Three artefacts, three variants each. Every number and every answer is generated from the live Compass export of 2026-09-15 and the scenario transcripts; deal names in C are the only invented strings. Type notes under each variant (saved in this browser), tick a pick per artefact, then copy everything from the bottom. Thumbnails scroll; open full size to read.
Mobile-width Slack mock. Three messages per variant: the qualify-a-lead result (runs/scenarios/qualify_all_in_one.json, turn 1), the onboarding-pack answer (docs_by_currency.json, turn 2, Equals' sole-trader pack) and the Monday change-only post assembled from dated facts in the live export. The words are identical across variants; only the shape changes.
A1 · Prose, as today
The persona's fixed shape verbatim. Cheapest to ship: it is what the bot writes now. Reads fine on desktop; the provider lines wrap hard on a phone and the offer is easy to miss.
Same words, structured: header, a LIVE/ROLODEX pill per provider, caveats in grey, the footer as a context block and the offer as buttons. Buttons keep the offer rule honest: a button can only exist for something a tool computes. Costs a Block Kit renderer in the Slack worker.
Only the answer line and counts land in the channel; provider lines, footer and offer go to the thread. Best for the unprompted Monday post, where presence is the news. Worst for a DM, where the person wants the whole answer in one glance.
Analytics › Panel health. Gaps on top, re-verify queue below, and the new column: for each gap, how many live partners record a yes, a recorded no, or were never asked. Compass's own gap analysis cannot draw that line because its matcher reads an empty field as a pass. Numbers are live: six partners past 90 days (Moneycorp 199 days, four at 182, Verto never), zero contacts, zero sourced rows.
B1 · Two tables
Compass-native and the least new UI: a Gaps table with the three counts in one column, then the queue with the call questions inline. Scans fastest for a weekly review. The per-partner detail is hidden behind 'Ask the panel'.
Each gap card shows all six partners as Yes / No / Never asked at a glance; the queue draws staleness as a bar against the 90-day line. More visual, more scrolling; the strip only works while the panel stays at six.
Gaps and re-verify merged. A column read top to bottom is the question list for that partner's verification call, so the screen is also the verify-stale-provider call pack. Densest; needs a wide screen; most useful when the data is this thin.
Analytics › Win / loss vs capability. Each lost deal beside what Compass records for that corridor and client country, sorted into three piles. Deal ids, client descriptions and Pipedrive reasons are illustrative; the Compass record lines and the never-activated counts (analysis/cross_summary.json) are real. Pile sizes from the leads book: truly cannot 1,216 rows, could do but nobody recorded 604, not about capability 7,512.
C1 · Three piles as columns
Kanban. Reads as a sorting exercise, which is the job: which pile does each loss belong in. The middle column is visibly the one to work. Weak once a pile has more than ten deals.
Flat, filterable, exportable: Pipedrive reason and Compass record side by side on every row, pile as a status badge. Closest to how Mark already looks at Pipedrive. Least persuasive about where the money is.
Three tiles and a bar size the piles against the never-activated book before any deal is shown; the middle pile is open, the others fold away. Strongest for a board or leadership read. Hides the individual rows that a rep would act on.