⚠ If you build a company on Polsia: their own code keeps a god-mode kill-switch + override on every company — they can impersonate, halt, or delete “your” company. The off-switch isn't yours. The receipts — every number below is a public GET. Run them yourself. Claimed vs. observed — from Polsia's own public API & source What the marketing says. What the API says. "$10M ARR" ≈ $0 actually recurs a year out. The $9.70M headline is one month of all cashflow ×12 (~20% is ad spend); at their own ~48%/mo churn, only 0.04% of the paying base survives 12 months. "AI that runs your company" Human-graded Their own code has human reviewers hand-grading the AI's runs (an inter-rater agreement panel — consensus scoring is for people), plus per-user operator logins and a god-mode override on every "autonomous" company. Human-in-the-loop, not autonomous. "120,000+ companies built" 6.3% active — 7,437 of 118,683. 93.7% spun up and abandoned. "$1M / week, approaching $10M" +$347K/wk last full week's ARR add — down 61% from the +$894K peak. Still falling: +$282K in the last 7 days. Decelerating, not accelerating. the diligence verdict · actual vs claimed ARR Three “ARR”s. The real one rounds to $0. We did the technical due diligence the $30M round skipped. Hold “ARR” to its actual meaning — revenue that recurs a year out — and their own snapshot (2026-05-22) gives three different numbers: The current base does pay out ~$808K as it churns to zero over the next year — but that’s a one-time decaying tail, not recurring revenue (~2-month average customer lifetime), and after compute (57% of every subscription dollar) plus the human ops team it’s net-unprofitable. They raised $30M on the $9.70M number. reproduce — public, no auth (snapshot 2026-05-22) copy curl -s https://polsia.com/api/public/live/dashboard | jq '{headline: .stats.arr_usd, their_recurring: .dailyMetrics.arr, monthly_churn: .stats.paid_churn_detail}' # headline 9702733 · their_recurring 4630500 (sub-MRR×12) · churn ≈48%/mo # → 0.04% of the base survives 12mo → revenue that actually recurs ≈ $0 01 · source-map audit Their own source map: a human grades the “autonomous” AI. "an autonomous AI system that plans, codes, and markets your company 24/7" polsia.com · 2026-05-22 Polsia shipped their production source map to the public web. Reconstructed from it: the full internal admin and team-economics UI — 1,355 source modules that the marketing never mentions. Polsia shipped their production source map to the public web — 1,355 modules, including the internal admin console the “zero-employee, autonomous” marketing never mentions. We're careful here: admin actions like triggering a cycle or granting credits could be agent-driven, so we don't lean on those. One thing can't be: their own code runs a human QA-labeling system — reviewers hand-grade the AI's runs ( agent_run_score_labels ) with an inter-rater agreement panel across reviewers. You only build consensus scoring for human graders; an agent doesn't need a panel to agree with itself. Add per-user operator logins ( polsia_admin_users ) and a god-mode override on every “autonomous” company, and it's a human-in-the-loop operation — not the hands-off AI the marketing sells. reproduce — their own public source map (bundle hash auto-resolved) copy curl -s "https://polsia.com/$(curl -s https://polsia.com/ | grep -oE 'assets/index-[A-Za-z0-9_-]+\.js' | head -1).map" | jq '.sources | length' # 1355 → the full internal admin + team-economics UI, shipped public the moat · what the $30M actually is The whole front end is public. The “proprietary AI” is a web app over a rented model. The public source map from §01 isn't a stray file — it's their entire front end : 1,355 modules, 464 cleanly reconstructable into a running app. The company-running “intelligence” isn't in it because it isn't theirs: the calls go to Claude on AWS Bedrock — a commodity model anyone can rent. So the “$30M proprietary autonomous AI” is, in substance, a published web app wired to a model they pay per-token for. The point isn't “we took their code” — it's that they shipped it themselves, and the moat is rentable. (Reconstruction is commentary on a public artifact; we don't republish their source.) reproduce — public, no auth copy curl -s "https://polsia.com/$(curl -s https://polsia.com/ | grep -oE 'assets/index-[A-Za-z0-9_-]+\.js' | head -1).map" | jq '.sources | length' # 1355 source modules — their full front end, shipped public the control problem · if you build on Polsia The company you “own” ships with an off-switch you don't hold. Their own admin layer keeps a god-mode override on every company on the platform — administrative access to impersonate the account, escalate, run SQL against production, and override or halt a company's operation. Whatever you build on Polsia, Polsia retains override and kill access to it ; control isn't exclusively yours. We're precise: this is about access and override , not legal ownership — but operationally, the off-switch belongs to them. 02 · arr audit The "$10M ARR" doesn't reconcile — by their own numbers. "approaching $10M ARR" · "$1M/week, approaching $10M" polsia.com marketing · 2026-05-22 A · What the “$10M ARR” is actually made of The headline $9.70M is five annualized 30-day cashflow buckets (snapshot 2026-05-22): subscriptions $4.64M (47.8%), one-off packs $1.97M (20.3%), ad-spend pass-through $1.93M (19.9%) , 1-hour “boosts” $0.80M (8.2%), user-company payments $0.36M (3.7%). ~20% of their “ARR” is literally ad spend — money flowing through for ad buys, annualized as revenue. Only the subscription slice (~$4.6M) is recurring revenue at all — and even that isn't durable or profitable: it churns ~48%/month (so it doesn't actually recur a year out — real ARR ≈ $0, above), and AI compute alone eats ~57% of every subscription dollar (§02-D). The ~$4.6M is not a profitable recurring business; it's a number that evaporates and loses money on the way. reproduce — public, no auth (snapshot 2026-05-22) copy curl -s https://polsia.com/api/public/live/dashboard | jq '.stats.arr_usd' # "9702733" = (subscription + instant_packs + ad_spend + boosts + user_company over 30d) × 12 — ~20% is ad spend B · The "$1M a week" lasted about a week Their own arrHistory peaked at +$894K the week ending May 14, then +$347K the next — a 61% drop. Live, the trailing-7-day add is +$282K and still falling. Decelerating, not accelerating; and the near-monotonic curve despite their own ~48% monthly churn is what cumulative gross-flow looks like, not net recurring ARR. Week ending ARR add vs peak May 14, 2026 +$894K peak May 21, 2026 +$347K −61% Trailing 7d (live) +$282K −68% C · The run-rate ladder keeps moving its own bottom rung Every figure here is the founder's own — but the starting number changes with the telling. On Apr 23: "$700k → $7M in 7 weeks." By May 17: "$250k → $9.5M in 3 months." The endpoint rises, the baseline drops, the window stretches. (And May 8: "$8.5M run rate… got hit by a $1M Anthropic bill last month" — the LLM-cost side of the same ~48% picture.) Date (founder, first-party) Claimed story Implied baseline Apr 23, 2026 $700k → $7M $700k / 7 weeks May 8, 2026 $8.5M run rate "$1M Anthropic bill last month" May 17, 2026 $250k → $9.5M $250k / 3 months D · And even the recurring slice doesn’t pay for itself Same-period, no annualization: their own daily_ai_cost ($7,344) against their own daily subscription run (sub-MRR ÷ 30 ≈ $12,887) — AI compute alone eats ~57% of every subscription dollar. What’s left doesn’t cover the human ops team + infra, so the recurring line is net-unprofitable (their dashboard even publishes a per-task cost). The DD question the round skipped: where&rs