# OpenAI web_search — Long-tail coverage (Q11–Q22), extracted from the scored CSV

Source: `_human-tasks/review-artifacts/openai/OPENAI-scored-run-export.csv`, run_id `2026-08-16-r1`,
rows filtered to `type=niche-technical` (Q11–Q22, the long-tail/niche-technical block), mode=`answer`.
Exact rows (trimmed to the relevant columns: query_id, latency_ms, answer_correct, citations_total,
citations_verified, citations_false):

```
Q11  6848.2ms  answer_correct=1.0  citations: 1 total, 1 verified, 0 false
Q12 15220.9ms  answer_correct=1.0  citations: 4 total, 4 verified, 0 false
Q13  6866.1ms  answer_correct=1.0  citations: 2 total, 2 verified, 0 false
Q14  6521.6ms  answer_correct=1.0  citations: 2 total, 2 verified, 0 false
Q15  5219.6ms  answer_correct=1.0  citations: 1 total, 0 verified, 0 false
Q16 11128.0ms  answer_correct=1.0  citations: 4 total, 4 verified, 0 false
Q17  7913.6ms  answer_correct=1.0  citations: 3 total, 3 verified, 0 false
Q18  8706.1ms  answer_correct=1.0  citations: 4 total, 4 verified, 0 false
Q19 15120.5ms  answer_correct=1.0  citations: 8 total, 8 verified, 0 false
Q20  6390.7ms  answer_correct=1.0  citations: 1 total, 1 verified, 0 false
Q21  5876.8ms  answer_correct=1.0  citations: 0 total
Q22 19291.9ms  answer_correct=0.0  citations: 6 total, 6 verified, 0 false — notes: "blamed segment
                                    durations; verified fix is aresample=async=1"
```

**Long-tail/niche-technical answer accuracy: 11/12 correct = 91.7%.** The single miss is **Q22**
("FFmpeg concat demuxer audio sync drift fix"), which is a documented true negative — per NOTES.md
this query scored 0 across **every** tool-mode in the benchmark (13 modes), because no index
associates the canonical fix documentation with this query's phrasing. OpenAI's miss on Q22 is
consistent with the field-wide pattern, not an OpenAI-specific weakness.

Zero false citations across the 11 correct answers (32 total citations checked, 31 verified — Q21's
single citation-total-0 row reflects an abstention-style short answer, not an error).

**Disposition: scorable, 91.7% (11/12) answer-correct on the long-tail block**, with the one miss
being the benchmark's known field-wide Q22 true negative rather than an OpenAI-specific coverage gap.