Each card below states the act being asked for first; the reasoning behind it stays visible on the page, not tucked under a fold. Five more items are already proceeding without a tap — say stop on any one of them and it stops. One sign-off at the end closes the page.
Every presenting agent was to get its own created voice. Daisy's was authored 30 Aug; Kai's was queued alongside it and never written. Kai currently runs on three plain rules (plain and numerate · claims labelled as claims · "the data doesn't show that" said unsoftened). Nothing is blocked by this — it affects how Kai reads, not what Kai does.
Proposal at 3. Agents/HI_Kai_LeadIntel_Skill_Proposal_2026-09-05.md, evidence behind it in 7. Research/KAI_Brief_2026-09-05_claude-lead-scraping-reality.md. Scores and enriches leads from sanctioned sources only (Apollo's official connector, Clay signals, your own exports) with a receipt on every enriched fact. No scraping, no social-layer automation, no autonomous collection — the pattern LinkedIn permanently banned a vendor for in March. Nothing is built; approval starts the repo layer.
The first brief found the evidence thinnest exactly where hi!'s own floor sits — the two rigorous surveys cover $250M+ and $25M+, and the only source speaking to $10–100M was vendor content that could not be verified. This is the gap under the qualification floor the strategy already published.
The strategy doc carries figures Kai has never verified — the Gartner AI-readiness split, "60% of marketing leaders trimming agency spend," referral-close-time and win-rate numbers. Three widely-circulated statistics were already rejected as unsourceable during other passes. These are client-facing if the strategy is ever quoted outward.
The approvals store wrote page, row, who, verdict, note and time — never the question, so the record could say a key was tapped but never what was asked. That was the root cause behind the divergent-wording problem.
This shipped while the row was sitting here waiting for you. Commit cc70729, “The store records the question, not just the tap” — the API now returns a question field captured at tap time from the row’s own wording, and a null reads as “not recorded” rather than “no question,” so rows written before tonight stay honest. The approval row has been removed rather than left asking you to authorise something already done. Nobody had tapped it, so no verdict was disturbed.
The structural half of the same fix. Mounted once, wording cannot diverge at all. Cairo already used this pattern for eight duplicate mounts on "Waiting on you" (ff78cf4). Four rows currently carry two wordings each.
The rule that decides which wording is authoritative is already a held row on tonight's page — /review/ tonight-2026-09-07, row row-wording. Kai's determination is in 3. Agents/HI_Kai_RowWording_Determination_2026-09-07.md; the tap is there, not here. Remounting it here would be the exact fault it rules on.