Imagine you built a home safe out of the best lock money can buy, then bolted it to a plywood door. That's the core tension Zirah maps: you can run a post-quantum signature algorithm on a $5 microcontroller, but the key has to live somewhere, and 'somewhere' on cheap hardware is unprotected flash memory behind a single configuration bit. The paper's honesty about this is its best feature. The committed claim is narrow but useful: you can build a complete post-quantum hardware signer — avalanche TRNG, ML-DSA-65 with side-channel masking, ECDSA co-signature, human confirmation via OLED — on an STM32F407 at 168 MHz, signing in under 0.8 seconds, and the whole thing matches the pq-crystals reference output across 1,976 test vectors. That's an engineering integration result, not a cryptographic breakthrough. The paper knows this. The randomness pipeline is where the measurement work lives. A reverse-biased 2N3904 junction in avalanche breakdown feeds raw noise through an LM358 op-amp into SP 800-90B health tests. NIST's tool measured 6.898 bits of min-entropy per byte on the non-IID track, while the firmware conservatively credits only 1 bit per byte — a 7× safety margin. But spectral lines near 116 Hz carrying about 5% of the variance remain unexplained. The paper flags this rather than hiding it, which is the right move, but it means the entropy source isn't fully characterized. The side-channel defense uses two-share arithmetic masking on the secret-key polynomial multiplications inside ML-DSA-65. The validation here is where you should squint: leakage assessment comes from an instruction-level emulator (Unicorn on Cortex-M4), not from actual electromagnetic or power traces on the physical board. The paper states this plainly — 'nothing here is certified' — but emulated leakage and real leakage are different animals. An EM probe on the running board is the obvious experiment that didn't happen. The ATECC608 secure element handles ECDSA P-256 as a co-signature but cannot store the 4,032-byte ML-DSA-65 key, which is the architectural bottleneck the paper can't solve. The post-quantum key sits in ordinary flash behind RDP Level 1 readout protection. The paper demonstrates that before RDP was enabled, a debugger extracted the key and produced a valid signature without the device. This is the plywood door. RDP is a fuse, not a vault — known fault-injection attacks can bypass it on STM32F4 chips. What makes this paper worth reading is the threat-model honesty, not the performance numbers. The 157 ms median signing time and the 1,976 passing test vectors are table stakes for a correct ML-DSA implementation. The real contribution is a worked example of what post-quantum migration looks like on constrained hardware: where the TRNG entropy budget sits, where side-channel masking is and isn't validated, where the key storage is genuinely vulnerable, and where a secure element's limitations force compromises. For anyone planning PQ migration on embedded devices, this is a useful map of the minefield. The paper is a single-author effort with full open-source release of hardware designs, firmware, host software, and measurement data on Zenodo. That's commendable. But the absence of physical side-channel measurements and independent replication means the security claims rest on emulation and self-assessment. The next step — and the one that would turn this from an engineering report into a security result — is physical leakage testing on the actual board.
