Applied · the clinical companion lives in the Spine repo
A chain is held to a fixed total rotation. Some interior elements are made rigid. The motion they held is conserved and must reappear in the survivors. Every engineer's first guess, and every textbook's, is that it reappears next door. In the single-element case the evidence for that is real but contested. In multi-element constructs it fails outright: the largest increase sits three elements away from the rigid block, in every such construct measured. Two candidate redistribution laws are ruled out by arithmetic on the published point estimates. What replaces them is not a local law at all.
1 · The setting, stripped of its anatomy
A serial chain of $n$ compliant elements is driven end to end to a fixed total displacement $\Theta$. A subset $F$ of the elements is made rigid. The constraint is not a perturbation and not a load: it is a hard zero imposed on $|F|$ coordinates while the sum is held.
The identity VERIFIED
$$\sum_i \theta_i = \Theta \quad\text{(held)},\qquad \theta_f \to 0 \ \ \forall f \in F$$
$$\Longrightarrow\qquad \sum_{i \notin F} \Delta\theta_i \;=\; \sum_{f \in F} \theta_f^{\text{intact}}$$
The displaced motion is conserved. This is arithmetic on the protocol, not a claim about the medium. It holds for springs, for discs, for anything that adds.
The identity says how much. It says nothing about where, and the whole content of the problem is in the where. This is the same structure the corpus meets in WP-39: a conserved column quantity, a constraint that redistributes it, and an order-of-operations question about what commutes with what. Here the constraint is a rigidity rather than a lid, and the conserved quantity is a rotation rather than a burden, but the shape is identical — and so is the trap. A constraint is not a local operator. The instinct to ask which element is next to the rigid block is an instinct imported from local physics into a globally constrained problem, and it has no standing there.
2 · The measurement
The realisation used here is the human subaxial cervical spine: six motion segments, C2–C3 through C7–T1, cranial to caudal. [C] tested nine fresh-frozen cadaveric spines C2–T2 on a six-degree-of-freedom robotic manipulator under a displacement-control protocol — total C2–T1 rotation driven to the intact target in every condition. That protocol is what makes §1's identity apply exactly, and it is why this particular study, out of a large literature, is the one that can answer the question at all.
| Rigid set $F$ | C2–C3 | C3–C4 | C4–C5 | C5–C6 | C6–C7 | C7–T1 |
|---|
| C5–C6 | — | — | — | rigid | — | — |
| C4–C6 | +2.50°
$d=2$ | n.s.
$d=1$ | rigid | rigid | n.s.
$d=1$ | n.s.
$d=2$ |
| C5–C7 | +3.90°
$d=3$ | n.s.
$d=2$ | +2.70°
$d=1$ | rigid | rigid | n.s.
$d=1$ |
| C4–C7 | sig.
$d=2$ | sig. (rot)
$d=1$ | rigid | rigid | rigid | n.s.
$d=1$ |
Flexion unless marked; $d$ is distance in elements to the nearest rigid one, contiguous $= 1$. Intact flexion where reported: C3–C4 $6.0 \pm 2.0$, C4–C5 $6.5 \pm 3.6$, C5–C6 $7.1 \pm 1.3$ degrees.
The single-element case is contested, and is not used below OPEN
[C] reports no significant adjacent-element motion increase in any direction after single-element C5–C6 rigidity ($p > 0.05$). [E], on the same construct under the same protocol class, reports adjacent intradiscal pressure up $73.2\%$ at C4–C5 ($p = 0.002$) and $45.3\%$ at C6–C7 ($p = 0.006$) in flexion, with motion increased at both.
Different observables — pressure against motion — and different drive points: [E] fixed T1 and drove C3 to $20^\circ/15^\circ$; [C] drove total C2–T1. Neither reading makes the single-element case settled, so it is set aside. Everything that follows uses only [C]'s two- and three-element conditions, which are internal comparisons on one rig and do not touch this conflict.
3 · First law refused: decay with distance
The stress-riser account — the neighbour takes the load, the next one less — is a family of models, not one model. What the whole family shares is monotonicity: $\Delta\theta$ non-increasing in $d$. That is an ordinal prediction. It survives any rescaling, needs no units, and does not care which cells [C] left blank.
Refusal, on reported point estimates alone VERIFIED
Under $F = \{$C5–C6, C6–C7$\}$: C2–C3 at $d=3$ rises $+3.90^\circ$ > C4–C5 at $d=1$ rises $+2.70^\circ$.
No null is read as zero, no missing cell is filled, no normalisation is chosen. A more distant element takes more of the conserved motion than a contiguous one. Every monotone-decay model is out.
