# Valyu — stability re-run comparison (resolves "no artifact" blocker)

The published Gate-1 closure note's per-tool stability table lists only 7 of ~13 tool/modes and omits
Valyu, backing its position with only the vague prose "~28–30%, comfortably inside the noise band."
Below are the actual computed numbers, using the same method that exactly reproduces the closure
note's published figures for the 7 tools it does list (verified before trusting this output).

**Run definitions:**
- **Run 1** = the union of run_ids `2026-08-15-r1` (freshness, Q42–Q47) + `2026-08-16-r1` (Q01–Q22) +
  `2026-08-16-r2` (Q23–Q41, Q48–Q52) in `_analysis/all-tools-scored.csv` — together these three
  batches cover all 52 queries once, with zero duplicates, and constitute the original pass.
- **Run 2** = run_id `2026-08-17-r3` in `_archive_run_2026-08-17-stability/_analysis/all-tools-scored.csv`
  — the genuine same-query re-run, one shot, 52 queries. (Note: this supersedes an earlier r3 archive
  that was a stale copy of r2 showing a false 100% agreement — see `_human-tasks/GATE-1-CLOSED-2026-08-17.md`
  §Pipeline note #3. This file uses the corrected, independently-rescored r3.)

| Mode | Run 1 top-3 (of 47 answerable) | Run 2 top-3 | Δ | Agreement (41 non-freshness queries) |
|---|---|---|---|---|
| valyu/search | 29.8% (14/47) | 27.7% (13/47) | −2.1 pts | 95.1% (39/41) |
| valyu/answer | 27.7% (13/47) | 31.9% (15/47) | +4.3 pts | 97.6% (40/41) |

- Agreement denominator excludes the 6 freshness queries (Q42–Q47), which are expected to legitimately
  drift between runs as world state changes — same convention the published 7-tool table uses.
- `valyu/search` mismatches (of 41 comparable): Q05 (hit in Run 1 → miss in Run 2), Q10 (miss → hit).
- `valyu/answer` mismatch: Q10 (miss → hit).

**Disposition:** Both figures sit comfortably inside the ±10.6-point noise band and 88–100% agreement
range established by the 7 published tools. This is the hard number that supports the closure note's
"comfortably inside the band" claim, and it holds — Valyu's ranking position is confirmed stable
across runs, same conclusion as before but now with a citable artifact instead of prose.