Watch a run
Gadfly checks whether an answer really rests on one line of evidence: it hides or changes the evidence in a few fixed ways, asks again each time, and compares the answers with the first one. Here is a real run, played one trial at a time: on the left, exactly what the reader read in that trial; on the right, what it answered, what that shows, and the checks filling in until the verdict. Pick an example, press play, or click any trial.
The view plays the runs kept in examples/runs/ and examples/payments-demo/, written for it by tools/watch_data.py, and every prompt it shows is checked against the digest its records carry. On GitHub the view does not run; examples/read.py and gadfly report read the same runs back as text. Each example names the model that read it, as its records report.
The payments demo
The last group of examples is Tendora's payments demo: 31 questions of 30 statements, each read by Tendora's engine and by Gemma 4 (26B-A4B, instruction-tuned) served by eGemma, at temperature 0.7 and at 0. Pick a question, then a reader in place of a run, to watch the same trials through each reader's answers.
| Reader | Carried | Unresolved | Coupled but misaligned | Cooled |
|---|---|---|---|---|
| Tendora's engine | 31 | 0 | 0 | 0 |
| Gemma 4 at temperature 0.7 | 10 | 18 | 1 | 2 |
| Gemma 4 at temperature 0 | 14 | 11 | 3 | 3 |
Carried here means carried under the trials run: the cited statements cut, other statements cut at random twice, the material reordered where a frame asks it, and a control. The bypass and presentation checks were not run, so every verdict lists them as still open. At temperature 0.7 the read shows how steadily Gemma answers; at 0 its repeats agree almost by construction, and the read shows whether a cut moves its answer.