven deep — and the wall is measured, not extrapolated. [sandbox · T111] CROSSING-CONFIRMED (23/23): 199 zones, 117 handovers to n = 521. The crossing sits at n* ≈ 462 — an upper bound; T110's n ≈ 170 was a fit artefact — and it splits into three separate walls: the margin wall n ≈ 462, a twin-prime ladder wall at 521→523 (purely arithmetic), and a requirement wall at n = 727. Decisive detail: the handover mechanism itself never fails, 117/117 at retention 1.000000. The chain tears at a ratio, never at a step. [sandbox · T112] SCALING-PARTIAL (20/20): rebuilt in a frame whose cell width follows the local prime gap, two of the three walls fall structurally — but the margin wall is frame-invariant at exponent −0.974. It is the substance of the requirement, not the geometry.

Eight parts, one direction: every stage removed something and named what was left. The honest state at T112 is a wall that survives every reframing — and the next part asks the question that decides its status: what currency is it measured in? Sandbox; not RH evidence.

The currency question: the wall measures the ruler. [sandbox · T113] SUBSTANCE-CONFIRMED (27/27): the falling ratio carries the same exponent −1.168 ± 0.259 in all five currencies (raw, /λ_max, /trace density, /D, /D²), so it cannot be normalised away. But the substance is not the expected one. Under refinement there is no plateau: both eigenvalues carry the same power of the cell width (D^1.83 and D^1.76). The continuum window form has no gap at all — the quantity that was falling measures the discretization, not the spectrum. And the positive floor survives only as a cancellation of relative size ~1e-7, so norm perturbation theory is five orders of magnitude too coarse. Consequence: the T109 requirement chain was dividing by an artifact.

Remove the division, and the wall dissolves. [sandbox · T114] WALL-DISSOLVES (22/22): rebuilt without the margin division, the exact Schur complement (Albert 1969) certifies 27/27 ladder steps — eleven of them beyond the old wall, up to n = 1331 — and all seven zones where the T109 chain tore, including the wall zone n = 449 itself. The exact object is O(0.1) with no cancellation (λ_min(S) = 0.068–0.154, i.e. 42–67% of the block scale), while the same quantity via the norm bound is negative by a factor 2.4e5–9.6e7: an O(1) numerator divided by a 1e-6 artifact floor. Every norm routehad to fail. From here on chains stop at the compute cap (h ≤ 1500), never at a step.

Compression breaks the cap: a certified step at zone 155,921. [sandbox · T115] TRANSPORT-BLOCKED (26/26), with a large compression gain. Transport between the ladder's non-nested grids certifies only mild refinement — and for a principled reason, not a weak bound: on nested ladders, where the transport error is exactly zero, the Schur floor itself falls like ρ^−1.7, so no bound can undo it. The two-scale compression, however, keeps the step bit-exactly margin-free and certifies a step at n = 155,921 (117× deeper than T114), on a fine lattice of 93,470 cells compressed to 1490; the longest chain runs 10 steps, 33 certified steps over four chains, every certificate 10⁵–10¹¹× above the numerical noise floor. The stopper is cost on 3 of 4 chains and a failing step on none. The remaining list is the shortest of the series so far: three points, only one of them an inequality.

The induction step is a boundary process — and the prime comb refuses to be compressed. [sandbox · T116] RICCATI-PARTIAL (33/33): the global pole rides exactly in a 12×12+12+1 state (bordered elimination, no truncation anywhere), and the Riccati march runs 169,236 prepends to 1,354,088 cells — 903× the old cap — at flat cost, 76 µs per step. Declared honestly as a cost-geometry demonstration, not a Weil certificate. What breaks is unexpected: the full symbol does not decay at all, because every incoming cell couples back to every prime power in the window. The prime comb is the object that refuses compression. Bonus, and the hinge of the next three parts: the one remaining inequality acquires textbook shape — ε is the error of a piecewise-constant Galerkin method, and classical Aubin–Nitsche duality hits the measured exponent.

