An evidence-grounding pass over any prose deliverable (answer, PR body, design doc). Each iteration locates one hedged or memory-based claim, verifies it against a real artifact, and rewrites it to carry that citation. The loop exits only when no hedging language remains and every surviving claim points at a checkable source — not a recalled prior.
Loop
- Goal
- every factual claim in the target text is backed by a verifiable citation (file:line, command output, or URL) with no hedging language left unresolved
- Exit when
- the check prints NO_HEDGES and every remaining claim cites a file:line, command output, or URL
Check between iterations
grep -rniE "\b(i think|probably|i believe|should be|presumably|likely|maybe|seems? to|as far as i know|iirc)\b" "$TARGET" || echo NO_HEDGES
$
Kickoff prompt
Start the "Ground Every Claim" loop.
Goal: every factual claim in the target text is backed by a verifiable citation (file:line, command output, or URL) with no hedging language left unresolved
Max iterations: 8
Between iterations run: grep -rniE "\b(i think|probably|i believe|should be|presumably|likely|maybe|seems? to|as far as i know|iirc)\b" "$TARGET" || echo NO_HEDGES
Exit when: the check prints NO_HEDGES and every remaining claim cites a file:line, command output, or URL
Step 1: List every factual claim in the target text and flag the ones asserted from memory rather than a checked artifact, then verify and re-cite the first flagged claim.
›