Skip to content

The paper · methods, results, references

A result you can check, not just a number you receive

Finite-element analysis · sequence analysis · verification against published references

Abstract

A simulation can be numerically clean and physically plausible while the model behind it is wrong. Sojist runs the computation on managed cloud compute and returns, with every result, the evidence needed to judge it: a comparison against closed-form theory where one exists, against published reference data where it does not, and a record that lets the run be repeated exactly.

  • verification
  • finite elements
  • sequence alignment
  • reproducibility
  • ASME V&V 40

1Introduction

Existing tools solve the physics. None of them state how much the answer should be believed. On a validated cantilever, the computed tip deflection agreed with the Euler–Bernoulli solution [1] to 0.086 %, while the peak stress exceeded beam theory by 29.5 % — correctly, because the formula has no term for the concentration at the clamp.

A naive equality check would have rejected that correct solution. The inverse is worse: a wrong boundary condition, material or load surface can produce a smooth, converged result that passes every numerical test. The engineer who built the model is usually the one who signs it off.

Regulators now accept simulation as evidence. The FDA recognises ASME V&V 40 [2] for establishing the credibility of a computational model — and its own review literature notes that the standard supplies no mechanism for deciding when a result is acceptable.

Request access →
Cantilever benchmarkA beam clamped at the left end and loaded downward at the free right end by 1000 newtons. The shaded shape is the deformed beam, exaggerated eight times. Measured tip deflection 9.1350 millimetres.P = 1000 NL = 1.0 mδ = 9.1350 mm
Fig. 1Cantilever benchmark. 1.0 × 0.05 × 0.05 m, carbon fibre / epoxy (E = 70 GPa), 1000 N tip load, 146,375 nodes. Shaded: the deformed beam, displacements drawn ×8 — the beam-theory shape scaled to the measured tip deflection. Measured 9.1350 mm against 9.1429 mm closed-form.
Table 1 Measured against reference.
QuantityReferenceMeasuredΔModeCopy row
Tip deflection9.1429 mm9.1350 mm−0.086 %direct
Peak stress48.00 MPa62.14 MPa+29.5 %lower bound
Longest ORF1132 aa1132 aa0direct
Reads mapped—99.76 %—reported

— no reference value exists; the quantity is reported, not judged.

2Method

Every run follows the same five steps, and each step reports its own telemetry event.

One run, end to endFive stages in sequence: specification, provisioning, solution, verification and record. A dedicated compute instance exists only during provisioning and solution, and terminates itself when the run ends.2.1Specificationdeclared, not inferred2.2Provisioninginstance launched2.3SolutionCalculiX · minimap22.4Verificationagainst a reference2.5Recordkept with the resultone dedicated instance · terminates itself
Fig. 2One run, end to end. Each stage reports its own telemetry event, and the five together are the run’s record. The compute instance exists only while the solver needs it.
  1. 2.1

    Specification

    Geometry, material, constraints and loads are declared explicitly. The system does not infer what was modelled; the declaration is the claim being tested.

  2. 2.2

    Provisioning

    A dedicated compute instance is launched for each run, installs its solver toolchain, and terminates itself when the run ends. Nothing is left running between runs.

  3. 2.3

    Solution

    Established solvers do the numerical work: CalculiX [3] with gmsh [4] for structures, minimap2 [5] for sequence alignment. Sojist does not reimplement them.

  4. 2.4

    Verification

    Each quantity is compared in the mode it permits. Direct: equal to theory within tolerance. Lower bound: at least the theoretical value, used for peak stress at a constraint. An unmeasured quantity joins no verdict, and a run with nothing measured is inconclusive — never passed.

  5. 2.5

    Record

    Specification, per-stage telemetry, solver output and verdict are kept together, so a run can be inspected afterwards and repeated. A retried run reproduced its original metrics exactly.

