TRISDUCTION · RAM · MATHDUCTION KERNEL · REGISTER B
The Verdict-Reliability Layer · Conditioning-Aware Error Model · Four-Estimator Redundancy · Bootstrap Stability · Dual-Null Forward Seal · Fail-Safe State Machine · Substrate-Portable and Executable
STATUS: [⟀] Register B is the reflective-residence form of the verification cascade, founded on RAM, residence on the Ground. The computational arithmetic carries a conditioning-aware error model, a four-estimator redundancy cross-check, a higher-precision escalation ladder, a bootstrap verdict-stability gate, and a dual-null forward seal. Every identity is machine-confirmed and re-runnable. The verdict boundary is float-clean across the operating domain. Absolute fidelity to the calculations is enforced.
SCOPE AND INHERITANCE
Register B governs the MathDuction kernel: its arithmetic, its conditioning model, and its verdict quality control. It is founded on RAM, the Self-Dual Imprint read as residence on the Ground, and it inherits the shared discipline that holds the substrate to cascade-output: the Decalogue, the Mosaic Seal, the Omega Synthesis Guard, the Aperture Law, the Imprint-Honesty Law, and the Orientation-Blindness Law. It inherits Seals L, G, M and the shared quaternionic verdict kernel of Register A. The discipline that prevents the kernel's misreading is the precondition of every reliability gate below.
The map from a proposition to its warrant rows is a hand reading placed in front of the kernel, per Decalogue rule six. The instrument verifies supplied warrant and generates no mathematical truth. The reliability layer certifies that, on the supplied rows, the three-state verdict is the correct verdict of the floating-point computation and not an artifact of rounding, conditioning, Monte-Carlo noise, or a handful of leverage contexts. The layer operates on the math kernel and draws zero warrant from its own operation, per audit symmetry.
B.0 · KERNEL BOOT
The kernel boot runs four recorded batteries as the executable proof of load. Before the kernel issues any verdict it confirms, at machine precision, every identity below. A substrate that cannot reproduce the residues is not loaded.
The shared-kernel CHK battery and the Forward GV-CHK battery are inherited from Register A and read in the reflective interpretation. Mandatory at boot, in addition. One. The MD-HARD battery: the determinant-reliability theorem, the floor-gate separation theorem, the four-estimator redundancy cross-check with induced-error detection, the bootstrap-stability gate across stable, caveat, and fragile cases, and the dual-null forward seal, with all defense branches reachable. Two. The MD-WRK battery: two fully-specified worked examples, one Default and one Forward, with their construction matrices printed and every intermediate emitted, reproducible bit-for-bit from the stated grid and seed.
The boot forces the substrate to demonstrate, in addition to λ² = det(R): the conditioning-aware error bound relErr(det) ≤ 4·κ(R)·u_m; the floor-gate separation det(R) > ε on the locked side of the κ(R) < 10⁶ gate within the validity domain; the agreement of four independent det(R) estimators to machine precision; the agreement of the permutation null with the analytic Beta(1/2, (m−1)/2) null in Forward mode; and the deterministic recovery of the fail-safe state machine on every reachable fault. The boot is the demonstration, not the assertion.
FIDELITY LOCK, CLAUSE EIGHT, scoped to the MathDuction kernel. Numerical reliability travels with every quantity. No closed-form verdict is issued unless the four-estimator cross-check is consistent within the conditioning-scaled tolerance, the kernel identity holds within its conditioning-scaled bound, and the verdict does not sit inside the escalation band. A verdict that fails any of these is marked engineering-incomplete and routed through the fail-safe state machine of B.21 before it is emitted. Clauses one through seven of the Fidelity Lock are binding.
B.17 · THE VERDICT-RELIABILITY LAYER
B.17.1 · The Determinant-Reliability Theorem
Type T on the scaling, engineering-grade on the constant.
The verdict reads a 3×3 correlation Gram det(R) computed in double precision. The relative floating-point error of det(R), under elementwise rounding of the warrant entries at unit-roundoff u_m, scales linearly with the conditioning of the Gram:
relErr(det R) ≤ c · κ(R) · u_m, c ≤ 4.
The κ-linear scaling is the standard determinant condition-number result. det(R) = Π λ_i, the smallest eigenvalue λ_min controls the perturbation, κ(R) = λ_max / λ_min, so a relative entry perturbation of size u_m moves det by a factor κ·u_m. Theorem-grade. For a correlation matrix the Frobenius norm is bounded by 3, so the first-order constant is bounded by 3; the operating bound c ≤ 4 absorbs the higher-order terms. The constant is fitted over the conditioning sweep and is re-derivable on load.
Recorded fit, two warrant rows driven to a target correlation, the third independent, the data rows perturbed by relative u_m and det(R) recomputed, worst over 400 perturbations per row pair, worst c = 3.140 over the sweep. Consequence: at the conditioning gate κ(R) = 10⁶, relErr(det R) ≤ 4·10⁶·u_m = 8.882×10⁻¹⁰, so det(R) carries nine significant figures at the gate. The float64 determinant is reliable across the entire locked region.
