Four capabilities. Five POCs. One arithmetic substrate.
Every platform POC on this site is built on some combination of four first-order capabilities of SolvNum and SolvSRK. This page is the technical reference: what each capability is, why it matters to platform teams and their regulators, and which POCs lead on it.
Cross-Platform Arithmetic Reproducibility
SolvNum produces bit-identical arithmetic across Windows, Linux, GPU, ARM, and WebAssembly. One SHA-256 receipt proves every party agrees.
How it works
Floating-point math disagrees across machines because different CPUs round differently, different math libraries approximate transcendentals differently, and different compilers reorder operations. Across thousands of arithmetic chains these gaps compound into visibly different outputs.
SolvNum uses a deterministic arithmetic representation that eliminates platform-dependent rounding. Every platform performs the same operations on the same data the same way. The canonical bytes of the output produce the same SHA-256 on every machine.
For platforms: the content-moderation engine, the ad-measurement pipeline, the creator-payout system, and the regulator’s audit replay all produce the same bits. A reviewer can independently re-derive the receipt on their own hardware.
Why it matters
- EU AI Act Article 12 requires reproducible audit logs for high-risk AI decision systems. SolvNum makes the decision-layer arithmetic bit-identical across platforms so the log hash is meaningful.
- DMA Article 5(9) requires advertisers to independently verify ad performance reporting. Same inputs must produce same outputs regardless of whose hardware runs the math.
- Creator-economy disputes (“your number doesn’t match my number”) become a single-hash verification instead of a customer-service escalation.
Validation status
PASS — SHA-256 receipt match verified across Windows-x64, Linux-x64, WSL2, dedicated Linux server, and Node 18 BigInt runtime on all four POC benchmarks. 32/32 receipts identical per POC across every host pair tested.
Three-Sided Verifiable Workflow
Platform issues a receipt. The counterparty re-derives the same number on their own hardware. An auditor does the same from the published artifact. Three parties, zero trust, one hash.
How it works
The platform runs a computation (attribution, payout, decision) and emits a receipt: input-bundle hash, per-stage hashes, output, and a final SHA-256 over canonical bytes. The receipt format deliberately excludes environment-dependent fields (wall time, raw libm draws) from the canonical hash.
The counterparty (advertiser, creator, or researcher) takes the same input bundle and re-runs the pipeline on their own machine. If the receipt hash matches, the math is verified. If it doesn’t, the per-stage hashes pinpoint exactly where the divergence occurred.
A third party (regulator, auditor) can do the same from the published artifact without access to either side’s code. The only dependency is the open-source SolvNum library and the documented receipt schema.
Why it matters
- Creator payout disputes become a deterministic verification instead of a human-reviewed escalation. The creator runs the verifier in their browser.
- Ad-measurement reconciliation: advertiser, platform, and regulator each independently derive the same attribution credit from the same inputs. Arithmetic-drift disagreements are eliminated by construction.
- Regulatory audit: DSA Article 40 researcher access and AI Act Article 12 logging both benefit from a receipt that any party can verify without trusting the platform’s binary.
Validation status
PASS — three-sided handshake demonstrated end-to-end: platform issues receipt on Windows, creator verifies on Linux (Python + Node BigInt), auditor re-derives from published artifact. Every stage hash and final receipt hash match across all three parties.
POCs that lead on this capability
Deterministic Differential Privacy
Deterministic DP noise generated from a committed key + data fingerprint. Reproducible, auditable, and still ε-DP against an adversary without the key.
How it works
Standard differential privacy adds random noise from a platform PRNG. The noise is not reproducible, so DP-protected reports cannot be independently re-derived — and the advertiser has to trust the platform added the right amount of noise.
Deterministic DP replaces the random source with a cryptographic DRBG keyed on a long-term secret. The same key + data fingerprint always produces the same noise, but the output satisfies computational differential privacy (cDP) against a polynomial-time adversary who does not hold the key.
SolvNum’s proprietary quantization absorbs cross-platform math-library divergence in the noise draw. The raw Gaussian samples may differ by fractions of a bit across platforms; SolvNum folds them into the same deterministic representation. The DP-protected output is bit-identical across platforms.
