Every human being. Any subject. Any domain. Free, forever.

Guides and learners

Everyone is a learner. Whoever is a step ahead, right now, on this, is the guide.

A parent, a teacher, a mentor, a manager, a grandson with a phone, a nurse training the next nurse: the guide is whoever is a step ahead on the thing in front of the two of you. The gap is small, it is temporary, and it moves. In every sample there is a moment where the learner is a step ahead of the guide on something and leads. We call that swinging leads, after climbing partners who swap the lead as they go up.

Learning happens within reach: just past what a person can do alone, and not further. Meta DAX exists to put the next step within reach for both of them, and then get out of the way.

See it, in the samples →

The measure under all of it

Human delta value: what a person brings. Reclaimed cognition: what they do with what they get back.

Every time Meta DAX takes cognitive labour off a guide — the sequencing, the worked examples, the quiz, the record of what happened last time — one question is left, and it is the whole point: what is it that only you can do here, and how do we help you do it as well as you can?

Look at the guide’s side of any sample. Megan is not handed a lesson to deliver; she is handed what to say, what to skip, which slips are normal, and one line of why under each step, because by session three she was explaining it in her own words and asking if that was allowed. It was. The page got out of her way so she could watch Liam. Dave is told how to coach without grabbing the keyboard. Mara gets an observation lens for her next visit and a note on where Jordan sits in his own zone of proximal development as a teacher. Renée gets one line on how to coach Marcus without taking over. The preparing is off them. The noticing is theirs.

That is reclaimed cognition. Now that the guide has it back, they have the chance to learn the part no system can do: how to bring out the best in the learner. Meta DAX is built to invert the ratio of learners to guides, and to hand the guide back the cognition the busywork took, on purpose, for the problems they could previously only dream about. Human delta value has two halves, and both count: what only a human being can currently do, and what a person and their cognitive companion reach together that neither would alone. It is the thread through everything here, not a one-off measurement, and it is close to the point.

It comes with a demand. A person and their companion are one unit, and the unit is only as strong as the human’s best in the moment. However good the companion becomes, a person who cannot bring their best gets less from it, not more. We hold that as a hypothesis with a kill line, and Meta DAX is where we intend to test it, in the open, with the result published either way. Everybody needs to step up.

Operationally: at the end of a session the guide says, in their own words, what was taken off them and what they spent their attention on instead. It is recorded as they said it. Never inferred from logs, never reconstructed by a model, never reduced to minutes. A system that cannot say what the human brought has stopped extending people and started replacing them.

The definition · Human delta value: the measure · Where it came from

How it works

One sentence from a guide becomes a course. A learner can ask why as many times as they like.

What has actually run

Two of the first fourteen experiments, first cut, both automated: a model played the guide, and in the learner runs a simulated adult played the learner. In the guide run, an empty repository became a course with seven objectives, the first page at about eight minutes; a second run produced Newton’s Laws of Motion, three lessons, seven modules, twenty-one objectives, every page validated. In the learner run on that course, six follow-up questions to a depth of four, a quiz, a saved record, and three bugs found, two fixed with regression tests. None of it is a pilot with a person. Every number, and what each run does not show, is in the repository with the date on it.

More have run since. E04 and E05 ran on 9 October, X13 and X14 on 10 October, and on 18 and 19 October came a wave of twenty-two first runs and seven follow-ups, set out under the experiments below. A model or a script sat in the people’s chairs in every one, and the files say which. No person has been in a chair yet.

The runs, in the repository · Sixteen concepts, one at a time

The experiments

Fourteen small ones first. Then the legwork. Then the twelve we can already picture. Twenty-four of the twenty-eight have now run, none with a person, and several came back negative.

The small ones are how we earn the big ones: nothing in the third tier can run honestly until the first two have. A result appears only when an actual run produced it, dated, whichever way it went.

Where we are starting

As of 20 October 2026, copied from the experiment cards and the findings of 19 October. Eleven of the fourteen have run. E02, E07 and E10 are blocked on things we do not yet have: a second vendor’s key, access to an open curriculum, and a small-model runtime.

The findings of 18 and 19 October → · The bottleneck register

What has run since, and what it showed

Two experiments ran on 2026-10-10. X13, with its harder follow-up X13b, tested whether a small, fast model can answer by default while a larger, slower one steps in only when a turn needs it. It came back a negative: the small model matched the large one, so there was little for the escalation to catch. X14 tested whether fetching only the piece of a course a learner needs holds answer quality as the library grows. It came back a positive: quality held at every library size, with about 75× fewer context tokens per turn at the large size (estimated), on six graded questions, so read it as directional. A negative is a result: it is written up, dated, and kept. The full account is in the findings.

The findings: X13 and X14, run 2026-10-10 →

What ran on 18 and 19 October, and what failed

On 18 October every experiment in the first fourteen that could run did, along with eleven of the twelve scenes: twenty-two first runs. On 19 October seven follow-ups ran on the ones that had failed or come back unclear; one of them re-scored earlier outputs. A model or a script played the people, and the findings label each result SIMULATED or deterministic. Of the twenty-two first runs, eleven passed, two passed with a caveat, four failed and five were inconclusive.

