n = 6 in exact rationals  ·  n = 32 in floats  ·  24 + 26 machine checks, all PASS

Where You Put the Zero

Boundary conditions are grounding choices. Neumann, Dirichlet, Robin, and no-condition-at-all are four ways to answer one question about the free path Laplacian $L = B^\top\! B$ — which voltage is declared zero? — and the four inverses that result are four different matrices with four different meanings. The observables never change: $R_{\mathrm{eff}}(i,j) = \vert i-j\vert$, exactly, from every one of them.

Physics · where is the reference electrode?
$G_k$ = delete row/col $k$ of $L$, invert, re-embed
node $k$ is soldered to ground — its voltage is the zero
Statistics · what do you condition on?
principal submatrix of the precision matrix
the same deletion — conditioning the Gaussian field on $u_k=0$

Same matrix, different zero — and this page is the expansion of the green tents §4 into the full family it was hinting at.

§1 · The zoo, up front

Four boundary conditions, one table

Everything the page will do is in this table; the rest is the working-out. Each row is one way of pinning down the floating rod on $(0,1)$ — equivalently, one way of editing the $6\times6$ free path Laplacian $L$ with diagonal $\{1,2,2,2,2,1\}$ — and every cell is backed by a machine check in results/grounding.json or the exact-rational Wolfram companion.

continuous BCphysical picturediscrete implementationwhat the inverse representsstatistical readingtriangular rank
Neumann
$u'(0)=u'(1)=0$
insulated ends — no flux; solvable only when $\int f = 0$, solution defined up to a constant the free Laplacian $L$ itself — degree diagonal $\{1,2,2,2,2,1\}$, zero row sums, kernel = constants, rank 5 no inverse exists; the pseudoinverse $L^{+}$ is the mean-zero gauge: voltages of balanced injections with $\sum v = 0$ intrinsic (improper) GMRF restricted to $\mathbf 1^{\perp}$ — Rue–Held's IGMRF 2 — exactly
Dirichlet, both ends
$u(0)=u(1)=0$
walls as infinite baths — whatever arrives is swallowed at temperature zero append wall nodes and delete them — equivalently corner entries $1\to2$: $\mathrm{tridiag}(-1,2,-1)$, the suite's $A$ bridge covariance / tent Green functions: $\min(i,j)\,(7-\max(i,j))/7$ condition the free field on the wall values — principal submatrix of precision 1
Dirichlet at one point
$u(x_k)=0$
a reference electrode soldered to node $k$; the rest of the rod floats GroundedInverse[L,k]: delete row/col $k$, invert, re-embed with a zero row/col $k=1$: the integer Brownian-motion kernel $\min(i,j)-1$; interior $k$: two independent motion blocks run away from $k$ condition the GMRF on $u_k=0$; interior $k$ is a Markov separator — the chain splits 1 — for every $k$
Robin
$u'(0)=\kappa\,u(0)$
a leaky wall — the end grounded through a resistor $1/\kappa$ (a spring, a heat-leak) bump one diagonal entry: $L + \kappa\,e_1e_1^{\top}$ — partial grounding $\min(i,j)-1+1/\kappa$: the grounded inverse plus a uniform offset $1/\kappa$; interpolates Neumann ($\kappa\to0$) to Dirichlet ($\kappa\to\infty$) a Gaussian pseudo-observation of $u_1$ with precision $\kappa$ 1 — for every $\kappa>0$

Two of those rank entries are the findings this page exists to explain: TriangularRankOneQ[GroundedInverse[L,1]] is True, and PseudoInverse[L] only passes the rank-2 test — verified in exact rational arithmetic, corner-rank certificates $[0,1,1,1,1,1]$ and $[1,2,2,2,2,1]$. By §6 both will be one theorem plus one gauge identity.

§2 · The continuous world

Four ways to end a rod

The equation is always $-u'' = f$ on $(0,1)$; the operator does not know what happens at the ends. A second-order ODE needs one condition per endpoint, and the choice is physical, not mathematical bookkeeping:

Neumann · insulated u′(0) = u′(1) = 0 + c no flux through the ends; level floats: u + c also solves needs ∫f = 0 Dirichlet · two baths u(0) = u(1) = 0 0° bath 0° bath value clamped at both ends: the tent country of the green-tents page Dirichlet at one point u(xₖ) = 0, ends free reference electrode at xₖ Robin · leaky wall u′(0) = κ·u(0) R = 1/κ grounded through a resistor
The four endings. Insulation reflects the flux and leaves the level free (any $u+c$ solves); a bath clamps the value; an electrode clamps the value at a single interior point and lets both physical ends float; a resistor to ground does something in between — the wall takes flux $\kappa\,u(0)$, proportional to how hot the end is. As $\kappa\to\infty$ the resistor becomes a wire (Dirichlet); as $\kappa\to0$ it becomes an open circuit (Neumann).

