The receipt

Every value TrialExtract extracts carries a receipt. Click a number anywhere in the evidence table and its receipt opens: the exact place it was read, how it was computed if it was computed, and whether a second pass agreed. You are looking at the answer to the only question that matters at revision, “where did this number come from,” before anyone asks it.

Three facts, kept apart

A receipt holds three distinct facts, and the interface never lets them blur into one another.

• Source evidence is provenance: where the number was read. You get the verbatim quote from the paper, gold-highlighted, with the source-block labels and a locator underneath. The specific reported numbers that built the value are lit up inside the quote. Any number the value depends on but that cannot be found in the quote is listed as “not located,” never quietly restated elsewhere.
• How this value was derived appears only when the value was computed rather than read straight off the page. It is a numbered cascade of steps ending in the reported effect, with its published method reference cited once beneath the cascade.
• Independent verification is whether a second, independent pass agreed with the value. It is a separate fact from provenance, and the receipt keeps it separate.

The point of the separation is the trap it prevents: a number can be located cleanly in the source and still be wrong. Provenance and verification answer different questions, so they never share a badge.

Reported or derived, native units first

The headline is always the native measure in the paper’s own units, tagged with its abbreviation: MD, HR, OR, RR, RD, and so on. Underneath, dimmed, sits the standardized effect as Hedges g with its confidence interval. The paper’s number leads; the comparable one follows. This is deliberate, so you never mistake a conversion for what the authors actually printed.

Two chips tell you what happened to get there:

• derived means the value was computed from reported inputs, not read directly.
• lossy means standardizing to an SMD discarded the native unit. It is a flag that a real quantity, the unit, was left behind in the conversion.

When a value is derived, the “How this value was derived” section shows every step. A common one: a between-group confidence interval becomes a standard deviation (SE-from-CI), the SD feeds the standardized effect, and a small-sample J correction turns Cohen’s d into Hedges g. The receipt renders each step and re-derives the result independently, so any drift from the reported value is flagged rather than hidden. The formulas themselves live in the notes: SE to SD and the Hedges g small-sample correction.

Located is not verified

The verification section is the one people misread, so the copy is blunt about it. When no second pass has run, it says: independent verification has not run for this value. Provenance located the number in the source; that is not the same as a second reader agreeing with it. Until a verification pass runs, the receipt tells you exactly that, rather than dressing a located number up as a checked one.

When the receipt has nothing to show

Honesty runs both ways. Where a value genuinely has no verbatim source span, the receipt says so instead of inventing one. Adverse-event counts, for instance, have no quotable source passage to cite, and the section states that plainly rather than fabricate a highlight.

Quality problems surface here too, not buried. A paper flagged for fabrication pressure carries a red line telling you to review it with elevated scrutiny, and any other quality notes on the value appear alongside it. The receipt is where a value has to answer for itself, so it never gets to stay quiet about its own weak spots.

Once you can read a receipt, you can act on one. That is the next stage: Reviewing evidence, where you accept, edit, or contest each value, and The quality queue, where the flagged ones come to you first.