How this actually gets done
Measure what can be measured. Say what would prove it wrong first. Publish the number and the receipts.
01
One shared paper trail
Every experiment writes to the same place: what got done, when, under what configuration, the raw numbers behind every result, and the stuff that didn't pan out. Even the hard hours are what makes this work fun. Every record has a sign-off field, and it stays blank until an actual human has looked at the work. We don't let AI take the glory of posting the results.
02
Where the line sits
| Agents | Humans |
|---|---|
| Research sweeps, weekly digests | Enjoying writing the code |
| Checking links, versions, spec editions | Designing and running the measurement |
| Corpus housekeeping | Making sense of what comes back |
| Drafting the shape of a write-up | Writing it up and signing it |
The obvious risk is the right column quietly leaking into the left. The fix is boring but it works: autocomplete stays off during hands-on work, and every session ends with an honest check on whether every line could actually be explained without help.
03
What gets published
Method, code, test fixtures, the numbers behind the numbers. Failed experiments too, since they're usually more useful than the successful ones and almost nobody publishes those.
Mistakes get corrected in place, with the original still visible. Everybody gets things wrong sometimes and there's no point pretending otherwise. Standards documents get cited to their publishers, never copied.
04
The house rules
- Cite anything that isn't obvious. Check anything that changes, like prices, versions, or spec editions, right before it's used, not from memory.
- Nothing touching live or physical gear runs without a dry run and a way to undo it. If you know, you know.
- Capability and safety get measured separately. A number with no trace behind it might be cool, but it's not data.