Stage 3 audit 2026-08-13: partial results coded successful, self-record links, quote-source provenance

A hand audit of 30 random `successful` and 15 random `failed` rows in the 8b3d01ec5ad7 render (seed 42) found no failure coded as success and clean evidence on the failed side, but three defects:

1. **Partial→successful drift (the substantive one).** 6 of 30 sampled successes are arguably `mixed`: the coder takes the authors' summary sentence at face value and never weighs a discordant clause later in the same abstract. Recurring shapes: a headline effect replicated while a secondary result ran opposite to the original (10.31234/osf.io/mhpu9), and partial coverage framed as success (11 of 17 proteins, 10.1016/j.bpsgos.2024.100332). Upper bound if all such rows re-coded: successful drops from 57% to ~46% of settled rows. Candidate fix: prompt the outcome coder to state whether any reported result contradicts the original before settling on `successful`.

2. **Self-record link.** 10.31234/osf.io/bqzj9_v1: `title_o` is the replication's own record (a version/duplicate of itself), not the original (Triplett 1898). A same-title, same-author-list candidate should be refused at link time.

3. **Quote-source provenance.** 10.17605/osf.io/5bms8: `out_quote_source = abstract` but the row's stored abstract is an OSF registration stub that contains none of the quoted text — the quote came from a document the row does not show.

Full audit table in the 2026-08-13 handover round.

Start with the 2026-08-13 handover round and its full audit table, using the 8b3d01ec5ad7 render with seed 42 to reproduce the three findings. Trace the outcome-coder prompt, same-title/same-author link check, and quote-source field; done means mixed outcomes are not settled as successful, self-records are rejected, and every quote source is represented by the row.

Written by the indexing model from the issue text.