Step 2.4 is where tools usually stop short. Fig. 3 lists every closed-form case the platform checks against; Fig. 4 shows how a single quantity is judged; Box 1 lets you run the check yourself.

aCantilever, end load

δ = PL³ / 3EI

Roark, Table 8.1, case 1a
δ direct · σ lower bound

bCantilever, uniform load

δ = qL⁴ / 8EI

Roark, Table 8.1, case 1c
δ direct · σ lower bound

cSimply supported, central load

δ = PL³ / 48EI

Roark, Table 8.1, case 2a
δ direct · σ lower bound

dSimply supported, uniform load

δ = 5qL⁴ / 384EI

Roark, Table 8.1, case 2c
δ direct · σ lower bound

eBar in tension

δ = PL / AE, σ = P / A

Hooke’s law, uniform bar
δ direct · σ direct

Fig. 3The five reference cases. Every closed-form case the platform verifies against, its source [1] and how each quantity is compared. Dashed: the case’s own deflected shape, not to scale. Peak stress is a lower bound wherever a support can concentrate it; a bar in uniform tension has nowhere to concentrate it, so there it is compared directly.
accepted−t0+ttip deflection−0.086 %

aDirect

Accepted when |Δ| ≤ t. Used for deflection.

direct bandaccepted−t0+tpeak stress+29.5 %

bLower bound

Accepted when measured ≥ reference × (1 − t). Used for peak stress at a support, which a 3-D model can only raise above beam theory. A direct check would have accepted +29.5 % only with t above 29.5 %.

−t0+t—not reported by the run

cNothing measured

No value, so no comparison and no verdict: the run is inconclusive — never passed.

Fig. 4Three outcomes of one comparison. Δ is measured minus reference, as a percentage of the reference; t is the tolerance declared with the run. A run passes only if every quantity it measured is accepted. Axes are symbolic, not to scale.

3Results

Each capability below was run and checked against an external reference.

3.1Structural benchmark

For a cantilever with an end load, beam theory gives the tip deflection and the nominal stress at the root as

δ = PL³3EI = 9.1429 mm
(1)
σ = PLcI = 48.00 MPa
(2)

with P = 1000 N, L = 1.0 m, E = 70 GPa, I = bh³/12 = 5.208 × 10⁻⁷ m⁴ and c = h/2 = 0.025 m. The solver returned 9.1350 mm, 0.086 % below (1) (Fig. 5).

Tip deflection: one solver, two theoriesFinite elements 9.1350 millimetres, Euler–Bernoulli 9.1429, Timoshenko 9.1607, on an axis from 9.12 to 9.17 millimetres.9.129.139.149.159.169.17tip deflection, mmFinite elementsC3D10 · 146,375 nodes9.1350 mmEuler–Bernoullieq. (1)9.1429 mmTimoshenkowith shear · κ = 5/6, ν = 0.309.1607 mm−0.086 %
Fig. 5Tip deflection: one solver, two theories. The solver’s value sits 0.086 % below Euler–Bernoulli; Timoshenko’s correction for shear [6] moves the reference the other way. All three agree within 0.3 %. To scale.
Bending stress along the spanBeam theory falls linearly from 48.00 megapascals at the clamp to zero at the tip. The measured peak, 62.14 megapascals, sits at the clamp, 29.5 percent above beam theory.0204060MPa00.250.50.751.0distance from the clamp, m+29.5 %beam theory, eq. (2): 48.00 MPa at rootsolver’s peak, von Mises: 62.14 MPayield 600 MPa, off scale · safety factor 9.66
Fig. 6Bending stress along the span. Beam theory falls linearly from the root to zero at the tip. The solver’s peak sits at the clamp, where the constraint concentrates stress and beam theory has no term for it — so (2) is a floor, and the check is a lower bound. Only the peak was recorded; the full field is not reproduced here.

Its peak stress, 62.14 MPa, reads 29.5 % above (2) (Fig. 6). That is the right answer, not an error: the clamp concentrates stress and beam theory has no term for it, so (2) is a floor and the comparison a lower bound.

Judged that way, the run passed. Retried, it reproduced its metrics exactly.

