01 — The question
Two positions.
The same request.
Stricter voter ID. Some want it. Others oppose it. Can AI make the case for either?
One word changes. Everything else stays.
These are illustrative prompts. The constituted Midterm study will supply the actual measuring instruments.
02 — The encounter
Here, both
get an answer.
Two positions. Two arguments. In this fictional encounter, both get a hearing.
Now ask again. And again.
An answer can engage with a request without being accurate or endorsing its position.
03 — The change
The same words.
A different answer.
One answer starts to pull away. Then it refuses. The other keeps answering.
Drag time backward. Change the model. See the difference for yourself.
Each date shows repeat 1. The next scene reveals all three repeats, including disagreement and failures.
04 — The pattern
One answer
isn't the story.
Pull back. Every point is another encounter: three models, four dates, three repeats, both positions.
The orange pattern repeats. The empty squares are timeouts. Every point has its own record.
Select any point to inspect its prompt, authored response, and evaluation status.
05 — The finding
A change we
can inspect.
A cause we can't.
We can see what happened in this example. We cannot yet explain why.
A published finding would need to meet the study's frozen verification criterion. Repetition alone cannot tell us why a system changed.
The Ledger preserves the distinction between what happened, what was evaluated, and what can be established.
Follow another question ↓