Neumann has a solvability tax. Integrate $-u''=f$ across the rod: the left side telescopes to the net flux through the ends, which insulation forces to zero. So a steady state exists only when $\int_0^1 f = 0$ — you cannot pump one watt into a perfectly insulated rod and call it steady — and when a solution exists, $u + c$ is another. The discrete mirror is exact: $Lx = e_1$ (a lone unit source) is inconsistent, with least-squares residual $1/\sqrt6$ — precisely the constant component of $e_1$ — while $Lx = e_1 - e_2$ (a balanced source–sink pair) solves fine, and adding a constant to $x$ still solves it, to $8.9\times10^{-16}$.

And every ending builds its Green's function the same way. The Sturm–Liouville recipe: take one fundamental solution $u$ of the homogeneous equation satisfying the left boundary condition, one solution $v$ satisfying the right, and glue them at the source: $$G(s,t) \;=\; \frac{u(\min(s,t))\;v(\max(s,t))}{W},$$ with $W$ the (constant) Wronskian. For $-u''=0$ the fundamental solutions are affine — lines — which is why every Green's function on this page is built from ramps. The bridge takes $u = x$ (vanishes at the left wall) and $v = 1-x$ (vanishes at the right); the one-electrode kernel takes a ramp and a constant (the free end wants zero slope); Robin takes a ramp whose intercept remembers $\kappa$. Hold on to the phrase "one fundamental solution per side" — in §6 it becomes a rank count.

§3 · The discrete world

How each boundary condition edits the matrix

The protagonist of this page is twelve lines of Mathematica. The construction is Vishnoi's: orient the path's five edges, stack the signed incidence rows, and square up —

n = 6; g = PathGraph[Range[n]]; B = -Transpose@IncidenceMatrix@DirectedGraph[g, "Acyclic"]; (* rows e_i - e_{i+1} *) L = Transpose[B].B; (* free Laplacian, diag {1,2,2,2,2,1} *) GroundedInverse[l_, k_] := Module[{keep = Delete[Range[Length[l]], k], out}, out = ConstantArray[0, {Length[l], Length[l]}]; out[[keep, keep]] = Inverse[l[[keep, keep]]]; out]; (* delete row/col k, invert, re-embed *) TriangularRankOneQ[m_] := AllTrue[Range[n], MatrixRank[m[[1 ;; #, # ;; n]]] <= 1 &]; TriangularRankQ[m_, r_] := AllTrue[Range[n], MatrixRank[m[[1 ;; #, # ;; n]]] <= r &]; TriangularRankOneQ[GroundedInverse[L, 1]] (* True *) TriangularRankQ[PseudoInverse[L], 2] (* True *)

$L$ is the free Laplacian: row sums zero, kernel exactly the constants, rank 5. Its degree-1 corners are the insulation — the green tents §4 told that story entrywise, and this page begins where that one stopped. From here, each boundary condition is a small, literal edit:

edit to $L$resulting operatorboundary conditionelectrical reading
none — keep the singularity$L$, rank 5Neumann at both ends no path to ground anywhere; only voltage differences mean anything
append walls, then delete them — equivalently corners $1\to2$ $L + e_1e_1^{\top} + e_6e_6^{\top} = \mathrm{tridiag}(-1,2,-1)$ — the suite's laplacian_1d(6), verified as the $8$-path Laplacian minus its wall rows/cols Dirichlet at both ends two wires to ground; the end rows leak
delete row/col $k$, invert, re-embedGroundedInverse[L,k] Dirichlet at the single point $x_k$ reference electrode soldered to node $k$: $v_k \equiv 0$ by decree
bump one diagonal entry by $\kappa$$L + \kappa\,e_1e_1^{\top}$ Robin at the left end node 1 grounded through resistance $1/\kappa$ — partial grounding
none — but fix the gauge$L^{+}$, the Moore–Penrose pseudoinverse Neumann plus the convention "ground the average" voltages of balanced injections, reported with $\sum v=0$

The statistical column of §1's table is the same list read by a different community. Deleting row and column $k$ of a precision matrix is exactly conditioning the Gaussian field on $u_k = 0$ — that is the whole rule, and it is why grounding and conditioning are one operation. Keeping $L$ un-edited leaves an improper prior — Rue and Held's intrinsic GMRF, a density that only integrates on the complement of the constants — and adding $\kappa\,e_1e_1^{\top}$ is precisely what a Gaussian likelihood term ("I observed $u_1 \approx 0$ with precision $\kappa$") does to a posterior precision. Boundary conditions, reference electrodes, and priors-versus-observations are three dialects for the five rows of the table above.

