The Question-Response Audit, as it actually stands

2026-07-28 · continues the 2026-07-23 evaluation at excel-arithmetic-malfeasance

1 · The state in one paragraph

The canonical file is QuestionAudit_PowerBI_TIMED_ONLY_2026-07-23.xlsx: 122 interactions, every one carrying a coach, an agent, a date, an asked time and an answered time. It is built by one command from a 154-row head file. 107 rows pass the gate directly, 15 are hand-read from Teams captures, and 47 are dropped and listed by name because no source on disk holds their times. 107 plus 47 equals the 154 that went in, so nothing vanished quietly. Median answer time across the 122 is 5 minutes, mean 8.7, 90th percentile 16, slowest 75.

2 · Three defects found and fixed today

The original evaluation named mixed-type columns as the defect most likely to make a Power BI dashboard misreport timing. That exact defect had come back in the file shipped yesterday, through the code path that appends hand-read Teams rows.

type check: 0 mixed-type columns (clean)
category check: 2 label collision group(s) -> [['DDE (admin)', 'DDE / DDF Removal'],
                                               ['Coach Stamp / OSHA', 'Coach Stamp / OSHA 500']]
recovered=5 kept=122 (from head 107 + appended 15) removed=47 GATE=PASS
reconcile: 107 kept-from-head + 47 removed = 154 (== rows_in 154: True)

3 · The finding that limits the numbers

The distribution of that unverified source across coaches is uneven, which is what makes it a ranking problem rather than a rounding problem:

Two thirds of Margarita's record comes from the unverified source and under a tenth of Toveta's does. Comparing the two on this file compares capture method as much as it compares coaching. The 47 open rows sit mostly on the same day, 40 of 47 on 7/22, so 7/22 is the day to finish before anyone quotes a per-coach number.

4 · Per-coach response times, both ways

Left block is all 122 rows. Right block drops the 39 video-OCR rows and reports only what came off verified captures. Both are honest; they answer different questions.

Every coach reads faster on verified rows alone. Margarita's sample drops from 29 to 10 and her mean falls from 8.2 to 6.1, which is the clearest reason to finish the 7/22 captures before publishing a leaderboard. The original evaluation flagged an unstable ranking built on a sample of two. The sample sizes are healthier now; the provenance is what is still soft.

5 · Response-time distribution

All 122 timed rows, minutes from agent question to substantive coach answer.

Just over half of all questions come back inside 5 minutes. The tail is thin: 9 rows of 122 run past 20 minutes, and the slowest is 75. Volume by day: 7/20 has 1 row, 7/21 has 39, 7/22 has 58, 7/23 has 19, 7/24 has 5. The 7/22 peak is also the day carrying every unverified row and 40 of the 47 gaps, so its shape is the least settled.

Question mix

TBFA and Medical-Indemnity validation is 50 of 122 rows, 41% of everything asked. Lost Time and TAW is 16. After those two, no category clears 10. The long tail runs: Claim Handling 10, DDE/DDF Removal 9, Companion Claims 8, 301 Field 7, Technical/Access/VA 5, then eight categories with 3 or fewer. Two of those small labels are the collision pairs named above, so the true tail is slightly shorter than 16 labels suggests.

6 · The 14 original defects, retested

The 2026-07-23 evaluation found no malfeasance and 14 named defects. Each one re-run against today's canonical file.

Ten fixed, two partly, one open, one holding. The malfeasance verdict from 7/23 is unchanged: the arithmetic was correct then and is correct now. What has changed is that the numbers are now reproducible from a single command with a stated gate, and that where evidence is thin the file says so instead of averaging it away.

7 · The 47 open rows

These are interactions that exist and are missing a clock reading. They are named individually in QuestionAudit_CAPTURE_WORKLIST_2026-07-27.xlsx with the Teams chat to open for each. None of them overlap the claims already covered by the 7/23 and 7/24 captures, so they need their own screenshots.

By date: 40 fall on 7/22, 5 on 7/21, 2 on 7/23. By gap: 30 are missing both times, 15 are missing only the asked time, 2 are missing only the answer. The 15 missing only an asked time are the cheapest to close, since the answer is already recorded and the question sits directly above it in the thread.

{"date":"2026-07-22","agent":"John Zapata Rodriguez","coach":"Tami Bales",
 "claim":"ARRAIAL, ANTONIO","claim_id":"2805522671","asked":"10:54","answered":"11:13",
 "question":"","answer":"","type":"Lost Time / TAW",
 "source":"Teams screenshot 7-22 (verified)"}

8 · Files, commands, provenance

Rebuild the canonical file

python3 ~/.claude/skills/question-response-audit/build_audit.py \
  --head ~/Projects/_outputs/audit/QuestionAudit_COMPLETE_2026-07-23.xlsx \
  --append-jsonl ~/Projects/_outputs/audit/teams_reads.jsonl \
  --out ~/Downloads/QuestionAudit_PowerBI_TIMED_ONLY_2026-07-23.xlsx

What would move this forward

• Capture 7/22 first. 40 of 47 gaps and all 39 unverified rows fall on that day. Closing it settles both problems at once.
• Re-read the 39 video-OCR rows off stills rather than trusting the 720p frames, or mark them excluded from SLA reporting in the model.
• Decide the two category collisions and pick a surviving label for each.
• Hold per-coach comparison until provenance is even. Section 4 shows how much the ranking moves when the unverified rows come out.