Marketing-Ops Platform: desktop → Telegram → web

Summer 2026: spend and reports. Meta spend is written to the tracker in each ad account's own timezone: one profile had accounts on two timezones, the day window drifted by ~10 hours, and part of a day's spend landed on the neighbouring day. A unique-clicks flag on the write removed a 10–15% overstatement. Writes are full-day totals and repeats are safe: 96 runs a day, and yesterday is recomputed after midnight while the network delivers its late hours. Sign-up → first payment reports per network campaign: spend and installs from the second network, registrations, payments and revenue from the tracker, joined on a six-character campaign ID carried in the campaign name; cost per install, funnel-step rates, ROI and the break-even point. The mature rate (registrations older than 6 hours) showed the same offer converting differently across apps, so the app is examined before the offer. Payments arrive over 24–48 hours, and the first-day rate is flagged in the report as understated.

Click-to-ad attribution. An fbclid prefix filter in the event collector dropped 492 of 672 clicks a day: a sale reached the Conversions API without its click ID, and of 4 sales a day the ad account saw 2. Checking the prefix family lifted the share to 672/672; then the prefix list was removed altogether, since such a list goes stale at the next format change. 48 sales from a 7-day window were resent with the same deterministic event ID, 48 of 48 accepted. Several campaigns run one post with one link, and the tracker merged their clicks into one bucket; the adid field inside fbclid matched the real ad ID in 0 of 137 checked clicks, so the mapping table is our own, seeded from clicks where Meta appends the IDs itself: 262 pairs, 22 ads, 0 collisions, independent check 361 of 361. An ad → ad set → campaign directory across 10 ad accounts (912 ads, 760 ad sets, 244 campaigns) with 0 integrity violations; the collector is idempotent by a high-water mark — before that, every run doubled its counters. The analytics bot doubled tracker figures on identically named campaigns and summed sales from a single stream — 10 against 13 in the tracker, which produced a false stop signal; all streams are now summed and the report matches the tracker to the cent. 34 tests on real samples, fully offline.

Rules engine. Five verdicts per funnel "tracker campaign × Meta campaigns", from an early stop to scaling. The engine proposes and the operator decides: the engine writes nothing to the tracker or the ad accounts. Three sources: the tracker as the single source of money and funnel data, Meta for rank, live spend and account health, the partner network for sales reconciliation as a safeguard: when the network reports more sales than the tracker, a stop is blocked. Thresholds were derived from 92 campaign-days and matched the operator's heuristics: they are the ROI sign-change points for unique-click price and landing-click price. The per-click reconstruction matched the tracker report in 1,704 of 1,704 day × campaign cells. A 90-day backtest with no look-ahead: the first version proposed stopping the flagship funnel, which went on to deliver 139% of the account's profit. A "never stop a profitable funnel" guard was added, the minimum sales for a CPA verdict rose from 3 to 8, and the worst-price-zone rule now requires at least 2 sales — it had issued 9 of 17 verdicts at zero sales. The rules' effect over 90 days is 0.8–1.8% of spend, and the report says so plainly. 49 tests, a daily Telegram report; the bot's campaign cards compute their verdict with the same engine and show the reason.

Analytics bot and vault. Meta profiles are added and switched from chat, and a token is validated and deleted from the conversation immediately; pause and resume work under any of the operator's tokens, and Meta rate limits trigger retries with 1, 2 and 5-second pauses. The web vault: scope and expiry checks, reveal only after re-entering the password, archive with restore, auto-lockdown, a log of login devices. An offer with an empty or placeholder URL cannot be created — added after 33 offers were found pointing at a placeholder.