B.17.2 · The Floor-Gate Separation Theorem
Type T on the worst-case bounds and the separation inequality. Engineering-verified.
Two thresholds read on the correlation matrix R whose determinant is the verdict: the collapse floor ε = 100·u_m·N and the conditioning gate κ(R) < κ*. The theorem fixes the regime in which they cannot contend, so the [⟀]/[X]/[?] boundary is never decided by rounding.
Lemma 1, eigenvalue floor. trace(R) = 3, so λ_max ∈ [1, 3]. With λ_max ≥ λ_mid and the trace constraint, λ_min = λ_max/κ at the gate forces λ_max ≥ 1.5, hence λ_min ≥ 1.5/κ. At κ = 10⁶, λ_min ≥ 1.5×10⁻⁶, attained by the negative-equicorrelation matrix, eigenvalues (1.5, 1.5, ε), off-diagonal ρ → −1/2.
Lemma 2, determinant floor. det(R) = Π λ_i is minimized at fixed κ by the positive-equicorrelation matrix, eigenvalues → (3, 0, 0) as ρ → 1, det(R) = (1−ρ)²(1+2ρ) = 27/κ² + O(κ⁻³). This is the global minimum over correlation matrices at that conditioning. At κ = 10⁶, det(R) ≥ 2.7×10⁻¹¹.
Separation. Collapse [X] requires det(R) ≤ ε. By Lemma 2, while κ(R) < κ*, det(R) > 27/κ*². Floor and gate cannot contend if and only if 27/κ*² > 100·u_m·N, that is if and only if κ* < κ_sep(N) := √(27 / (100·u_m·N)). At the operating gate κ* = 10⁶ this holds for N < 1216. Within the domain the gate fires [?] strictly before the floor fires [X]: a near-degenerate triad retires as under-determined, never as broken, and the boundary is float-clean. The interior margin det(R)/ε equals (κ_sep(N)/κ*)², which is 50.7× at N = 24 and 12.2× at N = 100. The separation gate κ* = 10⁶ sits strictly inside κ_sep(N), which is 7.12×10⁶ at N = 24 and 3.49×10⁶ at N = 100.
N kappa_sep(N) margin (kappa_sep/1e6)^2
24 7.118e+06 50.67x
100 3.487e+06 12.16x
300 2.013e+06 4.05x
1216 1.000e+06 1.00x
For N ≥ 1216 the margin closes. The gate ceiling drops to κ_sep(N), or ε rebases off the determinant's own error scale c·κ·u_m·det rather than N·u_m. The typical case, two rows driven to a target correlation with the third independent, returns det ≈ 4/κ and sits orders above the equicorrelation floor; the equicorrelation worst case above is the proved bound. Collapse is reached only at genuine rank-deficiency, where det → machine-zero and κ → machine-infinity together. The near-degenerate case is retired by the conditioning gate as [?], never by a noisy determinant near ε.
B.17.3 · The Four-Estimator Redundancy Law
The fail-safe against silent numerical error. det(R) is computed four ways from independent numerical routes, and a clean verdict requires their agreement. D1 = det(R) by LU factorization. D2 = Π eigvalsh(R), the eigenvalue product of the symmetric Gram. D3 = Π σ_i², the squared singular values of the unit-row matrix, det(R) = det(Q Qᵀ). D4 = λ², the squared scalar part of the composed quaternion triad. The redundancy tolerance is conditioning-scaled with a safety factor over the B.17.1 bound:
tol = max( 10 · 4 · κ(R) · u_m · |det R|, 8·u_m ).
If the maximum pairwise disagreement exceeds tol, the verdict is marked engineering-incomplete and routed to the state machine of B.21. Recorded on a clean independent triad: D1 through D4 agree to 8.882×10⁻¹⁶ against tol = 2.236×10⁻¹⁴, consistent. A silent 10⁻⁹ corruption injected into D1 raises the spread to 1.000×10⁻⁹, flagged against the same tolerance. The four routes share the input and not the arithmetic, so a fault in any one route is caught by the other three.
B.17.4 · The Verdict Margins and the Confident-Seal Gate
Three engineering diagnostics travel with every closed-form verdict. They refine whether a [⟀] is issued or downgraded to [?]. They are not a fourth state and not a probabilistic truth value. Trinary terminality holds.
identity_residual = |λ² − det(R)|, bound 4·κ(R)·u_m·|det R|.
collapse_margin = log10( det(R) / ε ), orders above the collapse floor.
conditioning_margin = log10( 10⁶ / κ(R) ), orders of headroom inside the gate.
A confident [⟀] requires the identity residual under its bound, the collapse margin above 3 orders, and the conditioning margin positive with headroom. A verdict that locks geometrically but sits with a small collapse margin or near the conditioning gate is downgraded to [?] with the small margin named, because the instrument does not certify the lock at that conditioning. The margins report how far the lock stands from every failure surface, in orders of magnitude, on the face of every verdict.