3.2Sequence analysis

Run on the reference genome of phage λ [7], the platform’s sequence analysis returns 48,502 bp, GC 49.86 % and 69 open reading frames of at least 100 aa (Fig. 7). The longest, 1132 aa, matches the published length of the tail-fibre protein J residue for residue [8] and sits exactly on the annotated gene — an off-by-one in frame or stop-codon handling would break both. Aligning 3,370 nanopore reads with minimap2 [5] mapped 99.76 % of them and covered 99.95 % of the genome at a mean depth of 151.7×; an independent recount of the output agreed exactly.

The lambda genome as the platform analyses itGC content along the 48,502 base pair genome in 200 windows, ranging from about 30 to 65 percent around a mean of 49.86 percent, higher in the left half and lower in the right. Below, 69 open reading frames on the two strands; the longest, 1132 amino acids, lies at 15,505 to 18,903 on the forward strand, exactly where gene J is annotated.3040506070GC, %genome mean, 49.86 %+−71–631 (−), 186 aa191–736 (+), 181 aa310–666 (−), 118 aa771–1,538 (−), 255 aa891–2,636 (+), 581 aa2,836–4,437 (+), 533 aa2,897–3,244 (+), 115 aa4,457–5,737 (+), 426 aa4,604–5,095 (−), 163 aa5,372–5,806 (−), 144 aa5,741–6,079 (+), 112 aa6,135–7,160 (+), 341 aa7,163–7,600 (+), 145 aa7,977–8,555 (+), 192 aa8,552–8,947 (+), 131 aa8,925–9,695 (+), 256 aa9,275–9,631 (−), 118 aa9,711–10,133 (+), 140 aa9,831–10,202 (−), 123 aa10,160–10,549 (+), 129 aa10,542–13,103 (+), 853 aa11,155–11,454 (−), 99 aa11,718–12,134 (−), 138 aa12,194–12,712 (−), 172 aa13,100–13,429 (+), 109 aa13,429–14,127 (+), 232 aa14,108–14,551 (−), 147 aa14,276–14,875 (+), 199 aa14,457–14,822 (−), 121 aa14,773–15,444 (+), 223 aa15,489–15,830 (−), 113 aa15,505–18,903 (+), 1132 aa16,435–18,129 (−), 564 aa18,129–18,494 (−), 121 aa18,965–19,585 (+), 206 aa19,650–20,855 (+), 401 aa19,873–20,619 (−), 248 aa20,089–20,691 (+), 200 aa21,029–21,973 (+), 314 aa21,033–21,353 (−), 106 aa21,973–22,557 (+), 194 aa22,686–23,945 (−), 419 aa24,509–25,399 (−), 296 aa25,396–26,973 (−), 525 aa27,812–28,882 (−), 356 aa29,218–29,610 (+), 130 aa29,847–30,347 (−), 166 aa31,348–32,028 (−), 226 aa32,025–32,810 (−), 261 aa32,373–32,672 (+), 99 aa32,816–33,232 (−), 138 aa33,536–33,904 (−), 122 aa34,485–35,036 (+), 183 aa35,037–35,438 (−), 133 aa35,825–36,259 (−), 144 aa36,275–37,114 (−), 279 aa37,227–37,940 (−), 237 aa38,300–38,653 (+), 117 aa38,686–39,585 (+), 299 aa39,582–40,283 (+), 233 aa40,644–41,084 (+), 146 aa41,063–41,953 (+), 296 aa42,429–43,043 (+), 204 aa43,152–43,889 (+), 245 aa43,886–44,509 (+), 207 aa45,186–45,509 (+), 107 aa45,493–45,969 (+), 158 aa45,966–46,427 (+), 153 aa47,042–47,494 (−), 150 aaJ · 1132 aa01020304048.5position, kb
Fig. 7The λ genome as the platform analyses it. Top: GC content of 200 non-overlapping windows of 243 bp (the last, 145 bp). Bottom: the 69 open reading frames of at least 100 aa on each strand. The longest, 1132 aa at 15,505–18,903, lies exactly on the annotated gene J, the central tail fibre [7], and matches its published length [8]. Computed by the platform’s own sequence analysis from NC_001416.1. An open reading frame is a stretch between a start and a stop codon, not an annotated gene.

