About — Zetamax · v1.2 · updated 2026-05-25

Mental-math drill, kept lean.

Two minutes. Open the page, drill, see how you did. Zetamac-compatible by default — same operand ranges, same auto-submit, same instant feel. Practice and Competitive share one engine; only what happens after the round differs.

Round length2:00120s, ± 2s tolerance

Sign-inoptionalPractice = none · Competitive = Google

Storageyour devicelocalStorage¹

Costfreeno paid tier

Practice

Local-only mode. The drill, with nothing after.

No sign-in, no backend. Your runs save to this device's localStorage. Stats are computed locally and live on the same device.

Competitive

Same engine, three flavours of post-round wrap.

Friend leaderboards on top of the same drill engine. Three flavours:

Rankedelo · margin-aware: every round counts. Server-issued seed, server-validated score, ELO updates per round with margin-aware math (a 47-30 win is bigger than 47-46).
Dailyall-time mean: one shared puzzle per day, one shot, leaderboard ranks the all-time mean. Reload mid-round forfeits that day; you can backfill past days you missed.
Leaguesshare url: share a URL, friends join, rank against just them. Open by default — no approval queue, anyone with the link is in.

Stats

Per-op accuracy, latency, trend, heatmap, focus.

Per-operation accuracy and latency, score trend across the last thirty rounds, a multiplication-fact heatmap, and a weak-pattern diagnostic that names the one thing you should drill next. Export to JSON or wipe at any time — it's your data.

How Learn works

Skill tags + pattern tags · empirical-bayes diagnostic · single Today's focus output.

Every problem you drill gets a skill tag (multi-carry addition, large multiplication) and zero or more pattern tags (by-9, near-square, doubles). Patterns trump skills — if a problem fits a known mental-math pattern, the pattern wins the attribution.²

After ~30 problems, the system starts looking for the tag that's consistently slow for you. Three guardrails keep it honest:

Log-transformed latencies

One eight-second blank can't dominate a tag's average — outliers stop being outliers in log space.

Empirical-Bayes shrinkage

Tags with few attempts get pulled toward your overall mean; tags with many attempts converge toward their raw average. No cliff at “I just hit 10 samples.”

Sample-size floors

30 problems total before the diagnostic wakes up; 10+ attempts on a specific tag before that tag is eligible. A tag is only flagged when the system is ≥70% confident it's genuinely slow versus your own baseline.

The result is a single Today's focus tag — the one weakness most worth drilling next. No top-3 dashboards; dashboards of deficits cause people to disengage. Just the next thing to work on.³

Where your data lives

Local stays local. Competitive scores go to the server.

Practice scores, lifetime totals, the multiplication heatmap, and the weak-pattern diagnostic all live in your browser's localStorage. Nothing about how you drill leaves your device. Competitive scores are saved to a server (so friends can see them on the leaderboard), but the per-tag diagnosis stays local.

Wipe everything from /me → Stats → Reset all stats, or export a JSON copy first.

Credit

Built by Thomas Ou. Inspired by Zetamac. Built to keep the keyboard feel intact while making the score stick.

Notes

[1]Web Storage API · domain-scoped, persists across reloads, browser-imposed ~5 MB cap.
[2]Attribution is versioned. When the tagging rules improve, prior rounds can be re-tagged without losing scores.
[3]Late-round rushing is filtered — events from the last 10 seconds of a round don't count toward per-tag latency.