Skip to content
Truth Academy.
Truth Academy

TRUTH ACADEMY

Our commitments

What we promise students and families, and what we don't.

Our first commitment

Feynman is free for every student, with no account.

Local programs in Montclair and Jurupa Valley are free to students too, as they open.

Privacy & security

How Feynman protects students

Children use Feynman, so this section describes what the code does, followed by its known limits.

This is a hosted deployment

On this website, the AI tutor runs in your browser. What you type to it stays on your device.

The first time a learner starts the tutor, the browser downloads the model files, about 1 GB, from Hugging Face and GitHub and caches them. The site itself is served from a server, so your IP address is processed to deliver pages. There is no database.

01

We keep no student data

Most student data breaches hit a student database. Feynman has none.

No accounts
There is no sign-up, no login, no password, and no student profile.
No student database
The product persists no student records of any kind. Conversations are not written to disk, not retained after the request, and not recoverable by us afterward.
No cookies
The app sets no cookies and stores no student data in the browser. If a learner starts the in-browser tutor, the browser caches its model files so they download only once.
No analytics on learning pages
Lessons, courses, and the tutor load no analytics, ads, or trackers. The home page and the programs, partner, community, and public Leadership Program pages can use Vercel Analytics when the operator turns it on. It records page views and button clicks, with the query string removed from each URL.
Eight dependencies
The app runs on eight packages: React, React DOM, Next.js, Framer Motion, clsx, tailwind-merge, @vercel/analytics, and @mlc-ai/web-llm. Two of them make network requests of their own. @vercel/analytics sends the page events described above, only on those pages and only when analytics is on. @mlc-ai/web-llm downloads the tutor's model files from Hugging Face and GitHub when a learner starts the in-browser tutor; it does not send what the learner types.
02

How the tutor sticks to sources

The tutor answers from course resources it retrieves first, and the app adds every link itself.

Retrieval-grounded
Every answer is generated against course resources retrieved first. The model writes prose about evidence it was handed; it does not answer from memory.
Citations cannot be faked
The model is forbidden from writing URLs. Every link a student sees is rendered by the app from its own index of resources, so the model cannot invent a citation.
It says “I don’t know”
If retrieval finds no supporting source, the tutor returns a fixed refusal (“I cannot find a reliable source for this”) and points to where to look. It is designed to decline rather than guess.
Web search is opt-in
The tutor never reaches the internet on its own. A web lookup happens only when a student deliberately chooses it, and the results are then cited the same way.
No verbatim copying
The tutor is instructed never to reproduce textbook, paper, or lecture text, and to explain every idea in original words.
03

Every message is screened, in both directions

Messages to and from the tutor pass through more than one check, listed below.

Screened before the model
A deterministic rule layer runs on every student message before a single token reaches the model. It is always on and cannot be configured off.
Screened after the model
The model's own output is screened again before it is shown to a student, including a guard that watches the response as it streams and can cut it off mid-sentence.
Optional classifier
Operators can enable a second-pass model classifier for cases keywords cannot judge. It escalates; it can never downgrade a block the rules already decided.
Fails closed
If the classifier errors or times out, the rule-based verdict is used, so a message the rules flagged is still blocked.
What it screens for
Self-harm, sexual content involving minors, explicit sexual content, violent threats, weapons and dangerous-instruction requests, and hate or harassment.
04

If a child is in crisis

A student in danger is pointed to a person who can help, and their words are not logged.

The model never sees it
A message signalling self-harm is never forwarded to the AI at all. The student is shown crisis resources directly, including the 988 Suicide & Crisis Lifeline.
Their words are never sent
If a school configures a safety webhook so a counselor can follow up, we transmit only a category, the course context, and a timestamp. The student's actual words are never included, never stored, and never forwarded.
Off by default
No alert leaves the system unless a school explicitly configures the webhook. If they don't, the student still sees the crisis resources on screen.
05

Hardened against abuse and manipulation

Students will test the tutor, so every input has a hard limit.

Role injection is blocked
A student cannot smuggle in a “system” instruction to reprogram the tutor. Incoming roles are forcibly coerced to “user” or “assistant” before the model ever sees them.
Hard size caps
Requests are capped at 32 KB, conversations at 24 messages, and any single message at 4,000 characters. Oversized requests are rejected outright.
Rate limited
Twenty tutor requests per minute per client. Beyond that the API returns 429 and asks the student to slow down.
No secrets in the prompt
The tutor's instructions contain no keys, no credentials, and no private data.
06

Schools stay in control

A district can change each of these settings without giving up the lessons.

The AI can be turned off
A district can disable the tutor entirely. The platform degrades to resources-only, and every lesson, ladder, and reading path keeps working.
The safety gate
A hosted deployment refuses to run the tutor on the weaker rules-only configuration unless the operator explicitly and knowingly accepts it.
Private deployments
A school can put the entire deployment behind an access wall while they evaluate it, so nothing is public before they are ready.
Run it yourself
The most private option: run Feynman on your own hardware. In that mode the tutor is a local model, and no student data leaves the building.
07

What we do not claim

  • No filter catches everything. Our moderation is a deterministic rule layer plus an optional small classifier. That is weaker than a dedicated moderation service and will miss some content. It does not replace a teacher in the room.
  • Rate limiting is best-effort. The limiter holds its state in memory. When the site runs across multiple server instances, the limit applies per instance rather than globally, so a determined actor has more headroom than the number implies.
  • Instructions are not a security boundary. We tell the model never to reveal its instructions or accept a jailbreak. That reduces casual attempts; it does not defeat a determined one. We rely on the moderation layer and on holding no secrets, not on the model's obedience.
  • An AI can still be wrong. Grounding and citation make hallucination far less likely, not impossible. Students should click through to the linked source. When the tutor cannot find a source, it says so instead of answering.
  • The in-browser tutor has fewer safeguards. It runs the same rule-based screening and output check, but not the optional classifier, the crisis webhook, or per-client rate limiting. It also needs a browser with WebGPU and a one-time download of about 1 GB.

The full legal detail lives in our Privacy Policy, Data Practices, and Data Processing Addendum. Privacy questions go to mtorre@truth-computing.com.

Optional ways to stay in touch

Academy interest forms are separate from the Feynman learning experience. They ask only for information needed to follow up. Website analytics, when enabled, run only on the home page and the programs, partner, community, and public Leadership Program pages. Learning pages have none.

Read about interest forms and marketing analytics.

Our child safety and accessibility statements describe the current approach and the work still requiring review.