Why it matters
- Post-ATT / post-Privacy-Sandbox ad measurement adds noise layers to protect user identity. Deterministic DP makes the noised output reproducible so the advertiser can verify the platform used the claimed (ε, δ) parameters.
- Composition accounting becomes tractable: same key + same data = idempotent release. Key rotation policy + Renyi-DP accounting give clean multi-release guarantees.
- The construction is a genuine research contribution — no clean production implementation of deterministic DP aggregation exists today. It is the most technically novel piece in the platform POC suite.
Validation status
PARTIAL — mechanism implemented and cross-platform receipt match verified. Formal privacy review by an external DP expert is a required prerequisite gate before any (ε, δ) claim is published. The privacy review charter is documented in full.
POCs that lead on this capability
ODE Integration Under Noise and Stiffness
SolvSRK is the only ODE integrator validated for reliable integration under real-world noise — 87–94% survival across 28 systems where standard solvers achieve 0%.
How it works
SolvSRK is a general-purpose ODE solver library. Caller supplies an rhs(t, y, dydt) callback; SolvSRK owns the solver orchestration. No new filter framework is invented — it is a drop-in ODE solver.
On easy regimes (non-stiff, short-window) SolvSRK matches RK45 accuracy at ~4× the cost — use RK45 in production. On hard regimes (stiff intermittent contact, stiff ODEs under multiplicative noise) SolvSRK is the only solver that survives every cell.
Four buyer-relevant dynamics regimes benchmarked: IMU preintegration (VR/AR), avatar physics with 22-DOF articulated contact (VR/gaming), long-duration attitude dynamics, and van der Pol μ=1000 stiffness survival under noise.
Why it matters
- VR/AR physics: production RK4 at 2 kHz dies in <100 ms on stiff contact (k ≥ 1e6 N/m). SolvSRK survives every stress axis at 100% survival rate. Limbs stop bouncing through the floor.
- Embedded dynamics for drones, ADAS, and inertial-nav: when the sensor noise gets aggressive or the integration horizon stretches, standard solvers diverge. SolvSRK stays bounded.
- The win is stiffness × noise, not throughput. On easy regimes SolvSRK is an expensive reference; on hard regimes it is the only option that works. The pitch is honest about both.
Validation status
PASS — 100% survival on every stress axis across four sub-POCs (IMU preintegration, 22-DOF avatar physics, attitude dynamics, van der Pol μ=1000). RK4 achieves 0% on stiff axes; RK45 dies at σ=1.0 on VdPol. SolvSRK is 130× more efficient than RK45 in RHS evaluations at σ=0.
POCs that lead on this capability
How they reinforce each other
Pair-wise and full-stack benefits
The pair-wise compounds are where most of the value lives. The full-stack combination — reproducibility + verification + privacy + stiffness survival — is the complete trust substrate for platforms under regulatory pressure.
R + V
Reproducible arithmetic plus three-sided verification — the core trust primitive. The platform, the counterparty, and the regulator each independently derive the same answer. No trust required, one hash settles it.
R + P
Reproducible arithmetic plus deterministic DP — the ad-measurement substrate. Privacy-protected attribution reports that the advertiser can verify used the claimed (ε, δ) parameters, on their own hardware.
V + P
Verifiable workflow plus privacy overlay — the regulator artifact. A DP-protected attribution receipt that any party can re-derive, with the noise parameters in the header and the privacy-review attestation linked from the receipt.
R + S
Reproducible arithmetic plus stiff-dynamics survival — the VR/AR substrate. Avatar physics that survive stiff contact, with bit-identical receipts so the twin on the cloud and the device produce the same trajectory hash.
R + V + P + S
The full platform substrate: SolvNum makes the arithmetic reproducible, the three-sided workflow makes it verifiable, deterministic DP makes it privacy-compliant, and SolvSRK makes the dynamics survive real-world noise. End to end, one hash, defensible under AI Act, DSA, and DMA.