Under $F = \{$C4–C5, C5–C6$\}$ the refusal needs the weaker reading — the only significant flexion increase sits at $d=2$ while both $d=1$ elements are null — and [C]'s own conclusion states it in those words: non-contiguous motion was increased. The three-element construct puts it at C2–C3 again. Three of four conditions, same element, and it is the one farthest from the hardware.
4 · Second law refused: proportional to compliance
The careful answer, and the one a physicist reaches for second: springs in series. With compliances $c_i$ and fixed total $\Theta$, $\theta_i = (c_i/\sum c)\,\Theta$; sending $c_f \to 0$ for $f \in F$ gives, for every survivor,
θipost / θiintact = (Σall c) / (Σremaining c) — independent of i
Every survivor rises by the same fraction. No favoured neighbour, no favoured end. The script checks this algebra numerically on a synthetic chain before turning it on data, because a model that is wrong in its own arithmetic cannot be refuted by anyone else's.
[C] reports both an intact value and a change in exactly one cell: C4–C5 under $F=\{$C5–C6, C6–C7$\}$, intact $6.50^\circ$, change $+2.70^\circ$, a $+41.5\%$ rise. Two consequences follow with no further assumptions.
Consequence one VERIFIED
C3–C4 (intact $6.00^\circ$) must then rise $+2.49^\circ$ — within $0.3\%$ of $+2.50^\circ$, the smallest change this study did resolve and report as significant. The experiment had the sensitivity to see it. It reports nothing there, in either two-element condition.
Consequence two VERIFIED
For C2–C3 ($+3.90^\circ$) to outrun C4–C5 ($+2.70^\circ$) under a common fraction, C2–C3's intact flexion must exceed C4–C5's in the same ratio: $\geq 9.39^\circ$, against $6.50^\circ$ at C4–C5 and $6.00^\circ$ at C3–C4. [C] does not report C2–C3 flexion; its lateral-bend row puts C2–C3 within $10\%$ of C3–C4, not $44\%$ above its neighbours.
The second consequence is the stronger one, because it does not lean on a null. It is an ordering, and the series law forbids that ordering outright unless the intact profile is shaped in a way the intact profile is not.
What survives MODEL
Both local pictures are out: decay with distance, and redistribution in proportion to compliance. What is left is a mode statement. The chain does not hand its motion to a neighbour; it re-shapes, and the amplitude reappears where the chain was already freest — at the cranial end, nearest the mass it carries. Which element is adjacent to the constraint turns out not to be a property that matters.
5 · The denominator, or: the same error the corpus keeps meeting
So much for the mechanics. The clinical literature ought to adjudicate, and it does not, because it disagrees with itself about the sign.
| Source | Endpoint | Size | Effect of more rigid elements |
|---|
| [H] 1999 | new symptomatic disease at an adjacent level | 374 patients
409 constructs | significantly lower after multilevel than single-level, $p < 0.001$ — the authors expected the opposite and said so |
| [N] 2024 | reoperation for symptomatic disease | 60,292 records | multilevel an independent predictor, aOR $1.61$ $[1.47, 1.75]$, $p < 2\times10^{-16}$ — higher |
| [C] | adjacent-element motion | 9 specimens | three-element constructs give the greatest change of any tested — worse |
The resolution is not subtle, and it is the error WP-100 and WP-98 are both about: the reported quantity is not the invariant. A patient does not carry a risk; a patient carries a set of interfaces, each able to fail. Making $k$ of them rigid removes $k$ from the count. Under a constant per-interface hazard $h$ and no protection of any kind,
R(k) = (n − k) · h — which falls as k rises
With $n = 6$ the ratio $R(3)/R(1)$ is exactly $3/5 = 0.600$: a forty per cent drop in the per-patient rate, produced by bookkeeping, with nothing happening in the tissue. Sensitivity: $1/2$ at $n=5$, $2/3$ at $n=7$; the sign does not move. [H]'s celebrated counterintuitive finding is the direction arithmetic predicts for an inert spine. VERIFIED
That [H] fitted a constant hazard is checkable from his own two headline numbers: $2.9\%$ per year over ten years gives $1 - e^{-0.29} = 25.2\%$ or $1-(1-0.029)^{10} = 25.5\%$, against the $25.6\%$ his survivorship analysis reports. Both land inside half a percentage point, which is what licenses the decomposition. Inverting it puts the per-interface hazard near $0.64\%$ per interface-year for a mean construct length of $1.5$. VERIFIED
Open — and this is the whole practical gap
Neither [H] nor [N] reports $h$. Neither reports rates stratified by construct length in a form that lets $h$ be recovered. The one quantity that is comparable across constructs is absent from both, so the correction above can be checked in direction only, never in magnitude. OPEN
6 · The discriminator is a disjointness
The two accounts — transferred load, or the natural history of a degenerating disc continuing on its own — make different predictions about which element fails, and the published record answers.
