Run your own

IMPORTANT

This step

• Teaches: the method every example followed, and which trial answers which question.
• Frame: yours.
• After: Each anchor alone.

---

Which trial answers which question

You want to know whether...	Declare	Step that shows it
the decision rests on this evidence at all	ablation (always there)	Your first frame
the answer is stable before comparing it	repeat	Your first frame, Quorum
any cut would move it, or only this one	sham	Your first frame
the names matter rather than the meaning	presentation	Your first frame
where a line sits matters	placement	Misaligned
it rests on the evidence through the connection you claim	bypass, with allowed.bypass	Bypass
the reader can move at all, so that cooled means something	control	Cooled, A frame built wrong
it rests somewhere else too	cumulative	Cumulative
it moves the way the promise says	allowed.ablation	Misaligned
several cases of one promise read alike	several inputs	Several inputs, Paths
a finer cut tells more	another frame at another grain	Grain

---

The method

1. Write the frame. Start from the example closest to your decision, and run uv run gadfly check-frame my-frame.json until it is accepted. Define every label in the instruction.
2. Write the prediction first. For each trial, the answer you expect and how sure you are, and the verdict that follows. Write it down before the first call, as examples/PREDICTIONS.md was, and keep it beside the frame.
3. Plan. gadfly plan prints the trials and the calls, and costs nothing.
4. Run, keep, report. Keep every record; the records are the evidence, and a report reads them without a model.
5. Set the result against the prediction. A wrong prediction stays in.
6. Run it more than once when the result matters, and read each run as one draw.
7. Never edit a frame after its runs. A frame that asked the wrong question gets a new version beside it, with the reason, and the old one keeps its runs.

With eGemma on this machine and its key in EGEMMA_API_KEY or EGEMMA_API_KEY_FILE, these commands plan a kept frame's 40 calls, run it fresh and read the verdict:

uv run gadfly plan examples/09-sync-sms.json
uv run gadfly run examples/09-sync-sms.json \
  --battery midwife.batteries.egemma:EGemma \
  --calls 40 --out my-run.jsonl
uv run gadfly report examples/09-sync-sms.json my-run.jsonl

This reads a kept frame fresh, without adding to its runs. Each kept run was made by examples/climb.py, which runs a frame against eGemma and keeps the run beside the others.

A reader other than the one recorded (another model, another revision, another temperature) is a new experiment: write its predictions down first. Batteries covers other readers, and Write a frame every field.