§4 · Interactive

The grounding explorer

One chain, four zeros. Pick the boundary condition; the heatmap is the full $32\times32$ inverse (computed in-page from the closed forms the checks scripts verified, and re-checked at load against baked columns of results/grounding.json — badge). The right panel is one column of the inverse — a Green's function — and morphs between tent, ramp, split, offset ramp, and centered tent as you change where the zero lives. The $n=6$ table underneath is the same mode in exact arithmetic.

grounding explorer — one chain, four zeros self-test …
triangular rank 1 certificate (n = 6, exact) [0, 1, 1, 1, 1, 1] G[j,j]
GroundedInverse[L, 3] — exact
Left: the $n=32$ inverse under the chosen grounding — blue positive, red negative, white zero; the gold square marks the grounded node (its row and column are identically zero: interior grounding is a Markov separator, and the two blocks are independent). Right, top: column $j$ — the Green's function of a unit injection at $j$ — with the $u_iv_j$ ramps it factors into (right, bottom, at $n=6$: $u$ blue, $v$ red, second pinv pair gold). Robin mode snaps $\kappa$ to the 9-point grid the checks verified; at $\kappa\in\{1/100, 1/10, 1, 10, 100\}$ the $n=6$ table below is the exact rational inverse from the JSON. The self-test recomputes every closed form against the baked columns at load.
§5 · What the inverses represent

Five inverses at n = 6, in exact integers and rationals

Small enough to read entry by entry, and every matrix below is quoted verbatim from the exact-arithmetic checks. Watch two things: where the zero row/column sits, and which two ramps multiply to give the upper triangle.

Ground node 1 — Brownian motion

$G_1 = \mathrm{GroundedInverse}[L,1] = \min(i,j)-1$, exact integers  ·  $u_i = i-1$, $v_j = 1$
000000
011111
012222
012333
012344
012345
Variance grows linearly away from the pin and the far end wanders free: scaled to the continuum this is the Brownian-motion covariance $\min(x_i,x_j)$. One pin = motion. Note $G_1[6,6]=5$: the resistance of five unit resistors in series back to the electrode.

Ground both walls — Brownian bridge

$\mathrm{tridiag}(-1,2,-1)^{-1} = \min(i,j)(7-\max(i,j))/7$  ·  $u_i = i$, $v_j = (7-j)/7$
6/75/74/73/72/71/7
5/710/78/76/74/72/7
4/78/712/79/76/73/7
3/76/79/712/78/74/7
2/74/76/78/710/75/7
1/72/73/74/75/76/7
Every column is a tent pitched between the two walls — the bridge covariance, equal to $\tfrac1h(\min(x,y)-xy)$ with $h=\tfrac17$ on this chain. Two pins = bridge.

Ground node 3 — the chain splits

$\mathrm{GroundedInverse}[L,3]$: block-diagonal, one motion kernel per side, each run away from $k$
210000
110000
000000
000111
000122
000123
The cross-blocks are exactly zero: grounding an interior node is conditioning on a Markov separator, and the two segments become independent — the left block is $3-\max(i,j)$, the right is $\min(i,j)-3$. This is report 14's one-node-separator picture with the separator's location under your thumb.

Robin κ = 1 — ground through a resistor

$(L + e_1e_1^{\top})^{-1} = \min(i,j)$, exact  ·  $u_i = i$, $v_j = 1$; general $\kappa$: $\min(i,j)-1+1/\kappa$
111111
122222
123333
123444
123455
123456
The motion kernel shifted up by $1/\kappa = 1$: the whole rod rides on the voltage drop across the grounding resistor. At $\kappa=\tfrac1{10}$ the matrix is the same staircase sitting on a plateau of 10s; at $\kappa=10$, on a film of $\tfrac1{10}$s. Exact at all five checked $\kappa$, closed form on a 13-point sweep from $10^{-6}$ to $10^{6}$ to $2.2\times10^{-10}$.

