Appendix A · Part IV: Reference
Reference Tables and Deadlines
The numbers from all nine chapters in one place. Sizes and standards are stable. Dates, qubit counts, and version numbers move, so confirm the ones you act on against the primary source.
In this chapter
- Post-quantum cryptography
- Confidential computing
- Supply chain
- Agentic systems
- Platform engineering
- Verification
- Quantum hardware and software
- Energy and carbon
- Runtimes and isolation
Post-quantum cryptography
What a cryptographically relevant quantum computer breaks.
| Primitive |
Classical security |
Quantum attack |
Status |
| RSA-2048, RSA-4096 |
112 to 140 bits |
Shor, polynomial |
broken |
| ECDH, ECDSA, Ed25519 |
128 bits |
Shor, polynomial |
broken |
| Finite-field Diffie-Hellman |
varies |
Shor, polynomial |
broken |
| AES-128 |
128 bits |
Grover, 2^64 coherent |
weakened |
| AES-256 |
256 bits |
Grover, 2^128 coherent |
adequate |
| SHA-256, collision |
128 bits |
limited gain |
adequate |
ML-KEM, FIPS 203.
| Parameter set |
Category |
Encaps key |
Decaps key |
Ciphertext |
| ML-KEM-512 |
1 |
800 B |
1,632 B |
768 B |
| ML-KEM-768 |
3 |
1,184 B |
2,400 B |
1,088 B |
| ML-KEM-1024 |
5 |
1,568 B |
3,168 B |
1,568 B |
| X25519 |
classical |
32 B |
32 B |
32 B |
ML-DSA, FIPS 204.
| Parameter set |
Category |
Public key |
Private key |
Signature |
| ML-DSA-44 |
2 |
1,312 B |
2,560 B |
2,420 B |
| ML-DSA-65 |
3 |
1,952 B |
4,032 B |
3,309 B |
| ML-DSA-87 |
5 |
2,592 B |
4,896 B |
4,627 B |
| Ed25519 |
classical |
32 B |
64 B |
64 B |
SLH-DSA, FIPS 205, SHA-2 variants.
| Parameter set |
Public key |
Signature |
Signing |
| SLH-DSA-128s |
32 B |
7,856 B |
slow |
| SLH-DSA-128f |
32 B |
17,088 B |
fast |
| SLH-DSA-192s |
48 B |
16,224 B |
slow |
| SLH-DSA-256s |
64 B |
29,792 B |
slow |
TLS handshake growth.
| Component |
Classical |
Post-quantum |
Growth |
| Client key share |
32 B |
1,216 B |
+1,184 B |
| Server key share |
32 B |
1,120 B |
+1,088 B |
| Leaf certificate key |
32 B |
1,952 B |
+1,920 B |
| Each signature in the chain |
64 B |
3,309 B |
+3,245 B |
| Typical total handshake |
~3.5 KB |
~16 KB |
~4.5x |
Deadlines as published in 2026. Confirm against the primary source.
| Requirement |
Scope |
Date |
| FIPS 203, 204, 205 published |
ML-KEM, ML-DSA, SLH-DSA |
August 2024 |
| HQC selected as a backup KEM |
code-based, in development |
March 2025 |
| CNSA 2.0 in new acquisitions |
US national security systems |
January 2027 |
| Key establishment migrated |
US federal high-value assets |
2030 |
| Signatures migrated |
US federal high-value assets |
2031 |
| CNSA 2.0 full transition |
US national security systems |
2030 to 2033 |
Mosca's inequality: if confidentiality lifetime plus migration time exceeds
time to a cryptographically relevant quantum computer, data sent today is
already exposed.
Confidential computing
| Technology |
Boundary |
Trusted computing base |
Overhead |
| Intel SGX |
process enclave |
enclave code only |
10 to 40 percent |
| Intel TDX |
virtual machine |
guest OS plus TDX Module |
2 to 8 percent |
| AMD SEV-SNP |
virtual machine |
guest OS plus SEV firmware |
under 5 percent |
| Arm CCA Realms |
realm |
realm software plus monitor |
varies |
| AWS Nitro Enclaves |
isolated VM, no network |
enclave image |
low |
| NVIDIA confidential GPU |
GPU plus CPU TEE |
driver plus enclave |
5 to 15 percent |
| Workload shape |
Confidential overhead |
| CPU-bound, small working set |
2 to 5 percent |
| CPU-bound, large working set |
5 to 12 percent |
| Network-heavy service |
10 to 25 percent |
| Storage-heavy service |
10 to 30 percent |
| Process enclave, frequent exits |
30 to 100 percent |
| Virtual machine start |
plus 1 to 5 seconds |
Attestation check order: signature chains to a trusted root, chip not revoked
and firmware current, debug disabled and policy flags acceptable, measurement
on the allowlist, and report data binds this session's key.
Supply chain
| SLSA level |
Requirement |
| 0 |
nothing |
| 1 |
provenance exists |
| 2 |
provenance signed by a hosted build service |
| 3 |
hardened isolated build, provenance unforgeable by the build |
| 4 |
under development |
| Runtime control |
Overhead |
| Seccomp profile |
1 to 5 percent |
| Capability drop |
0 |
| eBPF LSM policy |
1 to 3 percent |
| Identity network policy |
under 1 percent |
| eBPF runtime detection |
1 to 3 percent |
| Read-only root file system |
0 |
SVID lifetime: one hour typical, rotated at roughly the half-life. A stolen
credential expires within that window with no revocation infrastructure.
