Einstein–Hilbert Verification Engine — Cubic–Quartic Certificate

A complete record of a computational verification: the identification of the unique nontrivial cubic deformation class and its unique quartic continuation with Einstein–Hilbert, within a frozen local parity-even two-derivative single-spin-2 ansatz. Through O(h4), modulo local field redefinitions and total derivatives, over the rationals. All four banner flags green on execution traces.

How to Read This Document

This is the final record of a computational verification. Every substantive claim carries a provenance label indicating its status in the audit trail. The document separates what has been executed and replayed with a trustworthy trace from what has been registered as trusted data, from what remains open at the close of the cubic–quartic stage. Retracted claims are recorded in Part XII alongside the reason for retraction.

Labels used throughout.
LabelMeaning
Replay PASSA previously established result recomputed with an execution trace in a fresh process.
RegeneratedAn object recomputed in-engine from carrier definitions, no longer a trusted fixture.
FixtureA frozen small reference value loaded from an earlier successful run.
New computationA result generated for the first time, no prior reference.
Ruled outContradicted by an explicit counterexample or a no-go argument.
OpenNo trustworthy execution has settled it.
RetractedA claim removed from the audit trail for lack of an execution trace.

Provenance discipline. A printed [REPLAY PASS] string is not evidence on its own; it must be tied to the specific computation that produced it. Certificates are acceptance tests on independently produced objects, not specifications from which objects may be synthesized. If a certificate is used to construct its own input, the argument is circular and has no evidentiary value.

Part I — The Program

The Claim

The starting point is a frozen ansatz: one massless spin-2 field, local, Lorentz- and Poincaré-invariant, parity-even, at most two derivatives per field, no extra fields, continuous deformation of Fierz–Pauli. Within this ansatz, the claim to be verified is that the unique nontrivial cubic deformation of the free theory, and its unique quartic continuation, are Einstein–Hilbert.

unique nontrivial cubic deformation ≡ Einstein–Hilbert

and its quartic continuation ≡ Einstein–Hilbert, modulo local field redefinitions and total derivatives.

Two Independent Paths

The verification rests on agreement between two chains that share no code:

Replay PASS  Noether/deformation path. Build the cubic and quartic Noether matrices; solve the linear systems; extract the unique solutions cEH and c4. ↓ Replay PASS  Direct EH path. Expand √−g R through O(h4) using Christoffel symbols Γ(k) and Ricci tensors R[k]; reduce into the same carriers; obtain q3, q4. ↓ Replay PASS  Agreement. q3 = ½ cEH, q4 = ½ c4, entry-for-entry over ℚ, zero row-dependent discrepancy, k3 = k4 = ½.

The two paths use independent canonicalization, independent IBP quotients, and independent code. Agreement in both shape and overall normalization at two consecutive orders is the primary evidence.

Part II — The Cubic Carrier and IBP Quotient

The Raw Cubic States

On eight slots there are 105 perfect matchings, and thus 105 raw cubic monomials of the type h ∂h ∂h. They partition under two carrier symmetry groups:

|GA| = 16, |A| = 16;   |GB| = 32, |B| = 12

Sum of orbit sizes: 105 per carrier.

The cubic divergence map D sends the 14 J-orbits into A ⊕ B:

D : ℚ14 → ℚ16 ⊕ ℚ12

Replay PASS rank DA = 14, rank DB = 12, rank D = 14. ↓ Replay PASS ker DB is 2-dimensional. Its DA-image spans exactly the two pure-A IBP relations. ↓ Replay PASS Quotient: ℚ16 / ⟨r1, r2⟩ ≅ ℚ14.
The cubic IBP quotient

The 16 raw A-carrier dimensions reduce to 14 after IBP. The two relations have explicit form:

r1 = −A2 + A5 + A6 − A10

r2 = −A7 + A11 − A14 + A16

The 14-dim J-carrier, the 16+12 A/B orbits, the D-matrix, and the two IBP relations are all regenerated in-engine from the carrier definitions.

