The Two Disciplines
Integration is connecting existing components. Install pqcrypto. Import dilithium3. Call sign(). Call verify(). Wire it into your API handler. Deploy. The components are someone else's. You're assembling them.
Engineering is building the components. Implement the NTT butterfly. Write Montgomery reduction. Design the STARK AIR constraints. Build the FHE noise budget manager. Optimize for specific hardware cache lines. Prove constant-time execution. The components are yours. You're creating them.
Both are valuable. Both require skill. But they produce fundamentally different outputs.
What Integration Produces
A well-integrated post-quantum application has Dilithium signatures on its API responses, Kyber key exchange on its TLS connections, and maybe SPHINCS+ as a conservative backup. This is real security improvement. The classical algorithms are supplemented or replaced. The harvest-now-decrypt-later attack surface shrinks.
But the integration inherits the limitations of every library it calls. If pqcrypto's Dilithium is 2x slower than necessary, your application is 2x slower than necessary. If SEAL's BFV has a 23.6x overhead compared to a native implementation, your FHE operations carry that overhead. If the STARK library you chose doesn't support recursive accumulation, your audit trail verification is O(n) instead of O(1).
Integration is bounded by the best available library. Engineering is bounded by the best available math.
What Engineering Produces
H33's stack has zero external cryptographic dependencies. Here's what that means in practice:
FHE: 23.6x faster than SEAL. Microsoft SEAL is the most widely used FHE library. H33's BFV engine (H33-128) uses Montgomery-form NTT twiddles, Harvey lazy reduction (butterfly values in [0, 2q) between stages), and pre-NTT public keys at keygen. The result: 939 microseconds for a 32-user batch versus approximately 22 milliseconds in SEAL for equivalent parameters. Same math. Different engineering.
STARK: First re-randomizable proofs. Every STARK library (winterfell, arkworks, plonky2) generates proofs that the server can fingerprint. H33 built the first commitment-based re-randomization scheme for hash-based proofs. This doesn't exist in any library because it hadn't been invented before we built it.
Dilithium: 2x native performance. PQClean's C implementation goes through FFI. H33's Rust implementation is fully inlineable by the compiler. The NTT butterfly uses NEON 4-wide vectorization on ARM. Batch CBD sampling reduces RNG calls 5x. The result: approximately 2x faster than the C wrapper on identical hardware.
Pipeline: 38.5 microseconds. An integrator wiring SEAL + winterfell + pqcrypto together gets millisecond-range latency per operation, multiplied by the number of operations in the pipeline. An engineer who built all three from scratch, optimized for the same cache hierarchy, using the same field arithmetic, with fused operations that skip intermediate serialization — gets 38.5 microseconds for the entire pipeline.
The Compounding Effect
Each engineering decision compounds. Montgomery NTT makes polynomial multiplication 3x faster. Harvey lazy reduction eliminates one reduction per butterfly stage. Pre-NTT public keys eliminate a transform per encrypt. NTT-form enrolled templates skip another transform per match. Batch attestation amortizes Dilithium across 32 users. DashMap caching eliminates STARK re-generation.
Each of these is a 10-30% improvement. Compounded across the full pipeline, they produce a system that is 875x less expensive and 8 million times faster than assembling the equivalent from off-the-shelf libraries.
Integration adds post-quantum cryptography to your application. Engineering builds an application that is post-quantum cryptography.
When to Integrate vs. When to Engineer
Integrate when post-quantum crypto is one feature of your product. You sell a CRM and need quantum-resistant TLS. Install a library. Wire it in. Ship.
Engineer when post-quantum crypto IS your product. When the security guarantee is the value proposition. When the performance envelope determines whether the product is viable. When the proof system needs to be novel because the existing ones don't solve your problem.
Use H33 when you need the engineering without doing the engineering. One API call gives you the output of three years of cryptographic engineering in 38.5 microseconds.
See the engineering: post-quantum library attestations. Score your integration: HICS. The algorithm is the authority.