Agentic systems
Reliability falls geometrically with agent count.
| Agents in a chain |
At 95 percent per agent |
At 99 percent |
| 1 |
95.0 percent |
99.0 percent |
| 3 |
85.7 percent |
97.0 percent |
| 5 |
77.4 percent |
95.1 percent |
| 10 |
59.9 percent |
90.4 percent |
| 20 |
35.8 percent |
81.8 percent |
| Tool risk class |
Control |
Sandbox |
| Read, internal |
rate limit, per-user scope |
none |
| Read, external |
allowlist, timeout, size cap |
egress policy |
| Write, reversible |
audit log, idempotency key |
per-user credential |
| Write, irreversible |
human approval |
per-user credential |
| Code execution |
full isolation |
microVM, no network |
| Memory layer |
Store |
Lifetime |
| Working |
the context window |
one run |
| Episodic |
vector index plus metadata |
weeks to months |
| Semantic |
key-value or graph |
indefinite |
| Procedural |
prompt or policy store |
until changed |
| Metric |
Healthy direction |
| Deployment frequency |
higher |
| Change lead time, commit to production |
lower |
| Change fail rate |
lower |
| Failed deployment recovery time |
lower |
| Time to first production deploy, new engineer |
lower |
| Golden path adoption rate |
higher |
| Self-service ratio against tickets |
higher |
| Build and test wall-clock time |
lower |
| Flaky test rate |
lower |
| Platform net promoter score |
higher |
| Environment strategy |
Provision |
Isolation |
Fidelity |
| Shared long-lived staging |
none |
none |
high |
| Per-pull-request ephemeral |
5 to 15 min |
complete |
medium to high |
| Local with containers |
1 to 3 min |
complete |
low to medium |
| Shadowed on production traffic |
minutes |
read-only |
highest |
Verification
| Technique |
Finds |
Runs where |
| Example tests |
anticipated cases, regressions |
every commit |
| Property tests |
boundary and degenerate inputs |
every commit |
| Stateful property tests |
sequence-dependent bugs |
every commit |
| Coverage-guided fuzzing |
crashes, memory safety bugs |
continuously |
| Sanitizer builds |
silent memory errors |
nightly and under fuzzing |
| Fault injection |
error handling paths |
nightly |
| Deterministic simulation |
distributed interleaving bugs |
continuously |
| Chaos experiments |
operational failures |
production, scheduled |
| Formal specification |
design-level protocol errors |
at design time |
Five property patterns: round trip, invariant, oracle, metamorphic relation,
algebraic law.
A safety counterexample is a finite trace. A liveness counterexample is a
cycle.
Quantum hardware and software
| Property |
Superconducting |
Trapped ion |
Neutral atom |
| Two-qubit gate time |
50 to 500 ns |
50 to 300 us |
0.2 to 2 us |
| Two-qubit gate error |
1e-3 to 5e-4 |
1e-4 to 1e-3 |
1e-3 to 5e-3 |
| Coherence time |
100 to 500 us |
seconds to minutes |
1 to 10 s |
| Connectivity |
degree 2 to 3 |
all-to-all in a trap |
reconfigurable |
| Useful depth, 2q gates |
1,000 to 5,000 |
100s to 1,000s |
100s to 1,000s |
| Mitigation method |
Overhead |
Bias |
| Measurement mitigation |
fixed calibration |
small |
| Dynamical decoupling |
negligible |
none introduced |
| Randomized compiling |
10 to 100 variants |
none introduced |
| Zero-noise extrapolation |
3 to 5x circuits |
model-dependent |
| Probabilistic error cancellation |
exponential in total noise |
unbiased |
Track logical qubit count, code distance, logical error rate, and decoder
latency. A physical qubit count without an error rate indicates nothing.
Energy and carbon
| Component |
Idle |
Typical load |
Peak |
| CPU, 64 core |
40 to 70 W |
150 to 250 W |
350 to 400 W |
| DRAM, 1 TiB |
30 to 50 W |
60 to 90 W |
120 W |
| Accelerator, per device |
60 to 100 W |
400 to 700 W |
700 to 1,400 W |
| Whole server, no accelerator |
150 to 250 W |
400 to 600 W |
900 W |
| Whole server, 8 accelerators |
700 to 1,000 W |
4,000 to 6,000 W |
12,000 W |
SCI equals energy times carbon intensity, plus embodied emissions, divided by a
functional unit. The functional unit is what makes it a rate rather than a
total.
| Lever |
Typical reduction |
| Right-size over-provisioned resources |
20 to 50 percent |
| Raise cluster density |
20 to 40 percent |
| Scale to zero when idle |
30 to 90 percent of idle |
| Temporal shifting of batch work |
15 to 40 percent |
| Spatial shifting to a clean region |
up to 30x for that work |
| Algorithmic improvement |
10x to 1000x |
| Storage tiering and deletion |
50 to 90 percent of storage |
| Storage tier |
Relative cost |
Access latency |
| Local NVMe |
100 |
microseconds |
| Network block storage |
40 |
sub-millisecond |
| Object storage, standard |
10 |
tens of milliseconds |
| Object storage, archive |
1.5 |
minutes to hours |
| Deep archive |
0.5 |
up to 12 hours |
| Deleted |
0 |
not available |
Runtimes and isolation
| Boundary |
Cold start |
Memory |
Enforced by |
| Process plus seccomp |
1 to 50 ms |
varies |
kernel |
| Container |
100 to 500 ms |
30 to 200 MiB |
kernel namespaces |
| MicroVM |
~125 ms |
5 MiB plus guest |
hypervisor |
| V8 isolate |
1 to 5 ms |
2 to 10 MiB |
the JavaScript engine |
| WebAssembly module |
under 1 ms |
1 to 5 MiB |
the Wasm runtime |
Roughly 70 percent of serious vulnerabilities in large C and C++ code bases are
memory safety defects.
Replacement sequence, at every scale: interpose, dual path, shadow, canary,
switch, contract. Only the last step is irreversible.