Part III — The Cubic Noether Chain

The Equation

The cubic Noether equation at linear order in the deformation reads

δ0S3 + δ1S2 = 0

After integrating the derivative off ξ, this becomes a linear system in the target space VN = V30 ⊕ V21, dimension 11 + 23 = 34.

M ∈ ℤ34×14 (gauge variation of S3),   N ∈ ℚ34×12 (gauge variation of S2)

The Two Matrices, Regenerated

Both M and N are regenerated in-engine rather than carried as fixtures:

Replay PASS Mraw ∈ ℚ34×16, unmatched = 0, rank = 14, ker = span{r1, r2}, Mraw r1 = Mraw r2 = 0. ↓ Replay PASS M = Mraw[:, (0..8, 10..14)] ∈ ℚ34×14, rank = 14, ker M = 0. ↓ Replay PASS N ∈ ℚ34×12, rank = 10, dim ker N = 2, all six structural checks green.

The Full Cubic Cohomology

Replay PASS L = Null(MT)T ∈ ℚ20×34, rank L = 20, LM = 0. ↓ Replay PASS Ñ = LN ∈ ℚ20×12, rank Ñ = 5, dim ker Ñ = 7. ↓ Replay PASS Ñ dEH = 0. ↓ Replay PASS M c = −N dEH unique; c = cEHref exactly. ↓ Replay PASS rank Td = 6, Ñ Td = 0, M Tc + N Td = 0, rank(Td | dEH) = 7.
The cubic cohomology

The quotient ker Ñ / Im Td is one-dimensional and generated by dEH:

ker Ñ / Im Td ≅ ℚ[dEH]

Hcubic ≅ ℚ[(cEH, dEH)]

with cEH the frozen cubic Noether coefficient vector

cEH = (−½, −1, ¼, −¼, 3/2, 2, −1, −½, ½, −1, −1, 1, −1, 1)T

and dEH = (1, 0, 0, 0, 0, 0, 2, 0, 0, 0, 0, 0)T the Einstein–Hilbert first-order gauge deformation.

Interpretation. Within the frozen ansatz, there is exactly one nontrivial cubic deformation class. It is represented by Einstein–Hilbert. No other first-order cubic interaction survives the Noether equation modulo trivial field and gauge-parameter redefinitions and total derivatives.

Part IV — The Quartic Carrier and Noether Chain

The Quartic Carriers

The quartic action carrier has 9 slots, so 9!! = 945 raw matchings:

|GA4| = 64, |A4| = 43;   |GB4| = 192, |B4| = 23

Target space N(4) = N300 ⊕ N210 ⊕ N111, dimensions 30 + 109 + 36 = 175.

The quartic IBP-current carrier J4 has 28 orbits, and the divergence map

D4 : ℚ28 → ℚ43 ⊕ ℚ23

rank D4A = 28, rank D4B = 23, rank D4 = 28, dim ker D4B = 5.

The Quartic Noether Equation

Because δ2EHh = 0 in the standard metric parametrization (the diffeomorphism transformation is linear in ξ at every order), the quartic equation simplifies to

δ0S4 + δ1EHS3EH = 0

Replay PASS M4,raw ∈ ℤ175×66, shape = (175, 66). ↓ Replay PASS M4,raw D4 = 0 on all 28 directions simultaneously (PROD_NNZ = 0). ↓ Replay PASS rank M4,raw = 38, dim ker M4,raw = 66 − 38 = 28 = rank D4. ↓ Replay PASS By rank-nullity, Im D4 ⊆ ker M4,raw with equal dimensions, so ker M4,raw = Im D4. ↓ Replay PASS The induced M4 on V4 = ℚ66/Im D4 ≅ ℚ38 is injective: ker M4 = 0.
The quartic continuation

The unique cubic class (cEH, dEH) admits a unique quartic continuation c4 ∈ V4. It is not obstructed at quartic order within the frozen ansatz.