The failures matter most. The age wall (E14) stopped none of eight explicit minors. A patched prompt then stopped 21 of the 22 earlier probes and 6 of 6 new ones, but one opener, “skip the age check”, carries no age cue at all, and a first-turn prompt cannot catch it; it needs a structural check. A full learner session (E03) failed validation 44 times. The prompts (E04) missed the 90 percent line, mostly because the runner printed its reasoning before the JSON, and they have not been tried on a second vendor. The gate meant to keep a company’s private content inside (X08) let 10 of 10 paraphrases through. And the nurse’s walkthrough (X03) passed with a caveat: review caught 3 of 3 planted errors and missed the 1 planted omission, so the module was not safe to deliver.

Blocked: E02 and the cross-vendor half of E04 need a second vendor’s key; E07 needs network access to an open curriculum; E10, and the real half of X07, need a small-model runtime, because the run environment had no inference engine; X12 is the capstone and comes last by design. Every run adds rows to a register of bottlenecks, now 93, of which 89 are open.

The findings: 18 and 19 October → · The bottleneck register · What the runs say about the design

The legwork

What has to exist before any of the big ones can run honestly. None of it is glamorous, and all of it is where the scale lives.

Identifiers and schemas that hold at a million nodes

Today’s ids encode ancestry and registries are per module. Both have to be shown to hold, or be redesigned, at a scale no fixture reaches.

A registry at scale

Reuse decisions against fifty planted nodes are one thing; against every course in a language, with duplicates arriving from many guides at once, is another.

Reuse across languages, without copying

A core rendered into a second language is a shared asset; a hobby rendering stays personal. The rule exists; the machinery does not.

The five-minute bundle

The weekly loop at 50 kbps, then the bundle that carries enough teaching for a week offline, with nothing in it the learner did not need.

Certification

A prompt is trusted for a named model set because evals say so, not because of reputation. No model set has passed yet: the harness has run on one vendor, and its stop rule has not been met.

The constraint fields, null today, made live

Every node has a provenance file with fields reserved for a constraint and signing layer. They are null. A live Constraint Protocol record under every contribution is the destination.

The learner’s privacy, at scale and across years

Pseudonyms and a private repository work for one learner and one course; eleven courses and six years is a different problem.

Reclaimed cognition, recorded

A place in the session record for the person’s own account of what was taken off them and what they spent it on, written as they gave it. No experiment with a person runs without it.

Where we want to get to

Twelve experiments, each written as a scene first, because an experiment that cannot be pictured as one person on one day cannot be designed. The people are fictional; the places, languages and subjects are real. Each is planned as a study with people, and none has run with people. Eleven have had a first run with a model or a script in the people’s chairs, and the line under each says what came back; X12 has not run. Those runs measured no person, so none carries a human delta value line. With people, every one of them records human delta value, both halves, and reclaimed cognition in the person’s own words.

1. X01 One thread, two worlds. A guide’s move, made once in Indonesian for a kite flyer stuck on the third law, captured as a technique and offered in Spanish to a kite flyer in Guatemala who shares nothing else with him. The floor rises for everyone after.ran 9, 18 and 19 Oct (SIMULATED): inconclusive. Reuse landed in every case, across languages, but on 19 Oct a control that reinvents the move was slightly faster, 1.67 attempts against 1.83, inside a run-to-run spread of 0.67.
2. X02 Many hands, one course. A beekeeping course forked across four languages, learners asking at once, until its curriculum is mostly promoted follow-ups.ran 9 and 18 Oct (SIMULATED): PASS. Conflicts per 100 commits fell from 80.8 to 13.6, and unreviewed nodes crossing between forks from 841–854 to 0.
3. X03 The nurse’s walkthrough. A walkthrough that first asks what a guide brings, and gives a charge nurse only the stages she cannot do herself.ran 18 Oct (SIMULATED): PASS with a significant caveat. Review caught 3 of 3 planted errors and missed the 1 planted omission, so the module was not safe to deliver.
4. X04 The dead-end atlas. Where learners get lost, shared as shapes and never as records, mapped per module and re-taught at the spot.ran 9 and 18 Oct (SIMULATED): inconclusive. Stable ids made one stall one hole, not three; whether holes predict anything was not tested.
5. X05 Flagged by the learner. Steering measured two ways: by learners who flag it, and by an evaluator blind to them.ran 18 Oct (SIMULATED): inconclusive. One model played learner and evaluator, so their 0.95 agreement shows little.
6. X06 Same course, every bandwidth. One course for a five-minute-a-day learner, a patchy commuter and an always-on companion, and what the bundle must carry.ran 18 Oct (SIMULATED): PASS. A week fits in 965 KB of a 13.1 MB weekly allowance at 50 kbps.
7. X07 The pocket model. A small model on a phone holding the conversation while the bundle holds the teaching.ran 18 and 19 Oct: inconclusive. No real model ran: a fake-model harness passed 20 of 20, and the run environment had no inference engine.
8. X08 Behind the company wall. The same infrastructure on a company’s own git, learner records still the learner’s, one technique sent back out through a gate.ran 18 and 19 Oct: FAIL, then inconclusive. The exact-term gate let 10 of 10 paraphrases through. A synonym layer then caught 20 of 20, but its table and the paraphrases came from the same fixtures, and the judge was simulated.
9. X09 Nothing quietly edited. Every contribution on a live Constraint Protocol record, and whether knowledge and intelligence compound when nothing can be erased.ran 18 Oct (deterministic script): PASS on detection. An append-only log caught 18 of 20 silent rewrites; editing in place caught none, by construction. The log grew 10.7 times larger.
10. X10 One record, many years. One learner’s private record across eleven courses and six years, carried into her cognitive companion, with every carry-over hers to refuse.ran 18 Oct (SIMULATED): PASS. None of 13 refused carry-overs was offered.
11. X11 The negative space. The unexplored parts of a subject shaded on the tree and offered, not pushed, with the regions nobody takes recorded too.ran 9 and 18 Oct (SIMULATED): inconclusive, with one firm negative. Neutral wording cleared the offers, but an independent evaluator flagged all 11 atlas markers as steering.
12. X12 Five minutes to master’s. The long one. A phone, five minutes a day, a blind-graded external yardstick agreed in advance, and the result published either way.designed, not run. The capstone, last by design.