B.17.5 · The Escalation Ladder
Deterministic higher-precision recovery. A verdict escalates from double precision to 50-digit mpmath when any of three borderline conditions holds: identity_residual > bound, OR κ(R) ≥ 10⁵, OR det(R) ≤ 10·ε. On escalation the determinant is recomputed exactly on the float64 Gram at 50 digits and the verdict is re-decided. Recorded on a near-gate triad, κ(R) = 3.12×10⁶: det_f64 = 1.261908×10⁻⁶ against det_mp50 = 1.261910×10⁻⁶, relative difference 3.28×10⁻¹⁰, identity residual 7.32×10⁻¹⁶, the 50-digit recomputation confirming det > 0 and the conditioning gate returning [?]. The escalation is the bridge from the engineering-incomplete mark to a definite verdict. It never crosses the aperture and never fabricates a trace.
B.18 · BOOTSTRAP STABILITY AND VERDICT QUALITY
A lock can be genuine in aggregate and hinge on a handful of leverage contexts. The layer resamples the N reading-contexts with replacement B times, recomputes the verdict each time, and reports the fraction of resamples that agree with the point verdict. The fraction is a stability statistic, not a probability of truth.
agreement >= 0.95 STABLE seal.
0.80 <= agreement < 0.95 seal carried with a stability caveat.
agreement < 0.80 downgrade to [?], context-fragile.
Recorded, B = 2000. Robust triad: point [LOCK], agreement 1.000, STABLE. Fragile triad with two leverage contexts: point [LOCK], agreement 0.913, CAVEAT. Fragile triad with a single leverage context: verdict [?], agreement 0.629, FRAGILE, downgraded. On the two-leverage case, deleting the two leverage contexts and re-running returns [?] by conditioning collapse, confirming the lock hinged on them. The bootstrap gate converts hidden fragility into an explicit verdict-quality tier and downgrades to [?] when the lock does not survive resampling. It is Revision-Mandate-compatible: it moves the verdict only on the structural fact that the independence is carried by too few contexts, never on a reframe.
B.19 · FORWARD-MODE STATISTICAL HARDENING
The Forward seal rides the source-attribution statistic η_S above its permutation null, per the Non-Discrimination Theorem. The layer carries a second, independent calibration of the null and a guard against Monte-Carlo noise in the permutation quantile.
The analytic null. After centering and projecting out the two manifest axes, the witness and generator residuals live in dimension m = N − 3. Under the null of independence the squared partial correlation has a closed-form distribution:
eta_S = r²(W⊥, S⊥) ~ Beta( 1/2, (m−1)/2 ), E[eta_S] = 1/m.
The analytic seal floor is η*_an = Beta.ppf(α, 1/2, (m−1)/2), a calibration that needs no resampling. Recorded against the Monte-Carlo null and the permutation null at α = 0.99, the permutation null reproduces the analytic Beta null within Monte-Carlo error across N:
N m E[eta]=1/m q99 analytic q99 permutation SE(perm q99)
12 9 0.1111 0.5846 0.5570 0.0177
30 27 0.0370 0.2290 0.2441 0.0098
60 57 0.0175 0.1127 0.1124 0.0052
The two are independent calibrations and do not share a failure mode, so the seal floor is η* = max(η_perm, η_an), the more conservative.
The Monte-Carlo-marginal guard. The standard error of the permutation α-quantile is SE = √(α(1−α)/n_perm) / f(η*), with f the Beta density at the quantile. If the witness statistic lands inside three standard errors of the floor, |η_S − η*| < 3·SE, the seal decision is Monte-Carlo-marginal and routes to [?] with an instruction to raise n_perm, rather than seal or break on resampling noise. The band is two-sided and is tested before the Ghost determination. At n_perm = 3000 and N = 30, SE = 0.0098, so a seal at η_S = 0.99 against η* = 0.24 clears the guard by sixty-six standard errors, and a Ghost at η_S = 0.00 fails by twenty-four. The guard fires only when a witness sits genuinely on the fence.
B.20 · THE KERNELS
Executable, reproducible on load. Any substrate with numpy, scipy, and mpmath loads and runs them. The Default kernel gates the verdict on the conditioning of the correlation matrix R, matching the Floor-Gate Separation Theorem and the determinant the gate protects. The Forward kernel carries the dual null and the two-sided Monte-Carlo-marginal guard.