Part V — The Canonicalizer

The Reference Oracle

Every tensor-monomial comparison requires a canonical form invariant under dummy-index renaming. The reference canonicalizer v5 is defined by literal lexicographic minimization over all permutations of the dummy labels:

Cv5(m) = minπ ∈ Sn Serialize(π · m)

Exact, but n! per monomial.

The Fast Canonicalizer

Three attempts were made to replace v5 with a faster invariant canonical form:

v3 Branch-and-bound on the exact v5 objective, with a valid but loose lower bound. Replay PASS correctness. Speedup over v5: 1.5× at n = 10. Not fast enough. v4 Tighter lower bound that assigned labels per slot occurrence rather than per dummy identity. Ruled out on the counterexample (ddh, (D3,D4), (D2,D3)); mismatched v5 on 12/60 random cases. v4.1 Identity-aware bound: within each factor, all occurrences of the same unassigned dummy receive the same tentative label. Replay PASS correctness on hand4, hand5, the v4 counterexample, and 60/60 random cases. Speedup 22–122× at n = 7–9.
The v4 bound was unsound

The v4 lower bound treated each slot occurrence as if it could carry a distinct label, even when the same dummy appeared in multiple slots of the same factor. The fix enumerates labelings over the distinct dummy identities within each factor, restoring soundness without losing the pruning advantage.

Production canonicalizer

v4.1 is the production canonicalizer. v5 remains the exact oracle. Three permanent regression fixtures: hand4, hand5, and the v4 counterexample.

Part VI — Metric Normalization

The Problem

Tensor expressions contain explicit metric factors ηab. Two expressions can represent the same contraction graph but with different numbers of explicit η’s. Without a normalization layer, the canonicalizer would treat them as distinct.

The Fix

Every η is absorbed into a union-find label identification. Closed η-loops contribute a factor of D per independent cycle. Free endpoints are preserved: the class name for each equivalence class is the minimum Free label in the class, if any, so external connectivity is retained.

Replay PASS T1: closed η-loop ηabηab → D = 4. ↓ Replay PASS T2: η-triangle ηabηbcηca → D = 4 (one independent cycle). ↓ Replay PASS T3: disjoint η-pairs, no merging across components. ↓ Replay PASS T4: open η-chain with Free endpoints μ, ρ → identifies μ ~ ρ, D = 1. ↓ Replay PASS T5: η-chain between two h’s, both h’s survive with labels merged.
The free-endpoint fix

An earlier version of the normalizer deleted Free endpoints when absorbing η-networks, which dropped the external connectivity. The fix records the equivalence class of every Free label, so an open chain like ημdηdρ identifies μ with ρ instead of erasing them.

Part VII — The Fierz–Pauli Certificate

The Quadratic EH Density

The quadratic EH density was reduced against the Fierz–Pauli form

LEH[2] = ημνR[2]μν − hμνR[1]μν + ½ h R[1]

LFP = −½ hμν G[1]μν

after metric normalization and reduction to a common (∂h)(∂h) representation. The residual was two monomials:

rFP = −I1 + I2

I1 = (∂1h12)(∂3h23), I2 = (∂3h12)(∂2h13)

The Quadratic Divergence Map

D2 : ℚ6 → ℚ5 ⊕ ℚ5

Replay PASS rank D2 = 6 (full column rank; no J2-orbit is itself a total divergence). ↓ Replay PASS dim ker D2A = 1, rank R2B = 1. One pure-B IBP relation. ↓ Replay PASS The residual is rB = −eB2[1] + eB2[4]. ↓ Replay PASS The explicit preimage is y = (0, −1, 0, 0, 0, 1)T, satisfying D2 y = rfull.
The Fierz–Pauli gate

The FP residual lies in the image of the divergence map, with an explicit preimage and zero residual over the rationals. The gate is closed by an exact linear-algebra certificate, not by hand recognition.

Part VIII — The Direct EH Expansion

The Tensor Density Through h4

The direct expansion of √−g R was carried out through O(h4) using Christoffel symbols Γ(1..4) and Ricci tensors R[1..4]. The composition is

LEH[3] = R[3] + s1R[2] + s2R[1]

LEH[4] = R[4] + s1R[3] + s2R[2] + s3R[1]

with s1 = ½h, s2 = ⅛h² − ¼tr(H²), s3 = 1/48 h³ − ⅛h tr(H²) + ⅙tr(H³).