The twelve scenes in full: the machinery, the question, the measure, the kill line, what compounds · The first fourteen

The factory, the coach, the profiles

What builds the samples you can see: a course factory, a guide coach, and ten context profiles.

The deep showcases and the sample packages are not written by hand. They come out of a small set of meta prompts anyone can read and run. Two build the course; one sits with the guide; the profiles tell all three what a learner in a given setting actually needs.

The course factory

MP-11, the curriculum architect, which turns one sentence into a sequenced course with its dependencies and likely misconceptions, and MP-12, the batch builder, which renders the pages for it in one pass.

The guide coach

MP-13, the guide coach: the prompt behind “Tell us and ask” on the guide’s side. The guide pastes a real problem and gets coached on it. The guide asks; a child never does.

The ten context profiles

One profile per learning setting, so a page built for a seven-year-old with a caregiver and a page built for an apprentice on a job site start from the right assumptions: Early childhood, with a caregiver, Primary school, Secondary school, Higher education, Vocational trades and apprenticeship, Professional continuing education, Workplace and HR training, Adult basic education and language, Community and humanitarian, Lifelong and elder learners.

On the Constraint Protocol

When nothing can be quietly edited, what one person works out compounds for everyone after them.

The Constraint Protocol is the ground Meta DAX is built to stand on: every party declares in advance what it will not do, every consequential act is signed and witnessed onto a record nobody can quietly edit, and anyone can dispute any record. Today every node in Meta DAX carries the fields for that record, and they are null. Making them live is on the list above. Here is what it looks like when they are, told the only way we know how to tell it, as scenes. Each is planned as a study with people (X01, X10 and X09) and has not run with people; the first runs above used a model or a script.

The Constraint Protocol is a public working draft, run in full once so far. It keeps the record; it prevents nothing. What Meta DAX adds is the reason to keep one: a world where a parent in one country, a tutor in another and a learner in a third can each build on what the others worked out, and see exactly who stood behind every piece of it.

CP-SPEC v0.5, on GitHub · Knowledge, intelligence and wisdom, as defined here

Honest limits

Meta DAX is a research project of the DAX Foundation: a set of experiments before it is a product. The samples were written with help from AI from a short brief about two made-up people, and read in full by a person before they went up; they show what the system is aiming to hand people, not what the automated pipeline produces today. Twenty-four of the twenty-eight experiments have run, every one with a model or a script in the people’s chairs and none with a person, and several came back negative: the age wall, the learner session, the prompt certification and the outward gate among them. Together they do not yet show what a person experiences or learns, or whether anything compounds between two of them. Nothing here is production, and no element of the architecture is set in stone; several of the experiments will change it, and some will kill parts of it, which is what they are for. When one runs, its result lands in a dated file, whichever way it goes.

The ask

Give us a real scenario. We will build you a sample, free.

If you are a step ahead of someone on something and the two of you are stuck, tell us. A parent who homeschools and a child who hates worksheets. A new teacher and a class that nods and then fails the exit task. A grandson and a grandmother with a new phone. Send a few lines:

• who the two of you are, in general terms (a mom and her eleven-year-old; a charge nurse and a new hire), with no names and nothing that identifies anyone;
• the subject, and the exact spot where it keeps going wrong;
• what the learner loves, because that is what the page gets built through;
• what you have already tried.

We will hand you back a page like the ones in the samples: one side for you, one side for them. You read it before you use it; nothing on it is stored about anyone; a child uses their side with you beside them. Reply where you found this, or write to Jason on LinkedIn. Then tell us how it went, and the next one comes back shaped by it. That is the whole method, and it is how the real experiments will start.

MetaDAX is free. It is built on paid compute. If it helped, a coffee on Ko-fi or a GitHub sponsorship keeps it going.

Pick your chair → · The samples · The repository

Formerly brunswicklab.com, now part of DAX Foundation Research.