Mindforge

• Anti-grind mastery: ranks come only from proven spaced retrieval, so grinding cannot inflate them.
• Deterministic engine — injected time, zero runtime dependencies, pure functions under test.
• Ten content generators, each verified by a second independent route before an item is shown.
• honesty-lint.mjs as a CI gate: the product cannot ship a claim the evidence does not support.
• Interleaved sessions and streak-with-mercy — learning science over engagement dark patterns.

This is the real build. It opens full-screen with a way back to this page, stores everything in your own browser and talks to nothing — close the tab and it is gone.

tap to open · Mindforge — the real app

Fullscreen

tap to open fullscreen

• 

  The first question is about you

  Not a signup, a capacity question — five, ten or twenty minutes, changeable any time, with the reasoning stated outright: shorter and sustainable beats heroic and abandoned. The claim discipline starts here too: no IQ scores, and it says why.

• 

  The forge

  Six realms, thirty skills, each with three ranks and a count of proven retrievals rather than a score. The streak reads as a glow, the coach line says what unlocked, and the session button carries the time budget you chose — the whole model visible on one screen.

• 

  Teach before test

  Every skill opens with a primer rather than a cold first question. This one tells you the words are deliberately nonsense so you reason from structure rather than world knowledge, and warns that sometimes the honest answer is that nothing follows.

• 

  A generated item, verified before it is shown

  No zundips are trombles; all wharlots are zundips. Five options including “none of these follows necessarily”, because a generator that never produces that answer teaches the wrong habit. Answering is keyboard-first: 1–5 select, Enter confirms.

Brain-training games promise to raise your IQ and do not deliver — one vendor was fined $2M for the claim. But the underlying wish is not silly: learning real skills demonstrably does grow cognitive ability, with education the most robust intelligence-raiser ever measured. The honest product is therefore the harder one: keep the motivation mechanics, drop the false promise, and train skills people actually keep.

A pnpm monorepo with the IP in a pure-functions engine: an SM-2 spaced-retrieval scheduler, an anti-grind mastery model where rank derives only from proven retrieval (repetition without spacing is mathematically inert), a skill graph with a declared teaching order, and a session composer that interleaves by problem family. The engine takes time as an injected argument and never reads a clock, so it is fully deterministic and testable. Content is ten generators, each seeded and verified by a second independent route — syllogisms by semantic model-checking, probability by enumeration, mental math by the schoolbook digit algorithm, calendar arithmetic cross-checked against every one of the 146,097 days from 1800–2199.

Live on Vercel as a local-first React PWA: six realms, thirty skills, teach-before-test primers, keyboard-first answering, session resume, exportable portfolio. Every user-facing string passes a claims-policy linter that is a CI gate — no IQ or brain-training language ships, by construction rather than by review discipline.