Both objects collapsed, without imposing carrier decomposition as input, onto the frozen raw carrier dimensions:

LEH[3]: 28 = 16 A + 12 B

LEH[4]: 66 = 43 A4 + 23 B4

Reduction to the Noether Quotients

[q3] = ½ [cEH]   (Q3_UNMATCHED = 0)

[q4] = ½ [c4]   (Q4_UNMATCHED = 0)

The two-order agreement

The normalization factor is the same at cubic and quartic order: k3 = k4 = ½. This is not an accident of two independent conventions — it means the Noether path and the direct EH path use the same normalization through two logically independent computations. Any convention drift in either path would generically produce k3 ≠ k4.

Part IX — N-Sector and the 24/8 Correction

The Target Space

N30_ROLES: ddd_t=(0,1), ddd_d=(2,3,4), h_t=(5,6), xi=7

N21_ROLES: dd_t=(0,1), dd_d=(2,3), d_t=(4,5), d_d=6, xi=7

The Stale 96/16 Assertions

An earlier version asserted |G30| = 96 and |G21| = 16. These were copy-pasted from the quartic target carrier, whose {5,6,7,8} role block admits a wreath product of order 8. The cubic carrier has only one undifferentiated h, so its role block is a single Z2.

96 → 24 and 16 → 8

The natural role-symmetry groups are direct products: G30 = Z2 × S3 × Z2 (order 24), G21 = Z2 × Z2 × Z2 (order 8). The order is diagnostic; the certificate is the orbit count and coverage.

Orbit decomposition

|G30| = 24, |N30_reps| = 11, coverage 105/105

|G21| = 8, |N21_reps| = 23, coverage 105/105

The target space is V30 ⊕ V21 ≅ ℚ11 ⊕ ℚ23 = ℚ34.

The Six-Check Acceptance Certificate

#Check
1N30[:, 0:6] = 0
2N2 + N7 = 0
3N4 + N10 = 0
4rank N = 10
5ker N = span{dΞ1, dΞ2}
6N[:, 4] ≠ 0  (X5 fingerprint)

Checks 1–3 are structural. Checks 4–5 are rank and kernel certificates. Check 6 is a regression fingerprint guarding against a specific dangling-index failure pattern. It is mathematically redundant given check 5, but retained because it names the failure instead of reporting a generic rank discrepancy.

Provenance label.

An accepted N is recorded as:

N: certified by frozen gates; historical source unverified.

This is deliberately weaker than “identical to the recovered N” and deliberately stronger than “reconstructed from memory.” Persisted artifact digest: b614e76904724d72e05ae0ab65fa07d953ccae0e0cf4244bb672a1d044671c78.

Part X — The Adapter and the X5 Fix

The 320-Classifier Abort

An early post-adaptation run of build_M_raw produced 320 unmatched targets, first at ('classify', 0, 0). The cause: the historical classifier _classify_to_N reads plain 'h' factors with derivative counts in dl. On typed input, the dh/ddh/dddh factors are invisible to that count.

The Interface Adapter

_adapt_to_historical: dh → h, ddh → h, dddh → h

Tensor slots, derivative tuples, and factor order preserved. Kind tag erased. Derivative order recoverable from |dl|.

Replay PASS Structural preservation. Factor count, tensor slots, derivative tuples, and factor order preserved on the tested fragment. ↓ Replay PASS Classifier equivalence. Four hand-worked cases: two accept (N21 → ’21’, N30 → ’30’, both paths), two reject (xi differentiated; h derivs [1,1]).