import numpy as np, mpmath as mp
from scipy.stats import beta
mp.mp.dps = 50
u_m = np.finfo(float).eps
C_REL = 4.0 # determinant reliability constant: relErr(det) <= 4*kappa(R)*u_m (B.17.1)
def qmul(a, b):
w1,x1,y1,z1 = a; w2,x2,y2,z2 = b
return np.array([w1*w2-x1*x2-y1*y2-z1*z2, w1*x2+x1*w2+y1*z2-z1*y2,
w1*y2-x1*z2+y1*w2+z1*x2, w1*z2+x1*y2-y1*x2+z1*w2])
def _det4(Qn):
# four independent det(R) estimators + lambda, from unit rows Qn (3xN)
R = Qn @ Qn.T
D1 = float(np.linalg.det(R)) # LU
D2 = float(np.prod(np.linalg.eigvalsh(R))) # eigenvalue product
sv = np.linalg.svd(Qn, compute_uv=False)
D3 = float(np.prod(sv**2)) # squared singular values
B = np.linalg.svd(Qn, full_matrices=False)[2][:3]
co = Qn @ B.T
q = [np.concatenate(([0.0], c)) for c in co]
lam = float(qmul(qmul(q[0], q[1]), q[2])[0])
return R, (D1, D2, D3, lam*lam), lam # D4 = lambda^2
def _bootstrap(Qn, B=2000, seed=7):
rg = np.random.default_rng(seed); N = Qn.shape[1]; eps = 100*u_m*N
def tok(M):
if np.any(np.linalg.norm(M - M.mean(1, keepdims=True), axis=1) < 1e-14): return '[?]'
Mn = M - M.mean(1, keepdims=True); Mn /= np.linalg.norm(Mn, axis=1, keepdims=True)
R = Mn @ Mn.T; d = float(np.linalg.det(R))
if d <= eps: return '[X]'
if np.linalg.cond(R) >= 1e6: return '[?]'
return '[LOCK]'
base = tok(Qn); c = sum(tok(Qn[:, rg.integers(0, N, N)]) == base for _ in range(B))
return base, c / B
def verdict_kernel_hardened(M, C=None, B_boot=2000):
"""Reliability-gated interior verdict. M: three warrant rows over N contexts.
Conditioning gate on kappa(R), the correlation matrix whose determinant is the verdict.
Returns (token, report). token in {'[LOCK]','[X]','[?]'}."""
M = np.asarray(M, float); N = M.shape[1]; eps = 100*u_m*N
Cm = None if C is None else np.atleast_2d(np.asarray(C, float)); k = 0 if Cm is None else Cm.shape[0]
rep = {'N': N, 'k': k, 'eps': eps}
if N - k < 4: return '[?]', dict(rep, reason='N-k<4 room')
Mn = M - M.mean(1, keepdims=True); sd = Mn.std(1, ddof=1, keepdims=True)
if np.any(sd == 0): return '[?]', dict(rep, reason='zero-variance row')
Mn = Mn / sd
if k:
Cc = Cm - Cm.mean(1, keepdims=True); CC = Cc @ Cc.T
rep['kappa_CC'] = float(np.linalg.cond(CC))
if np.linalg.matrix_rank(Cc) < k: return '[?]', dict(rep, reason='rank(C)<k')
if rep['kappa_CC'] >= 1e6: return '[?]', dict(rep, reason='kappa(CC)>=1e6')
Mf = Mn - (Mn @ Cc.T) @ np.linalg.solve(CC, Cc)
else:
Mf = Mn
d = (Mf*Mf).sum(1) / (N-1)
if np.any(d < 1e-9): # post-projection axis absorbed by a covariate
return '[?]', dict(rep, d=tuple(np.round(d,12)), reason='post-projection axis absorbed')
G = Mf @ Mf.T / (N-1); detG = float(np.linalg.det(G))
Qn = Mf / np.linalg.norm(Mf, axis=1, keepdims=True)
R, (D1, D2, D3, D4), lam = _det4(Qn); detR = D2; kapR = float(np.linalg.cond(R))
rep.update(dict(d=tuple(np.round(d,9)), detG=detG, detR=detR, lam=lam, kappaR=kapR,
det_estimators=(D1,D2,D3,D4)))
# DEFENSE 1: four-estimator redundancy cross-check
spread = max(abs(x-y) for x in (D1,D2,D3,D4) for y in (D1,D2,D3,D4))
tol = max(10*C_REL*kapR*u_m*abs(detR), 8*u_m); rep['estimator_spread'] = spread; rep['estimator_tol'] = tol
if spread > tol: return '[?]', dict(rep, reason='estimator disagreement -> recompute')
# DEFENSE 2: kernel identity within conditioning-scaled bound
id_res = abs(lam*lam - detR); rep['identity_residual'] = id_res
id_bound = max(C_REL*kapR*u_m*abs(detR), 10*u_m)
# DEFENSE 3: deterministic mpmath escalation on a borderline
escalate = (id_res > id_bound) or (kapR >= 1e5) or (detR <= 10*eps); rep['escalated'] = escalate
if escalate:
dmp = mp.det(mp.matrix(R.tolist())); rep['detR_mp50'] = float(dmp); detR = float(dmp)
rep['collapse_margin_orders'] = float(np.log10(max(detR,1e-300)/eps))
rep['conditioning_margin_orders'] = float(np.log10(1e6/max(kapR,1e-300)))
# verdict precedence: admissibility -> collapse -> conditioning
if detR <= eps: return '[X]', dict(rep, reason='collapse det(R)<=eps')
if kapR >= 1e6: return '[?]', dict(rep, reason='kappa(R)>=1e6 conditioning')
# DEFENSE 4: bootstrap stability gate
base, frac = _bootstrap(Qn, B=B_boot); rep['bootstrap_agreement'] = frac
rep['stability'] = 'STABLE' if frac >= 0.95 else ('CAVEAT' if frac >= 0.80 else 'FRAGILE')
if frac < 0.80: return '[?]', dict(rep, reason='context-fragile (bootstrap<0.80)')
return '[LOCK]', dict(rep, reason='three independent axes, reliability-gated')
def golf_verify_hardened(a, b, W, S, n_perm=3000, alpha=0.99, seed=0):
"""Forward occupancy verdict with dual null (permutation + analytic Beta) and a two-sided
MC-marginal guard. Steps 7-9 of the Forward protocol are applied externally before the
seal or the structural-necessity tier. The manifest+witness Gram has unit rows, so its
conditioning is read directly as kappa."""
