The Desk

Post-quantum crypto on a radio link fails off a 0.16 dB cliff

Hold ~2 dB of link-budget margin above the error-correction waterfall so the post-quantum handshake never falls off the cliff.

An optimized post-quantum key exchange (ML-KEM) over an error-corrected radio link works flawlessly — until signal quality drops through a razor-thin threshold, where it fails…

Full framing

Hold ~2 dB of link-budget margin above the error-correction waterfall so the post-quantum handshake never falls off the cliff.

An optimized post-quantum key exchange (ML-KEM) over an error-corrected radio link works flawlessly — until signal quality drops through a razor-thin threshold, where it fails 100% of handshakes. That cliff sits right where an ordinary voice call still sounds fine, so a jammer needs only ~0.3 dB of added noise to silently kill the encryption while everything else looks normal.

SourceModel

Public-news source spine

The public record behind the subject

  1. NIST FIPS 203 (ML-KEM): decapsulation is defined over the exact ciphertext; a single corrupted bit yields a different shared secret (a decapsulation failure), which is why lattice KEMs have effectively zero tolerance for residual channel errors. csrc.nist.gov · open source →
Claim-bound findings What the evidence supports—and how strongly 3 bound claims
  1. supported

    Between Eb/N0 ≈ 1.2 dB and ≈ 1.8 dB the ML-KEM handshake goes from 100% failure to 0% — a 0.16 dB-wide cliff, not a gradual degradation.

  2. supported

    One residual bit error anywhere in the multi-kilobyte ciphertext frame collapses the lattice decapsulation, so the post-FEC frame-error rate must be driven to ~0 — far stricter…

    Full claim wording

    One residual bit error anywhere in the multi-kilobyte ciphertext frame collapses the lattice decapsulation, so the post-FEC frame-error rate must be driven to ~0 — far stricter than the 1e-3 a voice or video link tolerates.

  3. supported

    The cliff lands ~0.3 dB above where a voice-grade link is still fine, so degrading the channel just past voice quality silently breaks the crypto — the decryption-failure attack…

    Full claim wording

    The cliff lands ~0.3 dB above where a voice-grade link is still fine, so degrading the channel just past voice quality silently breaks the crypto — the decryption-failure attack surface.

0.16 dB
Failure-window width
Eb/N0 ≈ 1.2 dB
Handshake fails 100% at
Eb/N0 ≈ 1.8 dB
Fully recovers by
Eb/N0 ≈ 1.3 dB
Voice-grade link still works at
~0.3 dB
Jammer nudge to kill the crypto

Executive readout

Decision summary

Supported finding

  • Optimized post-quantum crypto over a contested radio link is a cliff, not a slope. Size the FEC and link budget to keep ~2 dB of margin above the error-correction waterfall — the…
Review 1 full note
  • Optimized post-quantum crypto over a contested radio link is a cliff, not a slope. Size the FEC and link budget to keep ~2 dB of margin above the error-correction waterfall — the handshake has no graceful-degradation region to fall back on.

Decision boundary

  • The cliff position (≈1.2–1.8 dB) comes from a calibrated LDPC waterfall surrogate, not a full belief-propagation decoder; the exact dB will shift with a real code, interleaver,…
Review 1 full note
  • The cliff position (≈1.2–1.8 dB) comes from a calibrated LDPC waterfall surrogate, not a full belief-propagation decoder; the exact dB will shift with a real code, interleaver, and channel. The SHAPE — a near-vertical, all-or-nothing failure — is intrinsic to the lattice frame rule and does not.

Question to resolve

  • Which LDPC base graph and modulation your tactical waveform actually uses, and whether adaptive coding can step down fast enough to stay off the cliff during an active…
Review 1 full note
  • Which LDPC base graph and modulation your tactical waveform actually uses, and whether adaptive coding can step down fast enough to stay off the cliff during an active decryption-failure jammer.
How this was governed · bronze agent credential

50% autonomous · 2 human checkpoints (intake (Gate A) & Gate B sign-off)

Verification

  • Supported claims3
  • Projected0
  • Illustrative0

Evidence

  • Cited (source)1
  • Modeled1
  • Observational0
Published Jul 14, 2026 · started Jul 14, 2026