What shipped, and why

The pricing page now runs the same near-black violet field and animated gradient layer as the landing page, with mint as its only call-to-action colour. It used to sit on a standalone green background with gold buttons, which read as a different product.

Scroll-spy alone fought your click near the bottom of the page, where short sections can never reach the top of the viewport. A click now wins until you scroll again.

Results stamp in, lists rise on a shared easing curve, and page transitions are consistent app-wide. Everything respects prefers-reduced-motion: the animation is skipped, never the content.

Display name, password, and avatar sit in their own cards alongside preferences that stay on your device.

Every solo drill is archived, and questions you got wrong can be re-served as a targeted set. Season resets are marked on the rating graph so a soft reset never reads as a collapse.

New accounts choose their target exam before landing in the lobby.

Accepting an invite link now carries you through sign-in, the host is notified without waiting on a realtime channel that never carried challenges, and a challenge code can be joined directly.

A retry after a lost connection re-derives your true position instead of resubmitting a stale question index, so a flaky network can no longer strand you mid-match.

Playing the bot no longer requires an existing queue row.

The in-app admin console and its API routes are gone. Question-bank work moved to the ingest scripts, shrinking the authenticated attack surface.

Matches were lockstep: both players waited on the slower one at every question. Now each player traverses their own nine questions at their own pace, and the match is decided once both have finished. Bot matches deliberately stay lockstep.

Data-interpretation sets are served whole — the caselet and its sub-questions together — with the reading window added to the first question's clock instead of being charged against your speed bonus.

A rolling profile of your accuracy by section, question type, and difficulty band feeds the coach with no extra model call per match.

If one section is thin, its slots are redistributed rather than producing a truncated match that could run past its own question list.

Three modes on one route: stat-grounded Q&A, a Socratic buddy that gives hints rather than solutions, and a seven-day plan. Plus a PDF solver that reads a question paper you upload.

TITA questions ship across matches and solo drills. The answer box accepts numbers only — the same constraint the real exam interface applies — because a lenient matcher would turn a correct answer with a unit attached into a wrong one.

Weakest-section-first question selection, unrated, with the worked explanation revealed the moment you lock an answer. Practice never moves your rating and never touches question difficulty.

Every Ninja route refuses to answer while the caller has a live match — including questions aimed at an older, finished match from a second tab. The rule keys on you, not on the match you name.

Questions carry embeddings, refreshed automatically whenever a question body changes. The search primitive returns identifiers and similarity only, never question text.

Answer rows are readable only by their own author. Previously a typed answer could be read by the other player while the question was still open — with type-in answers stored as plain text, that handed over the answer.

Ninja Bot no longer appears on the global ladder or in season results.

Unrated player-versus-player matches no longer nudge a question's difficulty rating — uncapped private challenges were a way for two accounts to distort the bank together.