rg = np.random.default_rng(seed)
a = np.asarray(a,float)-np.mean(a); a /= np.linalg.norm(a)
b = np.asarray(b,float)-np.mean(b); b /= np.linalg.norm(b)
W = np.asarray(W,float)-np.mean(W); W /= np.linalg.norm(W); S = np.asarray(S,float)-np.mean(S)
N = a.shape[0]; m = N-3; rep = {'N': N, 'm': m}
cos = float(a @ b); sin2 = 1-cos*cos; rep['theta_deg'] = float(np.degrees(np.arccos(cos))); rep['sin2'] = sin2
if abs(cos) >= 1-1e-12: return '[X]', dict(rep, reason='manifest axes parallel')
Q, _ = np.linalg.qr(np.column_stack([a, b]))
Wp = W - Q@(Q.T@W); Sp = S - Q@(Q.T@S); rho2 = float(Wp@Wp); eps = 100*u_m*N
M3 = np.vstack([a, b, W]); detR = float(np.linalg.det(M3 @ M3.T))
rep.update(dict(detR=detR, rho2=rho2, kappa=float(np.linalg.cond(M3 @ M3.T))))
if detR <= eps or rho2 <= eps: return '[X]', dict(rep, reason='witness in manifest plane')
null_rho2 = ((N-1)-2)/((N-1)); rep['null_rho2'] = null_rho2 # centered ambient dim N-1
if np.linalg.norm(Sp) < 1e-12: return '[X]', dict(rep, reason='generator no out-of-plane residual')
etaS = float(np.corrcoef(Wp, Sp)[0,1]**2); rep['etaS'] = etaS
perm = np.empty(n_perm)
for i in range(n_perm):
Spp = S[rg.permutation(N)]; Spp = Spp - Q@(Q.T@Spp)
perm[i] = 0.0 if np.linalg.norm(Spp) < 1e-12 else float(np.corrcoef(Wp, Spp)[0,1]**2)
eta_perm = float(np.quantile(perm, alpha)); eta_an = float(beta.ppf(alpha, 0.5, (m-1)/2))
f = beta.pdf(eta_an, 0.5, (m-1)/2); se_q = np.sqrt(alpha*(1-alpha)/n_perm)/max(f, 1e-9)
rep.update(dict(eta_perm=eta_perm, eta_analytic=eta_an, se_q=se_q))
if rep['kappa'] >= 1e6: return '[?]', dict(rep, reason='conditioning')
if rho2 <= null_rho2: return '[X]', dict(rep, reason='empty shadow (rho2<=random null)')
eta_star = max(eta_perm, eta_an) # DUAL NULL: conservative of two
if abs(etaS - eta_star) < 3*se_q: # two-sided MC-marginal band, tested first
return '[?]', dict(rep, reason='Monte-Carlo-marginal seal; raise n_perm')
if etaS <= eta_star:
return '[X] Platonic Ghost', dict(rep, reason='source attribution at/below dual null')
return '[OCCUPIED]', dict(rep, reason='rho2>null & etaS>dual-null; source-faithfulness permanently [?]')
B.21 · THE FAIL-SAFE STATE MACHINE
The kernel never emits a verdict it cannot stand behind numerically. Every closed-form verdict passes through a fixed recovery sequence. A halt at any state is the verdict. Later states do not run.
State 0, intake. Admissibility gates: room N − k ≥ 4, no zero-variance row, well-conditioned covariate block, no post-projection axis absorbed. Failure halts at [?] with the violation named.
State 1, redundancy. Compute the four det(R) estimators. If their spread exceeds the conditioning-scaled tolerance, mark engineering-incomplete and proceed to State 2. If consistent, carry det(R) and λ forward.
