What the 198 photos from 2026-07-28 actually held

Every time on this page was read off a capture. Nothing was inferred, defaulted or carried over from a prior file. Where a photo showed a question but never showed the coach's reply clock, the row is held rather than filled.

How a row earns its times

The threads run agent posts a claim → coach acks (“reviewing”) → coach answers. The ack is not the answer, so Time Answered is the first substantive reply and the ack only sets First Reply. Captures are a sequence: a question at the bottom of one frame has its answer in the next one, so the frames are merged into a single deduped timeline per chat before any pairing happens. Pairing inside one frame at a time cost 12 complete rows on this batch alone.

Validation against rows that were already verified

Two interactions in today's photos already exist in the 122-row workbook with independently verified times. The extraction reproduced both exactly: 09:01→09:10 (Raushanah Muro) and 08:59→09:04 (Toveta Jenkins).

Four things this pass found

1. A sixth coach was missing from the extraction roster

Jonathan Mosquera posts Reviewing, Calling and Per Coach JM, update Medical/Indemnity Payments Validated to Yes across 39 messages in this batch. The skill file names him (JM) as real, but the roster used for parsing held only five names, so his exchanges were read as agent-to-agent and dropped. Adding him took complete interactions from 128 to 140.

2. The append path double-counted rows already in the head, now fixed in the engine

build_audit.py deduped appended rows on (claim_id, asked), and it failed two ways. The seen set was built with claimid(), which extracts only a 9-to-12 digit run, while the incoming key used the raw value, so WC648G07903 never matched its head twin. And every head row with a blank Claim ID collapsed to (None, asked), so an appended row carrying a real id could not collide with the very row it duplicated. Six duplicates reached the first build.

The engine now keys on (date, coach, agent, asked). Agent is load bearing: on this head, (date, coach, asked) alone flags 4 collisions of which only 1 is real, because one coach is regularly asked by two agents in the same minute. A second pass collapses duplicates across the whole kept set, which the append path structurally cannot reach, and every collapsed row is listed on Build Notes. Backup at build_audit.py.bak-dedupe-2026-07-28.

2b. 135 messages were being thrown away before pairing

Teams draws a day divider only when the day changes, so a frame can hold real messages with no date anywhere on screen. Those were dropped: 135 of 967 messages, 33 of them substantive coach replies, across 21 frames that contributed nothing at all. Frames of one chat are a contiguous scroll, so an undated frame sits inside a known span. A date is now adopted when the nearest dated frames on both sides of the same chat agree, or when only one side exists, and a disagreement stays unresolved. That recovered 21 more complete interactions, 140 to 161. The 8 rows resting on an inferred date say so in Evidence Status.

3. One duplicate already sat in the head file, now removed

Toveta Jenkins, 2026-07-21, 11:05→11:08, Reshmi Saravanan, appeared twice under two wordings of the same question (Antoin Woodard, Medical/Indemnity to Yes). Both rows were marked verified and neither came from today's photos. On Brandon's instruction the thinner of the two was dropped and the fuller wording kept. The shipped workbook now holds 236 rows with zero duplicate (date, coach, asked, answered) groups, 1 removed across the whole pass.

4. Non-questions were being counted as questions

Teams pins an un-timestamped banner near the compose box, <Name> is out of office and may not respond. It sits in most frames and was being glued onto whatever message body came last, so it contaminated 41 of 161 questions and 51 answers, and 5 rows had a “question” that was nothing but the banner. The old filter allowed three characters before is out of office, so a real first-and-last name never matched it.

Worse, any long agent line counted as a question. It is now a request only when it carries an interrogative or an actual ask, so DDE please <claim> and Please delete DDE count while thanks Tami and Perfect, thank you Raushannah do not. Chrome like Press Ctrl+F to find in this chat, Type a message and left the chat is stripped from every body. The workbook now contains zero rows carrying the banner.

4b. Half the claim names were truncated

The Origami pattern ended in \s*\(?(\d10)?\)?, every part optional, sitting after a LAZY run. Lazy plus all-optional means the match stops at the minimum, so it captured PORTILLO, MA instead of PORTILLO, MANUEL and Reich, Jo instead of Reich, Joseph. 60 of 122 rows carried a truncated claim name while the full name sat in the question text a few characters away. Anchoring the lazy run against a required closing (id) forces it to reach the real end of the name, and truncation is now zero.

The same clipping split one agent into two roster entries, Juan Jose Rio and Juan Jose Rios. A name that is a strict prefix of a longer known name now collapses onto it, so one person is never counted as two.

5. A thank-you is not a question, but it does prove an answer

When an agent replies Perfect or thanks, a coach answered even if that answer bubble was never captured, and the answering coach is the coach in that chat. Those 16 rows take the acknowledgement time as the answer time, which is an upper bound: the thank-you lands at or after the real reply, so it can only make a coach look slower, never faster. They say so in Evidence Status.

6. Today’s photos do not cover the 47 open rows

See below.

Two tiers of trust

The 121 rows carried over from the spreadsheet are the ChatGPT extractions, and per the skill file they are trusted and not re-derived. The 115 rows read from today's photos are the unproven tier and are the ones to audit: every one names its source frames in Evidence Status, 8 rest on a date inferred from adjacent frames, and 16 take their answer time from an agent acknowledgement.

The 47 open worklist rows

0 of the 47 were recovered from today's photos. Of the 30 worklist rows carrying a claim ID, 0 of those IDs appear anywhere in the 198 captures, so today's batch is largely a different set of conversations rather than the 7/21-7/22 scroll-back the worklist needs.

Complete interactions read from today’s photos

123 rows carry all four required fields (date, coach, asked, answered) and pass the gate in build_audit.py.

Held: seen but not fully timed

These appear in the photos but a required field was never on screen. They are not in the workbook and no value was guessed for them.

Spread

Provenance ceiling

Read this before quoting any SLA number. These rows come from phone photos of a laptop screen, which OCR cleanly, so they sit with the verified tier rather than the 720p video tier. Coverage is uneven by coach and by day because it reflects which threads were photographed, not how much each coach handled. A per-coach ranking built on this file compares capture coverage as much as coaching.

Built 2026-07-28 · source 198 HEIC captures in ~/Downloads · macOS Vision OCR · engine build_audit.py · skill question-response-audit