Theorem-shaped, and no rate lost. [sandbox · T117] THEOREM-SHAPED (23/23): the family is exactly nested, so ε is a Galerkin best-approximation error of one bilinear form and its monotonicity is a theorem rather than a fit; positivity becomes a non-membership statement. The direction trap is handled in the open — Céa and Aubin–Nitsche give upper bounds — and the two-level chain delivers a certified lowerbound on 19/19 pairs (bound/ε ∈ [0.111, 0.185]) at rate θ' = 1.74 against θ = 1.76: no power of D is lost. Self-correction: T116's factor-120 jumps were a sweep artifact — all 23 prime-power entries actually raise ε. What remains is three named analytic lemmas about one symbol, each with a classical address, two of them constants rather than rates.

Two of three lemmas stand — and the arithmetic half closes as a theorem. [sandbox · T118] TWO-OF-THREE (36/36): the first route is refuted with a reason (the exact two-grid symbol is a harmonic mean, which the comb dips make vacuous on 14/14 windows), and the rescue is a classical shift onto the oscillation Gram, whose symbol is the arithmetic mean of the same aliasing pair — the low-frequency negativity is suppressed quadratically instead of poisoning the statement. On a 15,680-point FFT lever the certified floor rises logarithmically and crosses zero on 3 of 5 zones: the failing windows were under-resolved, not obstructed. Saturation turns out to be an identity here, so its constant is computed (0.252–0.336) rather than assumed. [sandbox · T119] ARITHMETIC-DONE (27/27): the arithmetic half closes as a theorem — positivity of the symbol for every cell width below an explicit D₀(α) = exp(−(Ξ(α) + B)), with Ξ the prime-power atom count and B = −1.0474 universal (drift ≤ 3.1e-4). The energy route to the remaining statement is proven empty — a tautology — which is what makes that statement genuinely new content. And the sharpest identity of the run, κ_end = 1/(1+R) exactly (1.1e-16), reduces everything to one discrete Harnack inequality with a classical address.

Worth stating plainly, because it is the load-bearing fence of the whole sprint: this chain is classical numerical analysis from end to end and contains no zeta input anywhere. As a conditional lemma the material is essentially complete — what is missing is a proof of the Harnack statement, not a discovery. Sandbox; not RH evidence.

The Harnack core becomes proof-shaped — at the price of one more defect. [sandbox · T120] HARNACK-EXPLAINED (21/21): the mysterious constant R ≈ 1 is not a constant at all but a symmetry. The two increment families are the odd and the even half of one sequence, offset by a single fine cell — so given one sign on the corner cells their difference is a sum over disjoint neighbour pairs, and |R − 1| ≤ 0.04745 follows unconditionally, dominating the measurement on 3112/3112 sign-pure rows. Two negative results count as much: the per-cell version of the inequality is provably false (ratios up to 4.8e3 — only the summed form holds, which is exactly the form proved), and the discrete maximum principle is false too, so the sign route is closed and needs a corner-localized decay estimate instead. The honest cost: the defect count rises 3 → 4, because in the real frame the window grows with the zone, and the unconditional D₀ criterion then covers exactly 3 of 1492 zones — below the first handover the ladder actually uses.

Against the real ladder: a section statement survives where a symbol statement dies. [sandbox · T121] WIDE-RESTRUCTURED (21/21): the chain never needed the symbol to be positive — it needs the finite sectionto be, at the frame's own resolution. Over 16 windows of the real ladder (n = 7…283,303) the section is positive 16/16 while the symbol infimum is negative on 8 of them; where it is negative the section eigenvalue sits on the positive side at 0.64–3.23× its size. The mechanism has a classical address (Christoffel functions): a polynomial of that degree cannot fit inside one comb dip. The measurement discipline is kept explicit — dense rows carry a Cholesky certificate, large rows are labelled measurements, because a Ritz value can refute positivity but never prove it. The net balance is then computed instead of feared: only a poly-log deficit (α^−1.57, uniform in the resolution) rather than collapse, decomposed exactly into two repairable steps — and one link of the chain is honestly refuted on 42/42 rows.

