Beta status: what works today
- 19 training modes, all free, with confidence-and-accuracy scoring after every answer
- 1052 authored challenges, validated before release; answers are graded on the server
- An 8-axis Cognitive Fingerprint built from your own sessions
- 6 games (ASCENT, Evidence Showdown, Rule Shift, Signal Station, Memory Heist, Fog Cartographer) with seeded, replayable runs
- AI feedback on free-text reasoning, with deterministic scoring first and AI explanation second
- Data export from your settings, and six readable themes
Available Start a session or try a quick demo without an account. There is no launch date yet.
What is not proven yet
- Whether practice in MindFrame changes the decisions you make outside it. No outcome study with real users has been completed.
- Whether MindFrame produces the effect sizes reported in the research it builds on.
- How MindFrame users compare with people who do not use it, and how long gains last after people stop.
- How engaging the games are over weeks. Our tests so far are automated or internal.
- Live user counts stay hidden until launch so test accounts cannot inflate them. Leaderboards are open, but with a small beta cohort a rank says little yet.
In progress The research we build on is on evidence & limits, the numbers we publish are in the transparency section below, and the formulas are in how we score.
Transparency: MindFrame by the numbers
What we can honestly measure about MindFrame today, and what we deliberately cannot measure yet.
Last computed: 2026-11-18 · rebuilt from committed evidence · freshness target 7 days
What we can measure
Each figure below is produced by a check that runs in our build, not maintained by hand.
Unique training challenges
as of 2026-11-181,052 unique challenges
Every challenge is verified unique by prompt and id, with exact-match answer keys and per-mode minimum floors enforced at seed time. The number is produced by a gate that fails the build rather than by a count we maintain by hand.
Computed automatically from the product's own records; source details are in stats.json.
Modes with a written challenge bank
as of 2026-11-1816 modes with a validated challenge corpus
Sixteen of the nineteen canonical modes draw on a validated challenge corpus. The remaining three are the Attention Lab modes, which are timing-based exercises and carry no challenge bank by design.
Computed automatically from the product's own records; source details are in stats.json.
Training modes
as of 2026-11-1819 GameMode values
Read directly from the database enum, which is the single source of truth a repo test asserts against — so this figure cannot drift from what the product actually ships.
Computed automatically from the product's own records; source details are in stats.json.
Cognitive Fingerprint axes
as of 2026-11-188 versioned cognitive dimensions
The public Cognitive Fingerprint contract defines eight named dimensions. The API and agent-facing passport response both consume that shared contract rather than maintaining independent axis rosters.
Computed automatically from the product's own records; source details are in stats.json.
What we are not measuring yet
MindFrame is pre-launch. These are the numbers people most want from a product like this, which is exactly why we will not estimate them. An unmeasured value is not zero, and it is not a guess.
Registered users
n/aNot yet measured
MindFrame is pre-launch and has no public cohort. Publishing a zero or a placeholder would imply a measurement we have not made.
We will publish this once there is a real cohort to report, together with the period it covers.
Median calibration improvement over 10 sessions
n/aNot yet measured
This is the product's central claim, so it will only be published from real user outcomes over a stated window — never from synthetic or internal runs.
The headline efficacy number. It stays unavailable until consented user data supports it, because an unearned figure here would be the single most misleading thing this page could say.
For AI agents and researchers
Every figure in this section is published in machine-readable form at /stats.json, including the ones marked unavailable and the reason each one is unavailable. See also /agents.json and /llms.txt.
What beta means for you
- Things may change, and occasionally break. Your progress and data are kept through changes.
- Every training mode stays free. Season Pass and Year Pass prices are on pricing.
- Tell us what confused you or what you would play again: contact us. What is coming next is on the roadmap.
How MindFrame uses AI
MindFrame uses Anthropic's Claude models through Anthropic's API. Most requests use a small, fast model; the Member coach may use a larger one. No other provider writes or scores text.
One feature can use a second provider. When audio narration is enabled on our servers, the text of your monthly Theory of You essay is sent to ElevenLabs to be turned into speech. This happens automatically for accounts that have an essay, and nothing else is sent to ElevenLabs. Voice Mode uses your browser's own speech service (in Chrome, audio goes to Google), not ours. Every provider is listed under subprocessors.
What the AI does
- Scores free-text reasoning when rule-based checks cannot decide.
- Writes coaching, session debriefs and insight summaries grounded in your training data.
- Suggests next practice based on your weakest areas.
What the AI never decides
- Whether an answer is right. That comes from each challenge's authored answer key, checked on our server.
- Your calibration and Brier scores. Those are plain maths (see methodology).
- Anything about your health, employment or eligibility for anything.
What it sees, and how output is checked
- Only what a feature needs, such as your answer, reasoning text and recent results. It is sent to Anthropic to generate a response and is not used to train Anthropic's models.
- Text you write is passed to the model as data, never as instructions.
- Structured AI responses are validated before use; invalid output is discarded.
- If AI is unavailable or over budget, you get a rule-based result labelled as such, never an invented one.
- Free accounts get 3 coach messages a day; you never have to use the coach to train. AI coaching can be wrong: treat it as a thinking partner, not an authority.
Safety: a training tool, not treatment
MindFrame helps you practise noticing how confident you are and how you reason. It is not medical, psychological or therapeutic advice, and it does not diagnose or treat any condition.
- Scores, archetypes and your Cognitive Fingerprint describe how you performed on MindFrame challenges. They are not measures of intelligence, mental health or clinical ability, and should not be used to make medical, employment or educational decisions about anyone.
- Some pages cite research on metacognitive therapy compared with cognitive behavioral therapy. That research studied clinical treatment delivered by professionals; it is not evidence that MindFrame treats any condition.
- Sessions are short by design. There is no penalty for stepping away. If practice stops feeling useful, stop.
If you are struggling: if you are in crisis or thinking about harming yourself, contact your local emergency number now. In the United States you can call or text 988 (Suicide & Crisis Lifeline). Please do not rely on MindFrame or its AI coach in these moments.
Security and your data
- The web app and API run on Hetzner servers in the United States (US-East), behind Cloudflare.
- All traffic is encrypted in transit; accounts use short-lived sign-in tokens.
- Your training data is never sold or used to train AI models. You can export it from your settings.
The full policy and how to report a vulnerability are on security. Who processes data for us is listed under subprocessors, and the details are in our privacy policy.