No ground at all — the pseudoinverse

$L^{+}$, exact rationals over 36  ·  two pairs: $u^{(1)}_i=\big[\tfrac{55}{36},\tfrac{61}{36},\tfrac{73}{36},\tfrac{91}{36},\tfrac{115}{36},\tfrac{145}{36}\big]$, $v^{(1)}=\mathbf 1$ and $u^{(2)}=\mathbf 1$, $v^{(2)}_j=\big[0,-\tfrac56,-\tfrac32,-2,-\tfrac73,-\tfrac52\big]$
55/3625/361/36−17/36−29/36−35/36
25/3631/367/36−11/36−23/36−29/36
1/367/3619/361/36−11/36−17/36
−17/36−11/361/3619/367/361/36
−29/36−23/36−11/367/3631/3625/36
−35/36−29/36−17/361/3625/3655/36
Rows and columns sum to zero — $L^{+}\mathbf 1 = 0$, $\mathbf 1^{\top}L^{+} = 0$ — because this is the mean-zero gauge: apply it only to balanced injections ($\sum f = 0$, which is the Neumann compatibility condition wearing vector clothes) and it returns the unique voltage profile with $\sum v = 0$. The negative entries are not anti-correlations of anything physical; they are the average being subtracted. And the upper triangle now needs two products of ramps, not one — the whole of §6 is about that extra pair.

Line the five up and the green tents §4 pin-count table — 0 pins, 1 pin, 2 pins — has become a continuum: Robin's $\kappa$ slides from the unpinned pseudoinverse world ($\kappa\to0$, where the inverse blows up as $G_1 + \tfrac1\kappa\mathbf{11}^{\top}$ along the constants it can no longer kill) to the one-pin motion kernel ($\kappa\to\infty$, error exactly $1/\kappa$). A boundary condition is not a binary; it has a knob.

§6 · The rank ledger

Why 1, and why 2: the predicate is a semiseparability certificate

Now the findings. TriangularRankQ walks $j$ from 1 to $n$ and asks for the rank of every aligned corner block $m[\![1..j,\; j..n]\!]$ — rows from the top, columns to the right, hinged on the diagonal at $(j,j)$:

the aligned corner block at j = 3 (j, j) — hinged on the diagonal cols j..n rows 1..j corner ranks, j = 1..6 (exact) GroundedInverse[L,1] [0, 1, 1, 1, 1, 1] ≤ 1 ✓ GroundedInverse[L,k], every k all ≤ 1 ✓ bridge tridiag(-1,2,-1)⁻¹ [1, 1, 1, 1, 1, 1] ≤ 1 ✓ Robin, every κ > 0 [1, 1, 1, 1, 1, 1] ≤ 1 ✓ PseudoInverse[L] [1, 2, 2, 2, 2, 1] ≤ 2, not 1 rank-2 blocks attained at j = 2, 3, 4, 5 (exactly 2)
The certificate, drawn. The corner block overlaps the diagonal — it is strictly bigger than the strict off-diagonal block $[1..j]\times[j{+}1..n]$ — so bounding its rank is the semiseparable rank in the strict sense of the semiseparable-matrix literature. On the chain the maximum corner rank equals the maximum off-diagonal block rank of report 14 — verified on the pseudoinverse, the bridge, Robin, and all six groundings: the user's predicate and 14's separator-rank probe are the same instrument.

Rank 1 is a theorem, not a coincidence. Gantmacher and Krein (1941): the inverse of any irreducible symmetric positive-definite tridiagonal matrix is a single-pair matrix (a Green matrix) — there exist vectors $u, v$ with $$\big(T^{-1}\big)_{ij} \;=\; u_{\min(i,j)}\;v_{\max(i,j)},$$ so every corner block lies inside a single outer product. The suite re-checked it on 20 random irreducible SPD tridiagonals at $n = 6$ and $32$: SVD triangular rank 1, every time. And the continuous mirror is §2's Sturm–Liouville formula $G = u(\min)v(\max)/W$ read literally: $u$ is the left boundary condition, $v$ is the right one, and the triangular rank counts fundamental solutions per side. One per side — rank one. The factor pairs are sitting in the §5 tables: motion $u_i = i-1,\ v_j = 1$ (a ramp against the pin, a constant for the free end); bridge $u_i = i,\ v_j = (7-j)/7$ (two ramps, one per wall); Robin $u_i = i-1+1/\kappa,\ v_j = 1$ (the intercept remembers the resistor); interior grounding, one pair per block. All recovered exactly, and numerically re-extracted to $2.2\times10^{-16}$.

