Evidence
Evidence for the tools themselves
Does the extra visibility help the next decision?
A fact about a program can be correct while the edit it suggests is wrong. So we test what a reader does with the evidence: find the real problem, keep the behavior the program must keep, and say when the evidence is not enough.
A boundary decision pilot, including its failed first round
Readers judged ten proposed edits across six programs from either the source or an operator-selected packet of tool records. The first packet failed by one answer; with the missing definition added, fresh readers on both sides answered all ten.
| Packet | Artifact correct | Source correct | Gate |
|---|---|---|---|
| Initial | 9 / 10 | 10 / 10 | fail |
| With the located definition | 10 / 10 | 10 / 10 | pass |
| Fresh CLI readers, same revised packet | 10 / 10 | 10 / 10 | pass |
The revised packet plus questions was 9,960 UTF-8 bytes against 30,485 for the source: input bytes, not measured model tokens.
A fresh CLI pair reported 17,952 tokens for the packet reader and 23,099 for the source reader.
One reader pair per round on selected tasks does not establish general product effectiveness. Method, reproduction and limitations · initial failing verdict · revised passing verdict · runtime trial verdict
Ask your agent about anything here; it has the detail.
Where this fits
What this is about: Fact Tooling