A curation gate for growing any catalog, registry, or dataset from research notes that may be embellished or wrong. Each iteration takes one unresolved candidate, re-verifies its core claim against the primary source (not the note), classifies it precisely, and runs a mechanical liveness check on its canonical URL — then records a verdict: included with evidence, or dropped with a written reason (unverifiable, duplicate, miscategorized). Distinct from ground-every-claim (which grounds prose in citations): this produces a per-item include/reject ledger, and treats a missing or social-only source as grounds to drop. A false omission is cheaper than a fabricated entry.
Loop
- Goal
- every candidate item carries a final verdict — included with a resolving official source and an independent re-check, or dropped with a written reason — and no included URL fails the liveness sweep
- Exit when
- the liveness sweep prints ALL_LIVE and no candidate remains without an evidence-backed include/drop verdict
Check between iterations
xargs -n1 -I{} curl -s -o /dev/null -w '%{http_code} {}\n' < sources.txt | grep -v '^2' || echo ALL_LIVE
$
Kickoff prompt
Start the "Verify or Drop" loop.
Goal: every candidate item carries a final verdict — included with a resolving official source and an independent re-check, or dropped with a written reason — and no included URL fails the liveness sweep
Max iterations: 8
Between iterations run: xargs -n1 -I{} curl -s -o /dev/null -w '%{http_code} {}\n' < sources.txt | grep -v '^2' || echo ALL_LIVE
Exit when: the liveness sweep prints ALL_LIVE and no candidate remains without an evidence-backed include/drop verdict
Step 1: List every candidate in sources.txt with its claimed source URL, pick the first unresolved one, fetch its primary source yourself, and record an include-with-evidence or drop-with-reason verdict.
›