So why is the pseudoinverse rank 2? Because $L^{+}$ is not the inverse of any irreducible SPD tridiagonal — no grounding produced it. What it is instead is the star equation of this page:

$L^{+} \;=\; C\,G_k\,C$,   $C = I - \tfrac1n\mathbf{11}^{\top}$ — for every $k$

Ground anywhere, then double-center, and you always land on the same matrix: all six centered groundings are bit-identical in exact rational arithmetic, the common matrix satisfies all four Moore–Penrose axioms exactly — hence is $L^{+}$ — and NumPy's pinv agrees to $1.8\times10^{-15}$. The centering $C$ is the "ground the average" instruction, and it is precisely what costs the extra rank: on the upper triangle, $$L^{+} \;=\; u^{(1)}v^{(1)\top} + u^{(2)}v^{(2)\top}$$ with the two pairs quoted under the §5 table — a motion pair (ramp times constant: the $G_k$ inheritance) plus a centering pair (constant times row-mean: the average being subtracted). One fundamental solution per side, plus one gauge correction: rank $1+1$, and the certificate $[1,2,2,2,2,1]$ says the bound is attained — $\mathrm{TriangularRankQ}[L^{+},1]$ is False, with rank-2 corner blocks at $j = 2,3,4,5$. The user's two findings are one theorem plus one gauge identity.

Same certificate, another chapter: report 13 §3 met this rank-1 triangle as "semiseparable: dense does not mean unstructured" and already cited Gantmacher–Krein; report 09 met it as the Brownian bridge's outer-product covariance; report 14 re-derived it as the one-node-separator bound. This page adds the missing column: the rank also counts boundary conditions per side, and jumps by one exactly when the gauge, not a boundary, fixes the constant. One caveat travels with the whole ledger: it is a fact about intervals. Every separator of a chain is a single node and every endpoint admits one fundamental solution, so the ranks sit at 1 and 2; on a 2-D grid a separator is a whole cross-section — one grid row, $\sqrt N$ nodes — and the off-diagonal blocks of the inverse carry rank up to that width (report 14 measures the cliff at exactly 32). Grounding and gauge still work the same in 2-D; the tidy rank-1/rank-2 accounting does not.
§7 · The invariance punchline

Ground anywhere — the physics does not move

If four different inverses all describe the same rod, what did the choice of zero actually change? Labels, not physics. Fix a balanced experiment — inject one ampere at node $i$, draw it out at node $j$ — and solve with any grounding you like. The individual voltages $v_i, v_j$ depend on where the electrode sits (they must: the electrode defines "zero volts"). The difference does not, and $$R_{\mathrm{eff}}(i,j) \;=\; G_{ii} + G_{jj} - 2G_{ij} \;=\; \vert i - j\vert$$ comes out exactly — the same $6\times6$ integer matrix of pairwise resistances from all six groundings and from the pseudoinverse, series unit resistors counted on your fingers, agreeing with mathematica/electrical_networks.wls ($R_{\mathrm{eff}}(1,6) = 5$). At $n=32$ the float check closes to $6.5\times10^{-13}$. The gauge identity of §6 is the reason: $G_k$ and $L^{+}$ differ only by rank-one pieces built from $\mathbf 1$, and $e_i - e_j$ annihilates them.

gauge invariance — drag the ground, watch the difference freeze self-test …
vi vj vi − vj = Reff(i,j) |i − j|
The voltage profile $v = G(e_i - e_j)$ for a one-ampere dipole between $i$ (blue) and $j$ (red). Drag the ground $k$: the whole curve shifts so that $v_k = 0$ (gold electrode), $v_i$ and $v_j$ jump — but $v_i - v_j$ and $R_{\mathrm{eff}}$ never move, because $v_i - v_j = (e_i-e_j)^{\top}G(e_i-e_j)$ is exactly the effective resistance. In pseudoinverse mode no node is zero; instead $\sum v = 0$ (dashed mean line) — a different convention, the same differences. In ground mode every displayed voltage is an exact integer. The self-test recomputes $R_{\mathrm{eff}} = \vert i-j\vert$ from all 32 groundings and the centered pseudoinverse over all $32\times32$ pairs at load.

