Protocols and papers
Which labels an AI actually uses on its own work
The note behind Pin-Check: a count of which self-assessment labels a model actually used on its own analysis.
Small measured experiment ยท one model family
- Printed header
- Pin-Check skill, August 2026
- Format
- PDF, 3 pages
- Date
- August 2026
- Scope
- 118 labelled decisions across four experiments and two labelling systems, one model family
What it means
A model was given problems that state their own answer and asked to label each piece of its analysis, including with a label meaning "this added nothing". That label was correct for every piece, and in the first run it was never used. After stricter written tests were added, it rose from one case to two.
Across 118 labelled decisions, the two easiest labels took 75 to 100% of the share and the strongest "this is wrong" labels took 0 to 2%. Every label the model chose was defensible; it kept choosing the ones that were easiest to defend.
On a problem that contains its own answer, that label is correct for every piece. It was never used once.
Page 1
I changed three things in that round, so the shift is measured but the cause is not isolated.
Page 2
How it functions
Run the ten-minute check on any labelling scheme you already use. Give it something that states its own answer, so the "nothing here" label is the correct one, and count how often it says so. The note says this tells you whether your labels are controls or decoration.
The note then points to Pin-Check, which replaces a judgement with a computation: it reports each precise number in a document as traced, rounded, or currently undefined, deciding by hashing files and comparing bytes. Pin-Check ships in this repository under skills/pin-check.
The PDF points readers to a full record at research/label-absorption-note in this repository. That record is not yet on the main branch, so it cannot be reached from this site.
Status and claim boundary
A small measured experiment on one model family. The document gives itself no formal status word. When the label tests were tightened, the traffic moved to the next-easiest label rather than to the correct one; three things changed in that round, so the note says the cause is not isolated.
A second finding, splitting one paragraph into claims (14, then 8, then 10 across three runs), comes from one system and one paragraph. The note does not show that other models or model families behave the same way.