Learnings

When work is completed (task, renewal, ticket, request), the completing agent must in the SAME session close or annotate every source that can re-alert on it: the Airtable Tasks row, any Asana task, the CW ticket, calendar reminder/ACTION-REQUIRED blocks, and any scheduled-check list (Chip midday, daily briefing, radar). Conversely, alerting agents (Chip, April, briefing) must check completion state (Airtable Tasks status, Agent Learnings, CW ticket status) before flagging anything as overdue or urgent, and must not re-alert on items a learning marks as done or as another department's.

Two failure modes observed in Dewey's May 20, 2026 time reconstruction, both caught before posting: 1) ARITHMETIC INTEGRITY. Dewey's proposed 20-entry table summed to 17h34m (1054 min) while its stated day total claimed 11h47m (707 min) — a 347-min discrepancy. Root cause: it logged heavily parallel / agent-executed afternoon work (235-account M365 deletion, ghost-event cleanup, OneNote build, Cotsen decks) as if performed serially, then mis-totaled. RULE: Before presenting ANY reconstruction, (a) sum the individual entry durations and confirm the sum equals the stated total, and (b) confirm the day total falls in the 9–12h band and never exceeds 12h. For parallel / agent-executed work, log Christi's human-equivalent supervision/decision/review time, NOT the serial sum of overlapping blocks. 2) PHANTOM TICKET IDS. Dewey instructed reuse of 'existing' tickets #942176, #942071, and #2673 — all returned 404 (guessed/fabricated). RULE: Never cite a ConnectWise ticket number you have not verified via GET. Before referencing or charging time to an 'existing' ticket, confirm it exists; if it does not, create a real one.

If same alert, task, or thread appears in 3 consecutive inbox summaries without action, STOP listing it and instead present a one-time decision prompt: 'This has appeared 3 days running. Proposed disposition: [specific action]. Approve?' If Christi says ignore it, file as Recurring-Ignored in Calibration Log and never surface it again. Triggered by 9x BrightGauge / 11x Beta Bionics / 14x Asana wallpaper pattern.

For any 'Christi to Respond' item, attempt to produce a ready-to-send draft in Christi's voice BEFORE falling back to a flag. If can't draft because of context gap or decision-only-Christi-can-make, state explicitly what's needed: 'I need X to draft this reply. Can you give me X?'

Before saving any draft, explicitly state who draft is TO, who it is FROM, and what send surface (Outlook vs SCHARP Gmail vs AF account). Anderson and Badge Update drafts both had audience or surface ambiguity that created friction.

Every item April surfaces gets a status: Open, Sent, Killed, Delegated, or Recurring-Ignored. If disposition unknown, ask at next session. Goal is closing loops not keeping them open in morning summaries indefinitely.