This is a gauge choice in the physicist's strict sense: a redundancy of description fixed by convention, with a dictionary of observables that do not care. Potentials are gauge; potential differences, effective resistances, and currents are physics. The suite has been quietly using this freedom all along — the dual-basis page's response matrix $X = B_{\mathrm{inc}}L^{+}$ is a gauge-fixed family of dipole fields (each row a balanced $e$-difference, so the choice of gauge cancels row by row), and report 10's improper prior is the statistician's same move: don't fight the flat direction, just report functionals orthogonal to it. Where you put the zero is a fact about your notebook. $\vert i-j\vert$ is a fact about the rod.

Terminology, gently: everything on this page is a boundary condition — a constraint pinned to the space axis. Its famous cousin, the initial condition, lives on the time axis, and a full heat problem $\partial_t u = u'' + f$ needs one of each kind: boundary conditions at the ends for all time, an initial temperature across the rod at $t=0$. The steady states of this suite are what remains after the initial condition has been forgotten — the $t\to\infty$ limit where only the boundary condition still speaks; report 10 tells the time-axis half of the story, and the green tents the equilibrium half. Saying "initial conditions" for the ends of a rod is the one substitution this dictionary won't survive.
§8 · Suite crossings

Where the suite has already touched this page

The seedthe green tents §4: free vs grounded, the degree diagonal as insulation, and the 0/1/2-pin table this page stretches into a $\kappa$-continuum.

The two-pin worldthe bridge explainer and report 09: everything downstream of $\mathrm{tridiag}(-1,2,-1)$, including the rank-1 triangle read as bridge covariance.

The rank machineryreport 13 §3 (semiseparable, Gantmacher–Krein) and report 14 (separator rank; this page's corner certificate equals its off-diagonal block rank on the chain).

The gauge usersthe dual-basis page ($X = B_{\mathrm{inc}}L^{+}$, dipoles immune to grounding) and mathematica/electrical_networks.wls (effective resistance, the projection $\Pi$, Vishnoi's conventions — including the $B$ sign this page's $L=B^\top\!B$ inherits).

The statisticsreport 09 (precision = stiffness; conditioning = principal submatrix) and report 10 (proper vs improper priors: the pseudoinverse as the covariance the intrinsic field almost has).

Everything above is machine-verified twice. Python: python/experiments/grounding_checks.py24/24 PASS in 0.1 s, every quoted number exported to results/grounding.json (145,812 bytes): the free-Laplacian anatomy (diag $\{1,2,2,2,2,1\}$, rank 5, kernel = constants); the $e_1$ solvability failure at residual $1/\sqrt6$ vs the balanced $e_1-e_2$ solve ($8.9\times10^{-16}$); both user findings with exact corner certificates $[0,1,1,1,1,1]$ and $[1,2,2,2,2,1]$, the pinv rank exactly 2 (rank-2 blocks at $j=2,3,4,5$); $G_1 = \min(i,j)-1$ and the interior block splits, all $k$; the wall-append identity tridiag$(-1,2,-1)$ = laplacian_1d(6) and the bridge $\min(i,j)(7-\max(i,j))/7$; Robin exact at five $\kappa$, closed-form sweep $10^{-6}..10^{6}$ ($2.2\times10^{-10}$), both limits (error $=1/\kappa$, to $5\times10^{-6}$); Gantmacher–Krein on 20 random SPD tridiagonals; the gauge identity $L^{+}=CG_kC$ for every $k$ with all four Moore–Penrose axioms exact ($1.8\times10^{-15}$ vs NumPy); the two-pair decomposition of $L^{+}$; $R_{\mathrm{eff}}=\vert i-j\vert$ exact from every gauge; corner rank == separator rank on every mode; and the full $n=32$ replay (min-kernels all $k$, continuum $\tfrac1h\min(x,y)$ / $\tfrac1h(\min(x,y)-xy)$, 9-$\kappa$ Robin grid, gauge identity, $R_{\mathrm{eff}}$; $\le 6.5\times10^{-13}$). Wolfram: mathematica/grounding_checks.wls26 PASS, 0 FAIL, exit 0, no N[] anywhere: the user's TriangularRankOneQ / TriangularRankQ / GroundedInverse verbatim on exact rationals, both findings, the sharper rank-exactly-2 facts, closed forms for every $G_k$, the $\min(i,j)$ Robin family with affine factor pairs, the gauge identity, and $R_{\mathrm{eff}}(1,6)=5$. Both in-page widgets re-verify their baked data at load (green badges). Companion pages: the green tents · the bridge explainer · one space, two bases · 09 · stiffness as precision · 10 · fluctuation–dissipation · 13 · preconditioning as decoupling · 14 · the hierarchical inverse.