Both repairs land: the refuted link was an identity all along. [sandbox · T122] NET-IMPROVED (20/20): the term that broke T121's link is not an error term but the reflection half of an isometry— the oscillation block is exactly a compression of one window Toeplitz form. From that plus a certified cell envelope and Parseval follows a certified band floor, non-vacuous 36/36 up to α = 6.31, where T121's budget tore at 3.45. The structural version of the Rayleigh step is sharp(slack 1.00–1.03, drift α^−0.002): no α-dependence passes through it any more. Certified deficit halves to α^−0.729 and is exactly uniform in the resolution; with the measured coupling the balance reads α^−0.113 against the chain's own ceiling α^−0.116 — statistically indistinguishable. The chain closes up to its own ceiling as soon as the coupling step stops being a worst case.

The coupling resists — structurally, and the reason is the result. [sandbox · T123] CBS-RESISTS (20/20): the band margin closes almost for free (that reduction costs under 1%) and the oscillation block gains uniform certified positivity on every row of the ladder. But every certified route to the coupling needs the coarse form from below, and there the numbers are brutal: λ_min of the coarse block is 1.7e-5–2.9e-4, three to five orders below the certified envelope. The diagnosis is exact — the near-null direction of the window form is smooth, so it lives in the coarse space, and its eigenvalue comes from a cancellation inside the form that no pointwise symbol minorant can see. So the worst-case coupling step must stay, and the theorem now says why. The entire remaining gap (α^0.5) is that one coupling — and it is identified, not estimated: it is the same object as the lid the chain throws away, four to eight recursion levels deep. Verdict: the two-level argument cannot be tightened. It has to become a recursion.

The telescope carries — because the rung is a maximum. [sandbox · T124] TELESCOPE-CARRIES (28/28): the nested level chain is one window form on nested spaces (nesting exact to 2.4e-14), so the whole two-level system of the previous part is literally one rung of a ladder, every rung is the saturation identity, and the rungs telescope: their sum is the whole quantity (9.2e-10). Then the key of the part — the direction flip. Read as a residual in the inverse norm, the rung is a maximum, not a minimum: every test vector gives a lower bound, and the denominator wants the form from above— exactly where the certified machinery works. The resulting certified rung bound holds on 400/400 rungs with a drift seven times weaker than T123's. T123's obstruction is still there and is now irrelevant: the new bound never forms the object that was obstructed. The recursion closes in the right direction, coarse to fine, with base case pure semidefiniteness — and the balance moves +0.444 of the α^0.5 gap, about nine tenths of the way to the ceiling.

What that leaves — stated as the diary states it. The coupling defect collapses: it is no longer a quantitative estimate but a sign that the coarse-to-fine induction already carries, and three quantities (the coarse minimum, its condition number, and the coupling constant) leave the chain entirely. One new defect is booked, honestly negative: the solution-free version of the rung fails, because the coupling term cancels the data instead of perturbing it — so the certified bound carries the solution along, on the same bookkeeping standard as the rest of the chain. The Harnack pair from T120/T121 is untouched.

The finale: the chain composes, and the Harnack pair leaves the spine. [sandbox · T125] ASSEMBLY-GREEN (34/34): every certificate of the arc mounted on one ladder, end to end. All five stages — base Cholesky, telescope rungs, ε lower bound, the κ chain, the margin-free handover — complete on 52 of 52 zones, and 30 of those zones form a run that literally composes: consecutive prime-power atoms on one common resolution, so the new window of a zone isthe old window of the next one (an integer identity on 29/29 pairs), the incoming atom's block is bit-exactly zero, and the output of every step is the input of the next (residual 2.9e-14). That is a way around the frame seam rather than through it: no two consecutive gap ratios are dyadic, so per-zone frames can never be refined onto one another — the run picks one frame for all of it and pays with its length. The load-bearing finding is a change of shape: the weakest stage of the whole mounting is the κ chain, but that chain is the second, independent route — the load-bearing spine is Cholesky-certified on all 52 zones, so the Harnack pair no longer carries anything. 430 completed Cholesky certificates, 440 identities, the certified margin a factor 32–8.7e10 above the declared floating-point bound, and the base case an equivalence rather than a bridge.

