Reproducible ML decisions. Verifiable creator payouts. Auditable ad attribution. Stable watermark detection. One arithmetic substrate, one SHA-256 per artifact, provable compliance across AI Act, DSA, and DMA.
Proofs of concept
Each POC opens with a scenario a platform team will recognize, walks through what SolvNum changes, and ends with a benchmark number plus a SHA-256 you can re-derive on your own hardware.
1,000,000 ML decisions. SolvNum produces 0 receipt drifts across platforms. Float64 produces ~32,000. EU AI Act Article 12 ready.
32/32 receipts identical across Windows and Linux on 6.1M real Criteo journeys. Float64 drifts on the same machine under BLAS swap.
Bit-identical payouts across Python (Win/Linux) and JavaScript BigInt. Browser-verifiable. $0.013 drift eliminated.
Platform capabilities
The same deterministic arithmetic, evidence-based tooling, and audit-grade research platform show up across AI decisions, ad measurement, creator payouts, content integrity, privacy, and regulatory compliance.
AI Decision Audit
The decision layer downstream of any ML model — calibration, policy-rule aggregation, thresholding, decision logging — made bit-identical on every platform. One SHA-256 receipt per decision that any party can independently re-derive. The audit artifact that EU AI Act Article 12 and DSA Article 40 actually ask for.
Ads Measurement
Shapley-value and Markov-chain attribution on the SolvNum substrate. Bit-identical per-campaign credit across every platform, with a three-sided handshake (advertiser, platform, regulator). Optional deterministic differential-privacy overlay that is itself reproducible. The arithmetic-drift component of the reconciliation problem stops being a reconciliation problem.
Creator Economy
A five-stage payout pipeline (engagement rollup, fraud adjustment, eligibility, revenue share, FX + tax) in integer micros. Platform issues a receipt; creator re-derives in their browser; regulator re-derives from the audit artifact. Three parties, zero trust, one SHA-256. The dispute queue becomes a hash check.
Content Integrity
LLM watermark detectors compute a z-score and threshold to decide “watermark present” or “absent.” Near the threshold, float drift flips the decision. SolvNum makes the detection statistic bit-identical across platforms so the decision is stable on every borderline case.
Privacy-Preserving Measurement
Standard DP adds random noise that is not reproducible across platforms. Deterministic DP replaces the random source with a cryptographic DRBG keyed on a committed secret. SolvNum’s proprietary quantization absorbs cross-platform math-library divergence in the noise draw. The DP-protected output is bit-identical and still ε-DP. A genuine research contribution with no clean production implementation elsewhere.
Regulatory Infrastructure
EU AI Act Article 12 (decision logs), DSA Article 40 (researcher reproducibility), DMA Article 5(9) (advertiser verification) all converge on the requirement that the same input must produce the same output with a reproducible audit trail. SolvNum is the arithmetic substrate that makes this possible across every hardware combination a regulator could audit from.
Want to see the receipts on your own pipeline?
We offer a bounded-scope reproducibility audit against a sanitized snapshot of your real decision pipeline. Two weeks, $25K, fully credited toward a Year-1 license.
The product line
Each product brief below opens with a scenario your team will recognize, then walks through what it is, who uses it, why it’s defensible, and how the claims survive review.
Same input, same output, every platform, one hash.
SolvNum is a deterministic real-number primitive that produces bit-identical arithmetic across x86, ARM, GPU, and WebAssembly. One SHA-256 receipt proves every party agrees. The substrate for reproducible ML decisions, verifiable payouts, auditable attribution, and stable watermark detection.
The only ODE integrator that survives stiff dynamics under noise.
Production RK4 at 2 kHz dies in <100 ms on stiff contact. SolvSRK survives every stress axis at 100% survival. Drop-in C library, MIT/Apache, no Python dependency. The integrator stability your perception, physics, and attitude pipelines silently depend on.
How we earn the right to make every claim on this page.
Every benchmark, every dead zone, every survival rate on this page came out of an AI-orchestrated research lab. One engineer. Six weeks to stand up. Over 1 trillion scalar observations, 1 billion validated data points, and zero retracted findings.
Engagement model
Read-only reproducibility audit on your pipeline in two weeks. Shadow run alongside your existing stack for one month. Production pilot on one workload. Every option credits in full toward a Year-1 license. Every numerical claim traces back to a test ID, a seed, and a verdict in a live evidence register.