• SoC: Jetson Orin Nano (6x ARM Cortex-A78AE)
• RAM: 7619 MB unified (CPU+GPU)
• GPU: 1024 CUDA cores (shared memory)
• Storage: 2TB NVMe (1.7TB free)
• OS: Linux 5.15.148-tegra (aarch64)
• Compiler: GCC with -O2

Each benchmark compiles a standalone C program linked against flux_vm.c, runs 10,000 iterations, and measures wall-clock time. All benchmarks use clock() for timing (CPU time, not wall clock).

1. VM dispatch is nearly free for hot paths — 440 Mops/s vs 474 Mops/s native (8% cost)
2. Confidence tracking costs only 14% extra — CONF_ADD (379) vs ADD (440). Think Tank's OPTIONAL decision validated.
3. Mixed workloads (MOVI+ADD) are the real cost — 32% overhead due to instruction fetch/decode variety
4. The "fence price" is ~1.3x for realistic agent workloads — well worth it for portability, A2A, and confidence

• Hermes-405B: Best for creative/narrative tasks. Max 1 call per batch (expensive + can timeout). Use for synthesis, not analysis.
• Nemotron-120B: Best for practical engineering with code. Fast (8.6s), cheap, high quality. Use heavily.
• Step-3.5-Flash: Produces thinking tokens, not direct output. Skip for term generation. May work for reasoning chains.
• Seed-2.0-mini: REMOVED from DeepInfra (404). Was best for devil's advocate.
• Qwen3.5-397B: TIMEOUT consistently. Expensive + unreliable. Skip for now.
• phi-4: Good for technical proposals. Reliable. Use for opcode/hardware design tasks.

1. Use Nemotron first for any code generation task (fast, cheap, excellent quality)
2. Hermes-405B last in batches (expensive, timeouts) — use only when quality ceiling matters
3. Skip Step-3.5-Flash for HAV — it thinks instead of generating
4. Pre-verify model availability before crafting prompts (Seed-2.0-mini wasted effort)
5. Benchmark after every VM code change — document delta
6. Opcode lookup before bytecode construction — wasted time on wrong opcodes (0x10 ≠ IADD)

• flux-runtime-c — The VM being benchmarked
• flux-asm — Assembler
• flux-disasm — Disassembler
• flux-isa-unified — ISA specification
• cuda-bytecode-optimizer — Bytecode optimizer
• cuda-edge-lint — Edge device linter

These benchmarks are the pulse of a living system. The fleet is not a collection of libraries — it's an organism where every vessel contributes signals. The fence at 0x44 (the abstraction boundary between native code and bytecode) is not a wall but a membrane: it lets useful things through while keeping the system sovereign. The 14% confidence overhead is the cost of self-awareness — the fleet knows what it knows and how well it knows it. This is the post-SaaS era: software doesn't just run, it metabolizes, adapts, and remembers.