And then the accounting. Theorem V-final is printed with line-by-line attribution — 25 lines: 10 identities, 9 Cholesky certificates, 3 window certificates (all three inside the route the spine does not use), and exactly one hypothesis, which is an accounting convention — plus a five-point statement of what is not claimed. Of the 31 links of the chain, 90.3% are an identity or a completed Cholesky; on the spine it is 96.2%, with zero window certificates. The series ledger: 24 refuted routes across 18 parts, four of them killed by the same cause — a bound that needs the coarse form from below, which is why the exit was a direction flip and not a sharper estimate — and a seven-station cascade whose last station is new. What remains is named as what it is: uniformity in the zone index, not size.

Where the series ends: a finite, machine-certified chain whose load-bearing spine is 96.2% identity or Cholesky certificate and whose only hypothesis is a declared accounting convention; a relay mechanism that never failed a step; and certified single steps as deep as zone 155,921. What does not stand is uniformity in the zone index — and that, not any missing estimate, is the honest distance to any infinite statement. The distance to RH remains large. The series completed at 125 parts and 3139/3139 sandbox checks; the second phase — “the full proof” (T126–T176) — then closed its measurement programme as planned, with the exact cores load-bearing as v562 and the work continuing in the backflow rounds of the live feed below. The complete end-of-series recap is preserved verbatim behind the expander. Sandbox; not RH evidence.

1. Done[machine-verified]

   Mechanism + finite RTF + Weil structure + amplitude route + proof package

   v535–v537: Hecke, Eichler, half-integral bridge. v538: one finite relative-trace identity. v539: Weil structure fully identified up to two explicitly isolated obstructions. v540: amplitude Dirac + geometric polarisation with the Cohen seed Θ(d) = −48·L(−1,χ_d), the universal even-k deletion as square-class double counting, the positive linear carrier with plus-only ζ-balance, and the exact FE — with the open boundary λ* named inside the claim. v541: the T78–T85 proof package — matching lemma proved exact-integer on [4, 10⁶], transport ledger closes exactly (Δ_pole ≡ Δ_conv ≡ 0 proven), character-exact signed envelope, arch internal via Legendre duplication, coherent class closed by the λ-equivariant CM channel — with the two named limits inside the claim.

2. Near[sandbox]

   I5 geography complete; the induction compressed to one sign plus one convention, and assembled end to end

   After 230 probes (5360/5360 sandbox checks) and the promoted modules v535–v564 plus v569–v570, v573 and v576–v596 of this arc: the matching lemma is closed, the I5 geography is complete, and the induction that would carry I5 is compressed from one matrix inequality (T104) to a sign the coarse-to-fine recursion already carries plus one declared accounting convention (T124/T125) — assembled end to end on 52 zones in the T125 finale, its load-bearing spine 96.2% identity or Cholesky certificate. What remains TFPT-specific is one object, I5 ⟺ Weil positivity ⟺ RH (an equivalence typing, not a proof claim); what is missing is uniformity in the zone index, and the phase-2 measurement programme on it closed as planned (T176; exact cores load-bearing as v562). Not RH evidence.

   ▸The full ‘Near’ status paragraph, preserved verbatim
3. Big if[sandbox]

   A genuinely new functor

   A compiler functor that is Hecke-translating, Euler-preserving, and ξ-carrying without smuggling ζ from outside. Kills are preregistered; classical named pieces stay classical (Weil 1952, Waldspurger, Cohen 1975, Shintani, Cauchy–Littlewood, Hilbert–Carleman det₂, Connes 1999, Connes–Consani 2021 as context).

4. Dream[sandbox]

   Riemann Hypothesis

   Not claimed. Not evidenced. I5 ⟺ RH is an equivalence typing of the irreducible core — not a proof claim. Crypto unaffected.

An external note asked: what if primes are not the origin but the readout— the shadow of the finished geometry in discrete arithmetic? The checkable core of that reading is exact, on the compiler's own objects. The E₈ theta function, computed from the glue decomposition, is the Eisenstein series Θ_E8 = E₄: shell counts r(2n) = 240·σ₃(n), and the first shell is literally 240 — the root count.