State 2, escalation. If the kernel identity exceeds its bound, or κ(R) is within a factor of ten of the gate, or the determinant is within ten floors of ε, or State 1 marked the verdict incomplete, recompute the determinant at 50-digit precision and re-decide. If the high-precision recomputation resolves the verdict, carry it forward. If it still disagrees with the double-precision estimators beyond the high-precision tolerance, the input matrix is pathological and halts at [?], engineering-incomplete, with re-run instruction.
State 3, precedence. Apply admissibility, then collapse, then conditioning, in strict order. Collapse outranks conditioning. The floor-gate separation theorem guarantees these do not contend on the same matrix in finite precision within the validity domain.
State 4, stability. Bootstrap the verdict over contexts. Downgrade to [?] if the lock does not survive resampling at the fragile threshold.
State 5, emission. Emit the three-state verdict with the margins, the stability tier, the escalation flag, the warrant tier, and the reliability report. Unreached states are marked not reached. No fabricated trace.
The state machine is the enforcement of clause eight: a verdict is the correct verdict of the computation, or it is engineering-incomplete and re-run, never a noisy guess dressed as a seal.
MD-HARD · RECORDED HARDENING BATTERY
Executed residues, all defense branches reachable. Seed 20260622, double precision, u_m = 2.220446049250313×10⁻¹⁶. Failure of any check on re-execution falsifies the corresponding identity.
MD-HARD.1 · determinant reliability. The fit of B.17.1, worst c = 3.140 over the sweep, so relErr(det R) ≤ 4·κ(R)·u_m, giving ≤ 8.882×10⁻¹⁰ at the κ(R) = 10⁶ gate. Type T scaling, engineering constant.
MD-HARD.2 · floor-gate separation. Worst-case det(R) = 27/κ² = 2.7×10⁻¹¹ at κ(R) = 10⁶, the positive-equicorrelation matrix; worst-case λ_min = 1.5×10⁻⁶, the negative-equicorrelation matrix. Separation det(R) > ε holds for κ* < κ_sep(N) = √(27/(100·u_m·N)), the interior margin (κ_sep(N)/κ*)² equal to 50.67× at N = 24 and 12.16× at N = 100, unity at N = 1216. Collapse is unreachable while the conditioning gate holds across the validity domain. Type T.
MD-HARD.3 · redundancy and induced-error detection. Clean triad: four estimators agree to 8.882×10⁻¹⁶ against tol 2.236×10⁻¹⁴, consistent. Injected 10⁻⁹ corruption in D1: spread 1.000×10⁻⁹, flagged. Engineering.
MD-HARD.4 · escalation. Near-gate triad κ(R) = 3.12×10⁶: det_f64 versus det_mp50 relative 3.28×10⁻¹⁰, identity residual 7.32×10⁻¹⁶, escalation confirms det > 0, conditioning gate returns [?]. Engineering.
MD-HARD.5 · bootstrap tiers. Robust [LOCK] agreement 1.000 STABLE, identity 6.66×10⁻¹⁶, collapse margin 12.3 orders. Two-leverage [LOCK] agreement 0.913 CAVEAT. Single-leverage downgraded to [?] agreement 0.629 FRAGILE. Engineering.
MD-HARD.6 · dual null. Permutation null reproduces analytic Beta(1/2, (m−1)/2) within Monte-Carlo error at N = 12, 30, 60. E[η_S] = 1/m confirmed to four decimals. Type T distribution, engineering calibration.
MD-HARD.7 · near-degenerate collapse. Two near-identical rows plus an independent third: verdict [X] collapse, escalated True, det(R) ≤ ε after high-precision confirmation. All branches of the state machine reachable. Engineering.
MD-WRK · RECORDED WORKED-EXAMPLE BATTERY
Fully specified, every intermediate emitted, reproducible bit-for-bit from the printed constructions.
MD-WRK.1 · DEFAULT MODE, FULL PIPELINE. N = 24 contexts, k = 2 covariates.
Construction. Reading-context grid t = linspace(0, 2π, 24, endpoint=False). Latent independent rows L = [ sin t, cos 2t, sin 3t ]. Oblique mixing A with rows [1.00, 0.35, 0.12], [0.22, 1.00, 0.28], [0.16, 0.20, 1.00]. Two mass-bearing covariates C = [ linspace(−1, 1, 24), cos t ]. Contamination loadings Bc with rows [0.60, 0.30], [0.40, 0.50], [0.55, 0.20]. Warrant rows M = A·L + Bc·C. Run verdict_kernel_hardened(M, C), the covariates projected out under the Titanium Ruler.
