How long until your AI reviewer trips alert fatigue?

It is a calculator that estimates how much false-positive comment load your team absorbs from an AI code review tool, then projects how many weeks of that load it takes to reach alert-fatigue onset. You pick a tool, set your PR cadence, your team size, and your reviewer tolerance — the calculator returns a signal-to-noise ratio plus the projected weeks until reviewers stop trusting the tool's comments. The output is rendered as a Windows 98 error dialog whose title bar tints by severity.

Each tool's base noise rate is an illustrative default — drawn from a mix of vendor-published precision claims (Sourcery and Korbit publish low-noise figures; Snyk reports high precision on its security findings), aggregated developer reports on HackerNews and dev.to, and conservative defaults for newer tools without public data. Treat the per-tool numbers as starting points, not ground truth — this page is not a peer-reviewed dataset. To get your real noise rate: read 50 random PR comments from the tool, count how many you would dismiss as not-actionable, and divide. That ratio replaces the default.

The 50-false-positive threshold is a heuristic from the broader alerting literature (security alerting, monitoring, code review) — once a reviewer has dismissed roughly 50 false positives from a system, mistrust generalizes to all comments from that system. The 50 number is a working heuristic, not a peer-reviewed constant; calibrate it against your own team. Weekly false-positive load per reviewer is computed as (pr_per_week × 8 comments_per_pr × effective_noise_rate) ÷ team_size, where effective_noise_rate is the tool's base rate adjusted by your selected tolerance (Strict applies the rate as-is; Permissive treats more borderline comments as signal). Weeks until onset is 50 ÷ weekly_FP_per_reviewer.

All four inputs pack into 20 bits = 5 hex characters: ?nc=<5hex>. 4 bits encode the tool (0–15), 7 bits the PR cadence (1–100/wk, clamped), 7 bits the team size (1–100, clamped), and 2 bits the tolerance (0–3). URLs round-trip exactly; pasting a teammate's URL pre-fills the inputs and re-renders the dialog. Wrong length, non-hex, or out-of-range values are rejected — tampered URLs do not silently mutate the form to a blank state.

A 5,000-sat one-time payment unlocks the AI Review Noise Reduction Playbook — 20 concrete items grouped into Configuration, Triage, Workflow, and Measurement. Each item is a config line, a command, or a process change with a P0/P1/P2 priority chip. The free calculator gives you the number; the playbook gives you the moves. The same 5,000-sat payment also unlocks the other priced features on Promptshelf (Promptle Hard Mode, Token Tetris Daily Challenge, Stumble Deep Cut, Vibe Code Security remediation guide, Comprehension Recovery Checklist, Token Budget Optimization Playbook) — the shared client-side verifier accepts any prior 5,000-sat tx to the address, so paying once gets you all seven. Verified client-side via mempool.space + blockstream.info; no account.

An AI code reviewer that drowns reviewers in nits is, in spirit, a system error — a process that produces noise instead of value. The Win98 error dialog is the visual canonical form of that experience. The title bar tints by severity (navy = healthy, mustard = warning, red = error, dark red = critical), the OK button refocuses the input form so you can adjust the cadence, team size, or tolerance and re-render, and the icon flashes a Win98 ! triangle. The dialog also screenshots well — easy to drop into a Slack channel or a retro deck.