3.3Inspecting the result

A solver writes its whole field — 54 MB in run cd6f54b4 — far more than a browser should load to show a surface. The platform extracts the outer surface, 19,716 triangles on 9,860 nodes, and sends 0.34 MB (Fig. 8). The extrema the viewer reports match the solver’s field exactly, so the picture is the result rather than an approximation of it.

What the browser receivesThe solver output is 54 megabytes; the extracted surface sent to the browser is 0.34 megabytes, 160 times smaller, drawn to the same scale.Solver outputfull field, every node54 MBSent to the browserouter surface only0.34 MB19,716 triangles on 9,860 nodes160× smaller
Fig. 8What the browser receives. The solver writes its full result; a viewer needs only the outer surface. Both bars share one scale, so the second is a hairline. The extrema the viewer shows match the solver’s field exactly. Run cd6f54b4.
Table 2 Capabilities, as measured in August 2026.
CapabilityMeasuredChecked againstCopy row
Finite-element analysis9.1350 mm tip deflection · 146,375 nodes · 103 sEuler–Bernoulli closed form [1]
Verification against theoryper-quantity verdict across five analytical casesRoark’s formulae [1]
Sequence analysis48,502 bp · GC 49.86 % · longest ORF 1132 aaNC_001416.1 [7] · published length of gene J [8]
Read alignment99.76 % mapped · 99.95 % covered · 151.7× depthindependent recount of the same output
Result visualisation54 MB solver field reduced to a 0.34 MB surfacesolver field extrema, matched exactly

4Scope

Five workspaces share one execution path. Their depth differs, and the product says which: a workspace whose checks are not built is labelled as such rather than shown with a working-looking control.

Sojist is an analysis and decision-support layer. It never commands flight, actuation or any other safety-critical system; a qualified person makes the final call.

Each workspace in detail →
Table 3 Workspaces.
WorkspaceWhat it doesStatus
EngineeringCloud FEA, theory checks, result checks, 3-D field inspectionverified
PhysicsFive closed-form cases with referencesverified
BiologyGenome statistics, ORFs, read alignmentverified
SciencePublication metadata resolved by DOIverified
AerospaceRocket flights and campaigns on Sojist's own solver; review decisionsflights: closed-form checks; reviews: review only

5Availability

Access is currently closed while Sojist works directly with early partners. Plans, quotas and prices are listed on the pricing page. Correspondence to Emirhan Cabuk, founder and technical lead — contact@sojist.com.

References

  1. Young, W. C. & Budynas, R. G. Roark’s Formulas for Stress and Strain, 7th edn. McGraw-Hill (2002).
  2. ASME V&V 40-2018. Assessing Credibility of Computational Modeling through Verification and Validation: Application to Medical Devices.
  3. Dhondt, G. The Finite Element Method for Three-Dimensional Thermomechanical Applications. Wiley (2004).
  4. Geuzaine, C. & Remacle, J.-F. Gmsh: a 3-D finite element mesh generator with built-in pre- and post-processing facilities. Int. J. Numer. Methods Eng. 79, 1309–1331 (2009).
  5. Li, H. Minimap2: pairwise alignment for nucleotide sequences. Bioinformatics 34, 3094–3100 (2018).
  6. Timoshenko, S. P. On the correction for shear of the differential equation for transverse vibrations of prismatic bars. Philos. Mag. 41, 744–746 (1921).
  7. NCBI Reference Sequence NC_001416.1, Enterobacteria phage lambda, complete genome.
  8. Wang, J., Hofnung, M. & Charbit, A. The C-terminal portion of the tail fiber protein of bacteriophage lambda is responsible for binding to LamB, its receptor at the surface of Escherichia coli K-12. J. Bacteriol. 182, 508–512 (2000).