[prodelt]

Why

A jury-style audit (COMPETITION_RULES §4.4.4, §6.1, rubric B1, E1) found:

1. Burst input counted as a test. One beforeinput (insertText) with the whole 59-character line gave «+10 XP, Вправу опановано» at 9 949 SPM; an automation tool's type_text (73 chars in 1.0 s, 4 264 SPM) counted too and became the personal best.
2. SPM counted wrong keystrokes. With k = 7 errors on 59 characters SPM read about 12 % above 60·N/T, outside the jury's ±10 %.

What

• SPM = 60 000 × min(correctCharKeystrokes, textLength) ÷ elapsedMs, i.e. 60·N/T on a finished attempt. A wrong key costs accuracy and time, never adds speed. Accuracy is unchanged (correct ÷ all, corrections included).
• Seam: an insertText carrying more than one character is ignored (reason burst), like a paste. IME compositions are still judged character by character.
• Plausibility verdict (packages/metrics/src/plausibility.ts, pure, unit-tested): metrics.implausible = 'burst' when the median interval between character keystrokes is < 25 ms, 'tooFast' when SPM > 1 500. Intervals run char to char, so a held Shift or Backspace repeat does not shorten them.
• Counts toward nothing: curriculum's typedByHand is the one predicate. The progress fold (history, Confidence, Mastery Rule, which it neither advances nor resets), Academy progress, XP/level, streak and daily minutes, personal best (result comparison, profile), speed trend and session sizing skip such attempts. The result screen says «Не зараховано: набір не схожий на ручний» with the reason. finish-race maps the same verdict to too_fast / implausible_interval (takes effect on the next functions deploy). DERIVED_VERSION is 2.
• Docs: Formulas page (uk + en: new SPM formula, new block «Набір руками»), ADR 0003 amendment dated 2026-10-06, CONTEXT.md (SPM, «Typed by Hand»), docs/security.md.
• Tests: the e2e harness types at 45 ms a key (about 1 300 SPM), and the CDP harness waits the same by default. No test-only escape exists in the product.

Thresholds and why

• 1 500 SPM ceiling: Klavogonki treats > 1 500 зн/хв as certain cheating and its top rank starts at 800 (docs/research/02-typing-trainers.md); Barbara Blackburn's record peak is about 1 060; finish-race already used 1 500.
• 25 ms median interval: a sustained 2 400 SPM. A record typist's median sits near 50 ms, while bursts, compositions and scripted event series land at 0–15 ms (the audit's tool: 14 ms). A median, so a human's occasional near-zero rollover or an IME word committed whole does not tip it.

Checks

• pnpm check: typecheck and lint are green; unit tests 835 passed. The one failure is the existing tools/architecture.test.ts (CoachMarks.tsx), which another PR fixes.
• pnpm build: green, initial JS within the 150 KB budget (index 123.4 KB gzip).
• e2e chromium, CI=1: formulas, jury, security -g "paste|counts for nothing" (new: an instant per-character series and a whole-line composition complete but are not counted): 12/12 passed. Harness consumers path-progress, stage2, academy, exercise, result, stage1-moves, review, session, groups, csp, settings, shell, plus real-input on cdp: all pass except failures that are already on main. Those are result.spec scenarios 4/5/6 and stage1-moves "weakest Transition" (both expect older coach copy) and the settings Profile axe audit (contrast of unearned badges).

Not here

• submit-attempt keeps its own older, stricter checks (a quarter of intervals < 40 ms or > 1 000 SPM rejects the upload). It stores the new verdict with the recomputed metrics.
• features/daily, features/sprint, features/home and features/owntext are owned by other PRs and were not touched. The sprint's own best and the daily challenge's "cleared" days do not read the verdict yet.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
typing-race	Ready	Preview	Oct 6, 2026 5:33am UTC