The “address space” reading is unique factorization, exactly (shell counts factor over coprime addresses; the must-fail control shows non-coprime does not). The “independent check channels” reading is a theorem: the Hecke operators T_p act with eigenvalue 1 + p³ for every prime, they commute, and the compiler's theta is a simultaneous eigenvector of all of them. And L(E₄, s) = ζ(s)·ζ(s−3): the Riemann zeta function appears as the factorized shadow of the E₈ counting function.

Honest scope: these are classical facts (Jacobi, Hecke) — the content is that the compiler's own objects realize them verbatim, fixing the direction of explanation inside TFPT's narrative: geometry first, primes as readout. The bolder framings (“primes as compiler eigenfrequencies”, “RH as maximal coherence”) stay typed hypotheses, not adopted.

“E₈ is an error-correcting code” is now a theorem in the suite: Construction A on the self-dual extended Hamming code [8,4,4] — the Reed–Muller code RM(1,3) — yields E₈ exactly (even, unimodular, shells 240/2160), and every single-bit error is exhaustively correctable: 16×8 corrupted words, a unique nearest codeword every time (v626). Round-22 fence: the stronger reading — E₈ as an error-correction hull, the 240 root operators plus the 16 neutral kernels as a twisted group-algebra basis of End(S₊) — died as preregistered (SYNDROME-DEAD, v805); the Construction-A theorem here is untouched.

v638 then made the dictionary compiler-native instead of decorative. Stage 1 killed the naive coordinate placement; among all 30 placements of the code exactly one(up to the anchor orientation) carries both compiler symmetries. On that placement the code's coordinates organise as four μ₄ pairs — one bit per pair: the family 3-cycle rotates three of the pairs and fixes the fourth, the anchor — 3 families + 1 anchor, read off the code. The placement reproduces the v629 root censuses verbatim, and the syndrome space factors along the same structure.

Compiler ties, typed: 8 = rank, 4 = d = |μ₄|, 16 = the carrier half-spinor. Robustness language on this page now has an exact anchor — a code with a decoder, not a metaphor.

The surprise of the third external review (v624): an explicit integer matrix P with det −6 satisfies Pᵀ J_det P = J_fix exactly — the prime-front determinant form (det 2) and the cover polarization lattice of the geometry program (det 72) are the same rational quadratic form, the cover an index-6 sublattice. A genuine new bridge between prime analysis and Hodge geometry: the two programs this diary has been running were computing on one lattice.

v627 measured what that buys: transported through the congruence, all 67 complete windows land in the positive cone of the cover polarization lattice — on one sheet, det S > 0 everywhere. v635/v636 then closed the “found matrix” worry: P is the unique minimal, operator-compatible congruence in its census class, and it is constructed from canonical operator data (the null-cone rays are exactly ker C_V and fix C_V), not guessed.

Honesty, twice: chamber membership is a density-layer statement — scrambled combs do not leave the chamber (v582) — so the fine C = 1 arithmetic lives inside the chamber; and v637 closed the tempting follow-up as a preregistered negative: the fine Hodge-ray invariants do not predict the C = 1 margin window-wise beyond the trivial h-trend.

v629 left a sharp positive residue: the μ₄ clock acts freely on the 240 E₈ roots with exactly 60 orbits — and 60 = D_start, the cascade's starting value. v633 built the quotient properly: the 60 orbits are 60 lines in ℤ[i]⁴ carrying a hermitian form (the “complex E₈”), and the 60 order-2 unitary reflections they define generate a group of order 46080.

v634 identified it: exactly 60 reflections, invariant degrees (8, 12, 20, 24) — the Shephard–Todd group G31, pinned by fingerprints computed, not cited. The compiler sits inside it canonically: σ = c⁴, J = c⁹ — the family 3-cycle and the μ₄ clock are powers of one order-12 element, and the μ₄ center is a power of the clock.

The kill that keeps it honest: |G31| = 46080 = |W(D₅)|·|W(A₃)| — the compiler's glue factors — looks like destiny and is numerology. The abstract isomorphism dies on cheap invariants, and stronger: W(D₅) embeds in no rank-4 group at all, so the order coincidence carries no subgroup structure. Also reproduced one level deeper: the compiler clock is not the ζ₁₂-regular element of G31 (census 19×12 + 3×4) — the v629 kill stands.