AI disclosure

AI-assisted prose, human numbersEvery benchmark from real silicon

Yeah I use AI. Mostly Claude (Opus 4.x), sometimes Sonnet for cheap edits. Pretending I don’t would be dishonest and you’d probably figure it out anyway, so here’s the actual split of what it touched and what it didn’t.

What AI helped with

• Drafting first passes of long-form prose (architecture, threat-model writeups) which I then edit.
• Code review on Rust PRs, mostly catching unwrap() slippage and clippy noise.
• Brainstorming experiments: BabyBear vs Goldilocks, Poseidon depth sweep, cross-ISA comparisons. Way more productive with a rubber duck that knows STARK params.
• Boring typing: Astro components, sidebar entries, TOML loaders. The kind of stuff where doing it from scratch is just busywork.

What AI did not touch

• The UMAAL inline asm. Wrote that myself after reading the ARMv8-M reference manual. Asm suggestions from AI at this level are mostly wrong, I checked.
• The benchmark numbers. Every result lives in benchmarks/runs/<date>-<slug>/result.toml, written by firmware running on a real Pico over a real USB cable. AI never produced a millisecond figure.
• Design calls. RP2350 as target, SRAM placement, Goldilocks over BabyBear, stopping Poseidon depth opt at sub-1ms delta. Human decisions on real data.
• Negative-result calls. When BabyBear landed at +69.7 % slower than expected, choosing to publish that as a finding instead of burying it was mine.
• The postmortems. Real bugs I chased. AI sometimes cleans the prose, never invents the bug.

How the numbers stay honest

The pipeline:

1. Compile on dev machine.
2. scp to the Pi 5.
3. Manual BOOTSEL reset, no auto-reset trick.
4. picotool load, firmware runs the bench, prints over USB serial.
5. dd if=/dev/ttyACM0 bs=1 count=N captures.
6. Parse to result.toml, commit.

No AI step in that chain. Number looks wrong? git blame the TOML file, the run that produced it is right there.

Models

Claude Opus 4.x for most work. Claude Sonnet sometimes for cheap edits where the heavy model is overkill. No other vendors. Not on principle, just hadn’t needed to swap.

Why bother

There’s a class of project that publishes AI-generated benchmarks pretending a human ran them. That poisons the well for everyone doing this properly. Easier to just be upfront.

If a number doesn’t reproduce, open a GitHub issue.