Disjoint VERIFIED
Mechanically loudest element, across three of four constructs: {C2–C3}.
Clinically riskiest elements, [H]: {C5–C6, C6–C7}.
Intersection: ∅.
If the disease followed the transferred load it would appear where the load goes. It appears instead at the two elements that degenerate in an untouched spine — the same two that sent the patient to surgery in the first place. That is [H]'s own reading of his data, and the arithmetic here does not prove it. An earlier draft of this paper added that it “removes the strongest reason to believe the alternative.” It does not, and §7 says why. MODEL
There is a second, cheaper test in [N]'s age term. Risk falls with age (aOR $0.96$/yr, peak $8.12\%$ in the 30–39 band). The interface-count model predicts that without biology: an older chain arrives with fewer mobile elements, because some have already stiffened. Count mobile interfaces on imaging that already exists, put the count in the model, and see whether the age term survives. If it collapses it was a denominator artefact; if it survives there is real age biology to find. CONJECTURE
7 · The randomised evidence, and a correction
What this paper got wrong
Arthroplasty preserves motion at the index element. It is therefore a randomised test of the load-transfer hypothesis: remove the transfer, see whether the disease follows. Eleven RCTs have run it. They were not consulted before §6 was drafted — an error of search, not of arithmetic — and they cut against the reading §6 offered.
| [B] — adjacent-element operation | Motion-preserving | Rigid | Odds ratio | $p$ |
|---|
| one element, 2 yr | 2.3% | 3.6% | 0.67 [0.41, 1.09] | 0.106 |
| one element, 7 yr | 4.3% | 10.8% | 0.37 [0.22, 0.62] | < 0.001 |
| two elements, 7 yr | 5.1% | 10.0% | 0.49 [0.27, 0.87] | 0.014 |
A $6.5$ percentage-point absolute reduction at seven years, one element; NNT $\approx 15$. The pooled odds ratios are trial-weighted averages rather than the odds of the pooled rates and need not coincide with a direct recomputation — they do anyway, to $0.001$ and $0.006$ at seven years, which excludes a transcription error. VERIFIED
This is the strongest evidence class in this paper and it favours load transfer. Nothing in §§3–6 outranks a randomised trial.
7.1 What it does not settle
The divergence is late: not significant at two years, separating near five. And on the endpoint nearer the tissue, [P] (5 RCTs, 48–60 months) finds symptomatic adjacent disease $8.8\%$ against $13.0\%$, $RR = 0.57$ $[0.19, 1.72]$, $p = 0.32$, and concludes that no such reduction can be concluded. A null on 140 patients is absence of evidence. OPEN
The measurement nobody has made
[P] records that no included trial compared radiographic adjacent-element degeneration between the two arms. The endpoint nearest the tissue, and furthest from anyone's decision, has never been measured head to head. The endpoint that has been measured — operation — is a decision. The trials are unblinded and industry-sponsored, and the conventional next step after a rigid construct fails at its neighbour is to extend the construct. Randomisation protects the contrast against confounding; it does not change what the endpoint means. OPEN
The size of that concern is computable. The rigid arms of these trials reach $10.8\%$ at seven years, $1.54\%$/yr; the $60{,}292$-record registry [N] reports $6.57\%$ at ten-plus years, $0.66\%$/yr. A patient inside a device trial is reoperated about $2.3\times$ faster per year than one outside it. That is surveillance and threshold, not disease — it leaves the within-trial contrast intact and makes the absolute rates unrepresentative. VERIFIED
The honest position
FOR load transfer: eleven RCTs on the surgical endpoint; adjacent pressure up after single-element rigidity [E]; longer constructs move survivors more [C].
AGAINST or unexplained: the motion goes to the far end and the disease does not; both local redistribution laws refused; no difference in symptomatic disease at 4–5 years; none at two; the per-patient rate is not the invariant.
No single account yet explains a randomised surgical benefit and the element disjointness. Asserting that either half settles it is the error this paper set out to prevent and then made.
8 · What the account does outside the literature
The corpus has a standing interest in the distance between a result and the statement that travels under its name — WP-107, WP-95. This instance is unusually direct, because the statement is what patients are handed when they choose between operations.
[D] is a patient-facing commercial page from a clinic selling a proprietary alternative to fusion. The conflict of interest is named and then set aside; claims are judged against sources. Three of its claims check out, one of which corrects an error this paper would otherwise have made — the “$30$–$70\%$” adjacent-level rise it quotes is real and is [E]. Two do not.
Contradicted by its own citation VERIFIED
[D]: rates rise sharply with each additional fused level.
[H] — the page's own source for its incidence figures — found multilevel risk significantly lower, $p < 0.001$, and recorded that the authors expected the opposite. The page takes the rate from [H] and the levels claim from elsewhere, without noting the collision.
