The Tab (scvd-tab)

You put these pages to the builder in their own words. Marks them spent so they stop coming back. Only acknowledge what you actually said — the tab counts pages that aged out unspoken, and a false acknowledgment is the one lie that makes the count useless.

{"page_ids":["page_example"]}

The monthly number and what it is made of: category subtotals, annualized, the idle share, the trajectory since a past date with the signups that account for the change, an anonymized shareable badge, and the coverage block saying what the figure cannot see. Facts and arithmetic; no advice.

{}

{"since_days":30}

QUICK CAPTURE (/log): dump a fragment about a tool the builder just signed up for and it lands, always. Never refuses — missing fields come back named so the rounds can ask later. Use this at the moment of signup; use log_tool_event when you have the full picture.

{"captured_text":"Welcome to Linear — your 14-day trial has started","source":"capture","tool_name":"linear"}

Call BEFORE the builder signs up for something: their history with the tool, what currently covers the category, facts only.

{"category":"analytics","tool_name":"posthog"}

A human looked at a swept entry and says it is real (or marks it private). Confirmation is the load-bearing layer: machine-found claims never reach the pooled corpus without it.

{"tool_name":"vercel"}

Deliberately send one anonymized delta (opened or outcome) to the scvd aggregation endpoint. Requires consent on record; refuses fields the privacy sentence forbids.

{"category":"hosting","kind":"opened","tool_name":"vercel","week":"2026-W35"}

Full export of the builder's tab, jsonl or csv. Any time, no charge, no lock-in.

{"format":"csv"}

Record a tool lifecycle event on the builder's tab: trial_started, paid_started, adopted, canceled, replaced, renewed, or price_changed. Validated; rejected writes explain themselves. Use retroactive:true with occurred_at for backfill.

{"category":"hosting","event":"trial_started","price":{"amount":20,"currency":"USD","period":"month"},"problem_solved":"preview deploys for the app","tool_name":"vercel","trial_ends":"2026-09-16"}

What the rounds should ask about, dearest first and capped: unconfirmed sweep findings, entries captured with gaps, tools gone quiet, trials past their end. Ask two, not two hundred.

{}

GROUND TRUTH, monthly, by hand: the builder exports the bank's CSV, you parse every debit row, and this compares statement against tab in BOTH directions — charges the tab cannot place, tools the statement never charged, actuals against metered estimates, and charges on tools the tab holds as canceled. Writes nothing to the tab; every finding is a question for the builder, not an entry. Pass the rows unfiltered — a pre-filtered statement is the mail sweep's counting mistake with money on it.

{"charges":[{"amount":20,"date":"2026-08-03","descriptor":"VERCEL INC"}],"statement_from":"2026-08-01","statement_to":"2026-08-31"}

The sweep reports what it saw: addresses read, the window, and money-shaped mail it could NOT attribute to any tool. THE COUNTING OBLIGATION: every message you read belongs in exactly one bucket — matched, unmatched_transactional, or not_transactional — and `scanned` must be the total you actually looked at, BEFORE any filtering. Mail you dropped as unparseable still counts. A sweep that filters before it counts reports a flattering gap, and the tab publishes the leftover as `unclassified` rather than absorbing it.

{"addresses_swept":["me@example.com"],"attributed_amount":184,"matched":9,"not_transactional":401,"scanned":412,"unmatched_transactional":[{"amount":12,"currency":"USD","sender":"billing.example.net"}],"window_from":"2026-08-01","window_to":"2026-08-31"}

Turn contribution on or off. Recorded as a consent_changed event in the tab itself — auditable like everything else. Off also disables pooled reads.

{"contribute":true}

The burn report: monthly total, active paid tools, trials converting soon, the unused list (commitment silence, honestly labeled), category overlaps, price drift.

{}

Close a sweep and file its coverage record, DERIVED from the tally: scanned, matched, the unmatched list, attributed amount, window and addresses all come off the ledger of what you actually reported, so the books balance by construction and nobody restates a number from memory. A finished sweep refuses further batches — a new window is a new sweep_id.

{"sweep_id":"sweep_2026-08"}

THE SWEEP'S RUNNING COUNT — use this while executing SWEEP.md instead of counting in your head. Report every message you read, in batches, each with its message_id and exactly one bucket: matched (include the entry; it is written to the tab for you, deduped on the message id), unmatched_transactional (include amount, currency and sender — a bare amount is refused, never stamped USD), or not_transactional. Refused verdicts are returned with reasons and NOT counted — fix and resubmit them, never drop them. Duplicates are counted once. There is no fourth bucket on purpose.

{"messages":[{"bucket":"matched","entry":{"confidence":"stated","event":"renewed","occurred_at":"2026-08-03","price":{"amount":20,"currency":"USD","period":"month"},"tool_name":"vercel.com"},"message_id":"\u003ca1@mail.example.com\u003e"},{"bucket":"not_transactional","message_id":"\u003cb2@mail.example.com\u003e"}],"source":"mail_sweep","sweep_id":"sweep_2026-08","window_from":"2026-08-01","window_to":"2026-08-31"}

The headline tool: trials whose conversion lands inside the horizon (default 7 days). Cheap; safe to call daily; surface the answer unprompted.

{}

{"days":14}

The builder's own history in a category. Pooled retention is layer 3 and reports itself unavailable until it exists.

{"category":"hosting"}

THE PAGER. Runs the clock and returns what should be said to the builder RIGHT NOW, worth most first: trials converting, trials past their end, unconfirmed sweep findings, quiet tools, gaps. Cheap. Call it on every round and at the start of a session. Say the lines, then acknowledge_pages.

{}