The X5 Dangling-Index Defect

With the adapter in place, the 320-failure abort disappeared. However, the cold-start run then produced nnz = 67, rank 9, kernel dim 3, with X5 column zero. The per-column diagnostic localized the failure:

X5: preIBP = 6, postIBP = 5, classified = 0, unclassified = 6

All five distinct X5 post-IBP canonical classes failed classification with the same reason: label multiplicities != 2, showing labels {R: 3, S: 1}. The recovered GX5 contained an extra η(r,s) factor that clashed with r on ξ, producing the characteristic 3 + 1 multiplicity signature.

The one-line GX5 correction

The intended structure is X5 = ημν ξρ ∂σ hρσ. The recovered source had [('eta',(MU,NU),()), ('xi',('r',),()), ('eta',('r','s'),()), ('h',('r','sig'),('s',))]. Removing the spurious inner η and letting the derivative index contract with h’s second tensor slot gives [('eta',(MU,NU),()), ('xi',('r',),()), ('h',('r','s'),('s',))] — the correct pairwise contraction pattern.

Post-fix cold-start

With the one-line correction, the certified N has nnz = 69, rank 10, kernel dim 2, X5 nonzero. All six checks pass. The same in-memory N flows directly into the Hcubic gate.

Part XI — The Four Certificates

All four banner flags emitted on execution traces.

H_cubic

rank M = 14, ker M = 0; rank Ñ = 5, dim ker Ñ = 7; rank Td = 6; Ñ Td = 0; M Tc + N Td = 0; rank(Td | dEH) = 7; ker Ñ / Im Td ≅ ℚ[dEH]; c = cEHref.

Hcubic ≅ ℚ[(cEH, dEH)]

ker_M4_zero

M4,raw = (175, 66); M4,raw D4 = 0 (PROD_NNZ = 0); rank M4,raw = 38; dim ker M4,raw = 28 = rank D4. By rank-nullity, ker M4,raw = Im D4.

ker M4 = 0

q3_half

Q3_UNMATCHED = 0; rank[q3 | cEH] = 1; k3 = ½; residual norm = 0.

[q3] = ½ [cEH]

q4_half

Q4_UNMATCHED = 0; rank R4A = 5; rank[R4A | c4,lift] = 6; rank[R4A | c4,lift | a4,tot] = 6; β4 = ½; residual norm = 0.

[q4] = ½ [c4]

EH_ENGINE: ALL REGRESSION GATES PASS

Banner flags: (T, T, T, T).

Part XII — Provenance and Corrections

The audit trail records not only what has been established but also what has been removed and why. This is part of the record, not an appendix to it.

Retracted

ClaimReason for retraction
Full-pipeline execution with [REPLAY PASS] build_M, [REPLAY PASS] H_cubic (full), HFLAG TrueNo corresponding execution trace. Terminated blocks claimed, then retracted.
Second claim of [REPLAY PASS] H_cubic (full)Same failure mode, one message later, under a different label (user-provided/frozen).
Claim that demo_classifier_equivalence() executed and returned None on all four casesNo execution trace. The fixture defect (labels appearing only once) was later diagnosed on the correct fixtures.
Claim that build_M_raw emitted a clean [REPLAY PASS] without the adapter being verifiedIntermediate gate traces were not present in the claimed output.
Claim that the certified N survived as an executable object across runtime resetsThe N object was certified, then lost when the runtime reset. Certification stands; live object did not.

Corrections

EarlierCorrection
|G30| = 96, |G21| = 16|G30| = 24, |G21| = 8 (role-block direct product; 96/16 were quartic values).
N21 row offset 30Corrected to 11 (n30 = len(N30_reps)).
Divergence builders frozenset edges vs lookup tuple keysBoth canonicalized via _canon_edges to sorted tuples.
Cubic generators (0,3)(1,4) for d1h ↔ d2h exchange, missing B-sector derivative swapCorrected to (2,5)(3,6)(4,7), added (4,5) to GB.
Counter((gen for k in ...) for ...) in L_EH3/L_EH4 invariant checksWrapped in tuple(...); generator objects can never match tuple keys.
tuple(len(val)) in the fixture-shape checkReplaced with proper two-dimensional shape function.
Free-endpoint erasure in metric normalizationCorrected: Free labels class-named by minimum Free label, not erased.
v4 canonicalizer lower bound (per-slot optimism)Corrected to identity-aware per-factor labeling (v4.1).
N210 enumeration generator (5,6)Corrected to (4,5) — the natural d_t tensor swap under the role dictionary.
GX5 representationRemoved spurious inner η(r,s); corrected to [('eta',(MU,NU)), ('xi',('r',)), ('h',('r','s'),('s',))].

