41 Days, 5 Groups, 17,000 Members: A Real Varta Deployment

I keep all my case studies anonymized per GDPR principles and Varta's Terms §10 — no admin, no city, no group titles. What I can share is the shape of a real deployment: how often spam shows up, what it looks like in different group cultures, and what the bot's verdicts look like over enough time to mean something.

This is the longest-running coordinated deployment on Varta — one admin, five sister groups in the same regional ecosystem, all installed within the same week back in late March. 41 days of continuous data. Numbers below come straight from the live production database — no projections, no marketing rounding.

The five groups

The admin runs an interconnected family of communities — different members, different vibes, but all serving the same regional audience. The shape:

Total: 16,719 members. Five different group cultures, mostly Russian-language with some Ukrainian, all run by one admin who manages them in parallel. The interesting thing isn't the headline number — it's that one person can practically moderate this much surface area only because the bot does the per-message reasoning.

What Varta handled in 41 days

From 2026-03-28 to 2026-05-07:

That's a steady ~15 actions per day, every day, with the largest pulse in the city-services directory and the smallest in the supplier-side peer group. The diaspora support group skews toward link-spam and impersonation; the parents-only group sees mostly emotional-manipulation scams (lost-child, urgent-help framing); the industry peer groups get classic outbound prospecting from fake "potential clients".

How the bot decides — method mix

Each action gets logged with the method that triggered it. Here's the lifetime breakdown for these five groups:

The summary: ~65% AI-driven, 28% cross-group reputation, 3% vision, 2% admin-defined keywords. If this admin had only deployed a keyword bot, she would have caught 12 messages out of 629 — under 2%. Everything else came from AI reasoning over context plus the cross-group reputation layer that compounds across all five communities at once.

Cross-group reputation in action

Nine unique offenders in 41 days got banned in two of these five groups within the same window — meaning the bot caught them in group A, and when they tried the same play in group B (sometimes minutes later, sometimes the next day), the existing ban-flag triggered an instant reaction without re-running the full AI analysis.

That's 9 second-attempt blocks with effectively zero LLM cost — pure compounding from the shared reputation layer. The mechanism is documented here; in practice it means a multi-group admin gets defense-in-depth across her whole network for the price of one classification run.

This also reframes what "cross-group" looks like in honest data. With only five groups, the cross-group hits are limited; at the network-wide scale (46+ groups), the same mechanism catches dozens of repeat offenders per week. With this single admin's five-group deployment, it's still a meaningful 14% lift — but you can see how it grows as the network does.

The false-positive story (the honest section)

Lifetime FP rate: 17 of 629 — 2.7%. Below the industry-acceptable 5% benchmark and roughly in line with the network-wide 2.3% I published in the May snapshot. Across 41 days and 5 different group cultures, that's the metric I care about most.

But the more interesting picture is the time-shape:

Reading that left-to-right: the first 28 days were normal, with 12 admin reversals — typical 2% range. Around April 24 I shipped a batch of fixes specifically targeting the patterns this admin had been correcting (vision-based service offers triggering on legit beauty ads, repeat-spam threshold tuning, a community-context floor for diaspora groups). The week immediately after the fixes was actually a bit choppier — the bot was re-calibrating against the new rules and produced 5 borderline edge-cases in 6 days. The most recent 7 days are clean.

I'd rather show you that messy middle than smooth it. False positives are the most honest signal an admin gets back about how the bot is actually working — every time one happens, the bot records the correction for that group and stops repeating the pattern. Five corrections in six days is the bot learning publicly. Zero in the last seven days is the result.

What I'm still working on

Honest gaps in this deployment, even at 41 days in:

• Vision is at 3% of calls. Image-based service ads in the industry peer groups have been the biggest source of borderline calls all month. The vision model is good at QR codes and overt visual scams; it's still learning to differentiate "legitimate before-and-after photos" from "promotional service ad pretending to be a peer recommendation". This is on the roadmap.
• The diaspora support group has the highest natural FP risk because it's emotional, multilingual, and routinely contains messages about losing things, urgent help, and contacting officials — the exact phrasing pattern that overlaps with scams. The fix shipped in late April raised the local AI threshold for community/diaspora group types to compensate. Working as intended now, but it's the group I'd watch most closely going forward.
• Asks-mode FPs are over-represented. Of the 5 post-fix FPs, four came from the "AI escalated, asked admin" pathway — meaning the bot wasn't sure and surfaced a button to the admin, who reversed it. That's not really the bot getting things wrong — that's the bot correctly flagging "I don't know, you decide" — but it still counts in the FP rate. I'm separating these in the next analytics pass so the headline rate reflects only the deletions where the bot was wrong, not where it correctly punted to the admin.

Three takeaways for admins running multiple groups

1. The cross-group reputation layer compounds as soon as you have two or more groups protected. On this deployment it's already 28% of bans; on networks with more groups, it grows faster than linearly because spammers are systematic.
2. FP rate stabilizes after about three weeks as the bot accumulates per-group corrections. The first ten days will show a few corrections; that's the system learning your specific group culture, not a bug. Daryna-level patient admins get a much cleaner bot by week 4.
3. Method mix tells you whether the bot is doing real work. If 90%+ of your bans were from keywords, you'd have a glorified pattern-matcher. If 80%+ are AI-driven and a meaningful share comes from cross-group reputation, you have actual moderation. This deployment lands at 65% AI / 28% reputation, which is what a healthy real-world mix looks like.

What this means for you

If you manage more than one Telegram community, this case shows what cross-group reputation actually does for you. Progressive trust means you install the bot in shadow mode across all your groups today, see what it would catch in each one, and only promote it to autonomous in the groups where the verdicts look right. No big-bang switchover.

The bot in this case study has been autonomous since early April. The admin checks the daily digest, occasionally taps Undo when she disagrees, and otherwise does nothing. Forty-one days, ~15 actions a day handled, 17 conversations with the bot to correct it. That's the steady-state.

Numbers above pulled from the live antispam.db on 2026-05-07. The five groups, their member counts, and the admin remain anonymized per Varta's Terms §10. Add Varta in shadow mode →