Evidence Portal
Evidence, not proof.
This portal documents how we benchmark ODE solvers, which flagship problems we've characterized, and how SolvSRK compares to 13 other solver arms - SciPy, SUNDIALS CVODE, and high-order reference integrators - across four objectives and six noise levels, including cases where we recommend a competitor or document total failure.
All of it is our own work: internally generated, self-reported, and never audited or replicated by anyone outside Resonix. We publish the method, the baselines, and the losses so you can judge it - and we would rather you reproduce the part you care about than take our word for it.
Three paths
Prove we lose
Start with dead zones and competitor wins.
Dead zones →
Find something like mine
Browse the Phase-1 flagship catalog by stiffness and domain.
Problems →
Cite this in a paper
Canonical methodology, the four objectives, verification status, and a citation block that names it a self-reported benchmark.
Methodology →
398
benchmark problems
14
solver arms compared
6
noise levels (clean + 5 σ)
7
dead-zone / honesty rows
Featured problems
All 398 →Wins and losses - not wins only. Each badge compares SolvSRK to the best competing arm we ran; what the verdicts mean.
Arenstorf orbit
DISADVANTAGESolvSRK leads clean precision, but at σ=0.1 it drops to 0% survival while SciPy arms hold 100%. This is a SolvSRK disadvantage - and we show it.
S0 · dim 4
E5 autocatalytic reaction
DEAD ZONEDEAD ZONE: SolvSRK and all 13 baselines fail this configuration. We publish the losses next to the wins - this is one of them.
S3 · dim 4
HIRES isomerisation
ADVANTAGEClean SCD is essentially a tie with SciPy DOP853 and Radau. The published gap opens under noise: SolvSRK 100%, all SciPy arms 0% at σ=0.1.
S3 · dim 8
Kepler two-body orbit
ADVANTAGEOn clean runs the high-order explicit Vern9 integrator is more precise (≈7.5 vs 6.4 SCD). SolvSRK's advantage shows only under noise, where it survives and the SciPy arms drop to 0%.
S0 · dim 4
Kuramoto oscillators (N=20)
ADVANTAGETwenty coupled phase oscillators - a medium-scale synchronisation benchmark. SolvSRK leads clean precision and is the only arm to survive at high noise.
S0 · dim 20
Linear stiff system (dim 10)
ADVANTAGEA dense ten-state linear stiff system from the DETEST benchmark suite. SolvSRK leads clean precision and holds survival under noise where SciPy collapses.
S2 · dim 10
Lotka–Volterra predator–prey
DISADVANTAGESciPy Radau is more precise clean, and at σ=0.1 SolvSRK survives only 30% vs 100% for SciPy. SolvSRK is the wrong choice here - published so you can see it.
S0 · dim 2
Robertson chemical kinetics
ADVANTAGEOn clean Robertson, SciPy BDF and Radau also finish. SolvSRK's edge is survival under injected noise - where every SciPy arm drops to 0% at σ=0.1.
S3 · dim 3
FAQ
- Has any of this been independently verified?
- No. There is no third-party audit, external replication, peer review, or certification behind these numbers - we generated all of them ourselves. That is why the method, the baselines, and the dead zones are published: so the claim can be checked rather than trusted. Verification status.
- Cherry-picked benchmarks?
- The catalog publishes all 398 problems in the current benchmark freeze. The highlighted set above is curated to include wins, parity, disadvantages, and dead zones - not a wins-only gallery.
- Is the reference tuned to favor SolvSRK?
- No. SCD uses an independent high-fidelity CVODE/Radau reference. SolvSRK is never its own reference. See methodology § accuracy.
- Why not publish all raw trial rows?
- Volume without context misleads. We publish governed per-problem medians you can read one page at a time - not a downloadable corpus dump. To fingerprint your problem against the full catalog, register for free SolvScout.
- Can I download the whole evidence table?
- No. The portal is an open book for browsing - not a bulk data product. SolvScout is the free path to the corpus when you want to check your own ODEs.
TRL 4–5 · simulation-lab validated · 398 problems · 14 solver arms · clean + 5 noise levels
Freeze: 2026-08-13 · scipy 1.14 · libsolvsrk 2.3.0 · Methodology
Self-reported by Resonix Labs · not independently verified · Verification status