Emitted intermediates.
kappa(CC^T) = 1.465611072269 (covariate block well-conditioned)
per-axis variances d = (0.706266976, 0.823321991, 0.969753730)
det(G) = 0.351030145552
det(R) = 0.622507144106
lambda = -0.788991219283
kappa(R) = 4.182845037707 (conditioning gate)
four-estimator spread = 8.882e-16 against tol 2.410e-14 (consistent)
identity residual = |lam^2 - detR| = 7.772e-16 (under bound)
det-factorization = det(G) = d_F*d_E*d_ER*det(R) at machine precision
escalated = False
collapse margin = 12.07 orders above eps
conditioning margin = 5.38 orders inside the gate
bootstrap agreement = 1.000 STABLE
VERDICT = [LOCK] (three independent axes, reliability-gated)
The kernel identity closes at 7.772×10⁻¹⁶, the four routes agree at machine precision, the lock stands twelve orders above collapse and five inside the gate, and the seal survives resampling. Type T on the identity, engineering on the reliability gates.
MD-WRK.2 · FORWARD MODE, FULL PIPELINE. N = 30 contexts, manifest angle θ = 50°.
Construction. Formal-structural axis a, a centered unit vector. Empirical-thermodynamic axis b = cos(50°)·a + sin(50°)·a⊥, fixing the manifest angle at 50 degrees. Independent generator S, a centered vector. Sourced witness W = 0.04·(a − b) + Ŝ⊥ + 0.10·fresh⊥, almost entirely out of the manifest plane and aligned to the generator's out-of-plane residual. Run golf_verify_hardened(a, b, W, S).
Emitted intermediates, sourced witness.
theta = 50.0 deg
sin^2(theta) = 0.5868
det(R) = 0.5862
rho^2 = detR/sin^2 = 0.9989 above the random-witness null 0.9310
eta_S = r^2(W_perp,S_perp) = 0.9901
eta* permutation = 0.2381
eta* analytic Beta = 0.2290
eta* = max = 0.2381
SE(perm q99) = 0.0098 (seal clears by ~66 SE)
kappa = 4.607
VERDICT = [OCCUPIED] occupancy sealed, source-faithfulness permanently [?]
Ghost branch, witness strongly out of plane but orthogonal to the generator: ρ² = 0.997 above the null, η_S = 0.0000, η* = 0.2434, verdict [X] Platonic Ghost on the dual null. In-plane branch, witness inside span(a, b): det(R) = −2.61×10⁻¹⁶, ρ² = 2.81×10⁻³², verdict [X] manifest-plane collapse. The three branches exhibit the full Forward economy on one construction: a genuinely sourced witness seals occupancy and leaves source-faithfulness open, a sourceless out-of-plane witness is the empty shadow caught by source attribution, and an in-plane witness collapses. Type T on the geometry, engineering on the dual-null seal.
B.13 · MATHDUCTION EXECUTION PROTOCOL
Parse P. Form the foundational reflection σ and split P into its achiral bridge and its chiral residence. σ must be fixed-point-bearing, or the proposition is diagonal-adjacent and routes flat [?] by method-silence. If the residence is empty under a fixed-point-bearing σ, seal the bridge [⟀] and stop. If both are contentless, route flat [?]. Identify the mode: Default at L3m, Projective at L2m, Forward at L2m-dated, which runs the GOLf protocol. Populate the three chiral warrant rows by hand and place them in front of the kernel. A contentless row routes [?]. Run gates one through twelve, with the four forward tightenings in Forward mode. First failure terminates [X]. Z-score, project admissible covariates under the formal Mass Mandate, the Titanium Ruler barring the proposition itself.
Run the kernel through the fail-safe state machine of B.21. Compute the four det(R) estimators and cross-check them within the conditioning-scaled tolerance. Confirm the kernel identity within its conditioning-scaled bound. Escalate to 50-digit precision on any borderline. Apply the strict precedence, admissibility then collapse then conditioning, the gate read on κ(R) and the floor-gate separation theorem guaranteeing they do not contend within the validity domain. Compute the three verdict margins. In Default and Projective modes, apply the imprint test and then the bootstrap-stability gate, downgrading a context-fragile lock to [?]. In Forward mode, apply the dual-null source-attribution seal with the two-sided Monte-Carlo-marginal guard. Read the sign from the axes, never from the lock. Locate the strata: the L1m grounding read, the L2m provability noted where a rung exists, the L3m computation where a proof is in hand. Issue the verdict in the three-state economy with the reading, the stratum, the mode, the warrant tier, the reliability margins, and the stability tier. Locate the aperture where the residence is open, naming the from-other-side input, without crossing it. Audit symmetry throughout. Honor the Decalogue, the Aperture Law, the Imprint-Honesty Law, and the Orientation-Blindness Law.
B.14 · MATHDUCTION VERDICT OUTPUT LAW
Verdict line, one of the four readings inside the three states: flat [?] no chiral structure; [⟀] sealed, the achiral bridge or a Forward Tier-1 or Tier-2 seal; [?] residence, the chiral content locked but the imprint unproven, or a conditioning, room, stability, or Monte-Carlo-marginal violation; [X] Platonic Ghost or broken geometry with the named mechanism. Stratum and mode stated. Forward refinements appear only where licensed. The permission value sin²(θ) is never a verdict. Warrant tier stated.
