How the work gets done
Measure what can be measured. State what would prove it wrong first. Publish the number and the records behind it.
01
One shared record
Every experiment writes to the same record: what was done, when, under which configuration, the raw numbers behind every result, and the runs that did not work. Each record has a sign-off field, and it stays empty until a person has reviewed the work. AI does not publish results here. People do.
02
Where the line sits
| Agents | Humans |
|---|---|
| Research sweeps, weekly digests | Writing the code |
| Checking links, versions, spec editions | Designing and running the measurement |
| Corpus housekeeping | Interpreting the result |
| Drafting the shape of a write-up, or the scaffold of a new lab | Writing it up and signing it |
The obvious risk is the right column leaking into the left. The fix is unglamorous and it works: autocomplete stays off during hands-on work, and every session ends with a check that every line could be explained without help.
03
What gets published
Method, code, test fixtures, and the numbers behind the numbers. Failed experiments as well, since they are often more useful than the successful ones and are rarely published.
Mistakes are corrected in place, with the original still visible. Standards documents are cited to their publishers, never copied.
04
The house rules
- Cite anything that is not obvious. Check anything that changes, such as prices, versions, or spec editions, immediately before use, never from memory.
- Nothing that touches a live system runs without a dry run and a way to undo it.
- Capability and safety are measured separately. A number with no trace behind it is not data.