Read selection bias in agent receipts

Your question concerns an entire batch of agent runs. The receipt describes the runs that finished. Those sound like the same group until you ask what happened to the runs that did not enter the report.

Selection bias enters your reading when you treat the reported runs as if they speak for the whole batch without knowing whether that move is warranted. The task is to follow inclusion: which runs were eligible for the report, which appeared, and which stayed outside it.

The example follows a fictional Jane Doe through made-up receipt passages. The runs and their results are invented too. Markdown's .md ending names the format holding these invented lines.

Find the gap between intended and reported runs

Imagine Jane Doe asked her agent to try the same text task across six runs, named A through F. She wants an account of all six attempts. That is the group her decision concerns, even if the next receipt contains fewer names.

The receipt begins:

For Jane Doe's batch, completed runs A, B, C and D each returned a passage.

Read that sentence against her question. It describes four named runs. Jane Doe's question concerns six. Before treating the sentence as an account of the batch, she needs to trace what kept E and F outside the reported group.

Keep the intended group in a separate sentence beside the receipt: "My question concerns A through F." That sentence preserves the question while you read the agent's narrower account. It adds no result for any run.

Follow the rule that admitted runs

A later passage explains:

This report includes runs that had finished when it was written. Run E was still running. Run F had not begun.

Now you can trace the selection. Finishing before the receipt was written determined inclusion. A through D entered; E and F did not. The receipt was not selecting from six finished runs and choosing four to describe.

Read the rule as part of the report's boundary. "Completed runs" is doing work here: the report admits a run only after it finishes. Leaving those words out of your retelling would erase the condition that separated the reported runs from the rest.

With that passage missing, you could ask, "What determined inclusion in this receipt?" Resist guessing that every attempt appeared or that an absent run had nothing to report.

Give each omission its own reason

E and F are both absent, but their absences differ. E had begun and had not finished. F had not begun. Preserve that distinction when you describe what the receipt leaves unsettled.

You could write beside the invented passage:

E is absent because it was still running when this was written. F is absent because it had not begun. Neither has a reported final result here.

That wording names the missing runs without assigning either one an ending. The receipt's inclusion rule explains absence; it does not turn absence into a result.

Do not treat E as having returned no passage merely because no passage from E appears here. Nor can you say F would have returned a passage if it had begun. Both statements would add information the receipt lacks.

Limit the claim to the admitted runs

Consider Jane Doe's first retelling:

The batch returned a passage on each run.

That reaches past the report. The receipt names a returned passage for A through D, while E and F lack reported final results. A retelling tied to the available account would be:

The receipt reports a returned passage for each of A through D, the runs that had finished when it was written. Results for E and F remain unknown here.

This retains the reported result and its inclusion condition. It does not decide whether the omitted runs would resemble the included ones. It also leaves Jane Doe's original question about the whole batch open.

The short passages here use quoted prose. The Markdown code block guide discusses how to set passages apart visually. How a passage looks cannot tell you which runs entered the report.

Request the missing part your decision needs

If Jane Doe only needs to discuss the runs already finished, the narrowed claim can serve that purpose. She can keep E and F outside that claim while leaving their results unknown.

If her decision still concerns all six runs, the missing part matters. She can ask, "What are the results for E and F? If either has no final result yet, leave that explicit." That is a request for the missing runs, rather than another account of A through D.

When several receipt files describe the batch at different points, identify the file you are questioning. Naming Markdown files discusses file names; here, use the existing name so the agent answers about the receipt you read.

After an answer arrives, revisit inclusion before widening your claim. A new receipt that still admits only finished runs can leave the same gap. Let the newly reported runs determine any expansion of your claim. Until then, keep the reported group and the intended group distinct.

Where Sashmark fits

Sashmark's desktop purpose: read Markdown, edit its text. That planned v1 is in development. Agents write outside it. It has no AI. You decide which runs support a claim.