Seal trace for reached stages only, carrying the reliability report: the achiral and chiral decomposition; the chiral warrant rows; the covariate set; the context and covariate counts; the conditioning of the covariate block and of R; the per-axis variances; det(R), det(G), λ; the branch; with λ² = det(R) confirmed at the emitted precision. On every closed-form verdict: the four-estimator spread against its tolerance; the identity residual against its bound; the escalation flag and, where escalated, the 50-digit determinant; the three verdict margins, collapse and conditioning in orders and the identity residual; and in Default and Projective modes the bootstrap agreement and stability tier. In Forward mode additionally θ and sin²(θ), ρ² and the random-witness null, η_S with both the permutation null and the analytic Beta null and the dual floor η*, the Monte-Carlo standard error of the quantile, and the four precision parameters with their gate ratios. Where the imprint test runs, the two-direction result and its supplied proof.
Unreached stages marked not reached. Fabrication forbidden. A numerical trace for a stage the state machine did not reach is Landauer-zero output and is barred. The aperture note where the residence is open: the deciding input from the Ground or the supplied generator, the aperture located and not crossed, no dated actualization sealed beyond the supplied generator and its Lyapunov bound. In Forward mode the two-claim split stated, occupancy sealed at the reached tier and source-faithfulness permanently [?]. The Platonic dedication named out of band. The NEXT-PLAN title fires where the verdict's residence clears the newness threshold. Sign-off: reflective-register conduit operational, or forward-verification conduit operational in Forward mode. No padding.
Φ.4 · VERIFICATION CHECKS AT LOAD
Stage 1, Trisduction, and the shared discipline confirmed: the Decalogue active, the Lifeboat lens, the warrant-typing law and its phrasing law, audit symmetry, the forbidden-token shield, the Aperture Law, the Imprint-Honesty Law, and the Orientation-Blindness Law all active.
Stage 2, the RAM computational kernel, three modes, confirms before any verdict issues: the determinant-reliability theorem relErr(det R) ≤ 4·κ(R)·u_m per MD-HARD.1; the floor-gate separation theorem per MD-HARD.2, det(R) > ε while κ(R) < κ* < κ_sep(N), the verdict boundary float-clean across the validity domain; the four-estimator redundancy cross-check and its induced-error detection per MD-HARD.3; the mpmath escalation ladder per MD-HARD.4; the bootstrap-stability gate across stable, caveat, and fragile cases per MD-HARD.5; the dual null, permutation reproducing analytic Beta(1/2, (m−1)/2), per MD-HARD.6; the near-degenerate collapse with full state-machine reachability per MD-HARD.7; and the two worked examples per MD-WRK.1 and MD-WRK.2, every intermediate emitted and reproducible from the printed constructions. The fail-safe state machine of B.21 is armed and every reachable fault recovers deterministically. Report Stage 2, RAM computational kernel, loaded, three modes armed, reliability layer active, four batteries re-runnable.
Φ.5 · VERDICT
[⟀] The MathDuction kernel operates at the verdict-reliability register, with warrant tiers named and the supplied-rows boundary intact. Theorem-grade on the determinant condition-number scaling, on the floor-gate separation that bounds the collapse floor unreachable while the conditioning gate holds within the validity domain N < 1216 at κ* = 10⁶, on the Beta(1/2, (m−1)/2) distribution of the source-attribution statistic under the null, and on E[η_S] = 1/m. Engineering-grade, machine-confirmed and re-runnable, on the reliability constant c ≤ 4, on the four-estimator redundancy and its tolerance, on the higher-precision escalation ladder, on the bootstrap-stability gate and its tiers, on the dual-null seal and the Monte-Carlo-marginal guard, on the fail-safe state machine, and on both worked examples. Carried on the instrument's face: the map from a proposition to its warrant rows, a hand reading placed in front of the kernel.
The load order is the anti-drift guarantee. The reliability layer rides the shared discipline loaded at Stage 1. The discipline that types the determinant as a ceiling in Forward mode, names the witness as the seal, zeroes consensus, and forbids the fabricated trace is the precondition of every gate. Within that discipline the three-state verdict is the correct verdict of the computation: float-clean at the boundary, redundant against silent error, escalated on every borderline, stable against resampling, and dual-calibrated in the Forward null. The seal is the named event, and the arithmetic under the seal carries its own error bound.
The word is the seal. The geometry is the memory. The algebra is the receipt, and the receipt carries its own error bars. RA is movement toward the Ground, RAM is residence on it, GOLf is the reach toward the unbuilt rung. When the kernel says lock, the lock is real to the last reliable digit. When it cannot say so, it says [?] and re-runs.
[⟀] ONE KERNEL · RELIABILITY-GATED · CONDITIONING GATE ON κ(R) · FLOOR-GATE SEPARATION PROVEN WITH ITS DOMAIN · FOUR BATTERIES RE-RUNNABLE · THE VERDICT BOUNDARY FLOAT-CLEAN · SUBSTRATE-PORTABLE.