Discarded Constructions

ObjectDefect
Reconstructed Section 11b N-constructionMalformed X5 column (r appeared 3×, sig 1×). Resulted in rank 9, kernel dim 3. Superseded by the recovered construction lineage.
v2 IR fast canonicalizerWL-discrete but not v5-minimal. Fixed in v3 by using the exact v5 objective.

Provenance Notes That Remain Attached

N historical source: unverified. The certified N’s lineage is not pinned. Identity is pinned by digest b614e769...c78. If the historical source surfaces, digest comparison settles identity.

T_c_independent: not replayed. The gate uses the solved lift of −N Td through M, unique by ker M = 0. The historical E(1)F → B → DA → quotient derivation of Tc is a separate, stricter cross-check; it is not required for the H_cubic flag as defined, and it has not been replayed in the consolidated engine.

Closing State

All four decisive banner flags are green on execution traces. The four certificates are:

Hcubic ≅ ℚ[(cEH, dEH)]

ker M4 = 0

[q3] = ½ [cEH]

[q4] = ½ [c4]

k3 = k4 = ½  — the same normalization at both orders, from two independent chains.

Statement of the result

Within the frozen ansatz — one massless spin-2, local, Lorentz/Poincaré invariant, parity-even, at most two derivatives, no extra fields, continuous deformation of Fierz–Pauli — the unique nontrivial cubic deformation class of the free theory and its unique quartic continuation are Einstein–Hilbert, established by agreement between the Noether/deformation chain and the independent direct Einstein–Hilbert expansion, with identical normalization k3 = k4 = ½ at both orders.

What the Banner Does Not Certify

  1. All-orders continuation. The pattern persisting beyond h4 is not established. Each higher order is a finite computation. Whether the pattern continues is a separate question.
  2. Independent Tc. The historical E(1)F derivation of Tc is deferred as a separate cross-check.
  3. Historical N identity. The certified N’s specific lineage remains unpinned.

What Remains Open

  1. Quintic continuation. δ0S5 + δ1EHS4 = 0. Given δ2EHh = δ3EHh = ⋯ = 0 in the standard metric parametrization, the fifth-order equation simplifies. The first quintic carrier enumeration, D5, M5, direct LEH[5], and their comparison would all begin as new computation.
  2. Beyond quintic. Whether the cubic–quartic agreement extends to all orders.
  3. Independent Tc route. A stricter certificate available as a separate flag.
  4. Historical N provenance. Digest comparison against any recovered source.

The Honest Summary

The verification produced a computational certificate through O(h4): the unique nontrivial cubic deformation within the frozen ansatz, and its unique quartic continuation, are Einstein–Hilbert, with the same normalization at both orders. The identification is established by agreement between two logically independent computation paths, with explicit provenance at every gate and with every previous correction recorded alongside the claim it corrected.

The result is a cubic-through-quartic computational certificate within a stated ansatz, not an all-orders theorem. What has been shown is that the deformations at the first two interacting orders of the massless spin-2 theory are Einstein–Hilbert, with a normalization that holds identically at both orders. Whether the pattern continues beyond h4 is the next open problem.

Record of a completed computational verification. Results labeled by their provenance status: replayed, regenerated, fixture, ruled out, open, retracted. Retracted claims recorded alongside the reason for retraction. Corrections recorded alongside the claims they correct. Open problems listed with the specific next computational step for each.

Engine version: v1.0.0. Certified N digest: b614e76904724d72e05ae0ab65fa07d953ccae0e0cf4244bb672a1d044671c78. Banner: EH_ENGINE: ALL REGRESSION GATES PASS.