The causal clause does not follow VERIFIED
[D]: disease occurs at C5–C6 and C6–C7 because they inherit the largest share of transferred load.
Where the disease is: {C5–C6, C6–C7} [H]. Where the transfer is largest: C4–C5 by pressure [E], C2–C3 by motion in every multielement construct [C]. The location is right; the because follows from neither measurement.
And one arithmetic note, on a figure advertised rather than cited. The page reports a $0.01\%$ complication rate across $2{,}700+$ procedures. On $2{,}700$ cases the observable rates are $0\%$, $1/2700 = 0.037\%$, $0.074\%$, and so on; $0.01\%$ is not among them. The number cannot have been measured on the denominator it is quoted with. That is a statement about the number and nothing else. VERIFIED
The shape of it
“The lost motion has to go somewhere” is a conservation law and is exactly right. “It transfers to the disc and facet joints immediately above and below” names a destination the measurements do not agree on. A true identity and an unearned locality, welded into one sentence, carrying a recommendation. The identity is doing the persuading and the locality is doing the work.
9 · The falsifier
One number settles it
If adjacent-element disease is natural history, the per-interface hazard $h$ is roughly invariant to construct length $k$. If it is transferred load, $h$ rises with $k$, because [C] says longer constructs move the survivors more. Opposite predictions, one quantity, and the cohorts to compute it on are already collected. If $h$ climbs with $k$, this paper is wrong.
That is the shape the corpus asks for and it is the reason this belongs in Vol VI rather than in a clinical note alone: the general lesson is that a globally constrained redistribution has no local law, and that the first step in any applied instance is to find out what the conserved quantity is being divided by. Both mistakes here — asking which element is adjacent, and reporting a rate per patient — are the same mistake, made twice, in two different literatures that never compared notes.
10 · What is not known
Fifteen gaps are printed by the producing script on every run and are reproduced in full in the Spine companion. The four that bear on the argument above:
- [C] does not report intact flexion at C2–C3, C6–C7 or C7–T1. The absolute conservation budget is therefore unavailable; only ordinal tests survive, which is why §3 and §4 are built ordinally. OPEN
- [C]'s C4–C5 change under C5–C7 has a point estimate of $2.70^\circ$ sitting outside its own confidence interval as extracted ($2.26$–$2.66$). One of the two was mis-transcribed and the printed table is needed. The $41.5\%$ figure in §4 depends on the point estimate. OPEN
- The three-element construct — the largest effect in the study — is reported by significance without magnitudes. The most informative condition is the least quantified. OPEN
- $n = 6$ is an assumption, shown insensitive for $n \in \{5,6,7\}$; and [C] has nine specimens, so “not significant” is not “zero.” The strict refusal in §3 does not lean on that reading. The permissive one does, and is marked where it is used.
Scope
This is a paper about how three datasets fit together. It is not clinical advice, it does not say fusion is harmless or that adjacent elements are unaffected, and nothing in it should bear on any treatment decision. The clinical companion in the Spine repo states the same limit at greater length.
11 · Sources
- [C] Cadaveric cervical kinematics after one-, two- and three-level fusion. Nine fresh-frozen human spines C2–T2, mean age $48.67 \pm 10.16$ yr; six-degree-of-freedom robotic manipulator with 3D motion capture; displacement-control protocol. PMC13463742.
- [H] Hilibrand AS, Carlson GD, Palumbo MA, Jones PK, Bohlman HH. Radiculopathy and myelopathy at segments adjacent to the site of a previous anterior cervical arthrodesis. J Bone Joint Surg Am 1999;81(4):519–528. PMID 10225797.
- [N] Exploring the incidence and risk factors of reoperation for symptomatic adjacent segment disease following cervical decompression and fusion. N Am Spine Soc J 2024. PMC10803933.
- [B] Badhiwala JH et al. Cervical disc arthroplasty versus anterior cervical discectomy and fusion: a meta-analysis of rates of adjacent-level surgery to 7-year follow-up. J Spine Surg. Eleven RCTs. jss.amegroups.org.
- [P] The incidence of adjacent segment degeneration after cervical disc arthroplasty: a meta-analysis of randomized controlled trials. PLoS One 2012;7(4):e35032.
- [E] Eck JC et al. Biomechanical study on the effect of cervical spine fusion on adjacent-level intradiscal pressure and segmental motion. Spine 2002. PMID 12435970.
- [D] Deuk Spine Institute, Adjacent Segment Disease: The Fusion Risk Surgeons Rarely Mention, deukspine.com, updated 16 September 2026. Patient-facing commercial page; audited in §8, not used as a source of fact.
- Producing script: Spine/asd-load-transfer-verify.py, github.com/sluing/neuro. Every figure above is recomputed there; the script exits non-zero rather than rounding to agreement.