Measure, upgrade, and defend the mind.

01 · In briefWhat this is.

MindTech is building a belief map. A guided conversation produces the first map of the beliefs a person runs on, and the map is read two ways. Read defensively, it shows where a person is most easily moved and by which tactics. Read for performance, it shows which beliefs are limiting them. The work is one founder, and the ask this autumn is three beta partners, each with a small cohort of adults. The purpose of the beta is to establish the map's reliability.

If you open doors, the useful introduction is to a head of clinical or product at a consumer health company, a research lead at a clinic with an ethics route, a head of leadership development, or a security training lead. Each has a section below.

02 · Why nowOne-to-one persuasion, adaptive and in real time, now runs at population scale.

Scripture, propaganda, and advertising have reached minds at scale for a long time. What is new is that a large language model can hold a persuasive conversation with millions of people at once, one to one, adapting to each person as it goes. The bet of this company is that the same capability, pointed the other way, can read a mind well enough to measure it, at a cost and depth that were not available before. That is a hypothesis. The rest of this page says what has been built to test it.

Two minds meet at that interface and both have weak points. The human mind is open to manipulation, distortion, and false belief. The AI mind is open to being poisoned, bent, and turned. MindTech works both, because a defence for the first that ignores the second does not hold.

Further reading. The Human Mind Is Up for Grabs, the essay behind this section

03 · The problemBelief structure can be measured today. It is slow, self-reported, and taken once.

Instruments exist. The Young Schema Questionnaire, the Dysfunctional Attitude Scale, and repertory grids all map belief structure, and routine outcome monitoring tracks symptoms week to week. They are long, they rely on what a person can report about themselves, they are administered once or twice, and they do not update between sessions. Outside of research settings most practitioners do not use them at all.

The gap is not that nothing exists. It is that nothing runs continuously, nothing is cheap enough to use every week, and nothing reads the layer beneath what a person volunteers.

Further reading. An X-ray for the Mind, on why measurement is the base of the stack

04 · Who this is built withFour settings, one rule on who may see a map.

The instrument gets its range from being used on very different people, which is why the work is done with partners rather than for one market. Each setting brings something the others cannot.

The rule. An individual map, susceptibility score, or tactic list is released to the subject and to no one else. No employer, partner, insurer, or agency receives one under any commercial arrangement. Cohort readouts are aggregate only. This sits in every agreement and is not negotiable.

Further reading. What we build, the Work page

05 · Who builds thisOne founder today. Medical doctor and AI systems engineer.

Joshua Tanner trained at Stanford in medicine and in AI and data science, 2018 to 2025, where he was Co-Executive Director of HealthAI and founded and taught the graduate course Deep Learning in Medicine. Eight years building AI systems. Fifteen or more healthcare AI projects across ten countries, including a dermatology triage AI deployed commercially in Singapore and a stroke-routing imaging system.

Agent verification

Psychological red-teaming and verification of AI agents, in production with two organisations through a partner platform. The engagement is under NDA. The counterparties can be described on request, subject to the partner's consent.

Team

MindTech is one founder, with contractors per build. A responsible clinician and a psychometrician join for the beta.

Further reading. About Josh, the longer version

06 · What we buildThe belief map. A bounded set of named constructs, and a rule for what gets on it.

A guided conversation produces the first map of the beliefs a person runs on, including beliefs they have not put into words and beliefs absorbed from outside. Beliefs are ordered by depth, from the reactions a person reports day to day down to the commitments underneath them. That ordering follows the cognitive-therapy model of automatic thoughts, conditional rules, and core beliefs.

The deep layer is anchored to a fixed set of constructs from the literature rather than invented. Early maladaptive schemas from schema therapy, the three core self-verdicts from cognitive therapy, and the assumptive-world beliefs from trauma research. A surface belief earns a place on the map only if the chain of why-do-I-believe-this lands on one of those anchors. A belief that will not ground is parked, not forced.

What a reading looks like. The output is prose, one belief at a time, with the utterance it came from and the anchor it traced to shown alongside. It is not a lab panel and is not presented as one. Establishing its reliability is the first beta's purpose, which is why the beta is small.

How much evidence is that? At the start, not much, and the page does not pretend otherwise. The conversation seeds the map. It does not stand as its evidence base. The read of a person's writing is a second angle, and from then on each program's evidence log and each defence exercise is written back against the map, so the base grows with use instead of being taken once. The rule for adding an angle is that it does not predict the angles already in use. A solid object cannot be rebuilt from a single shadow, and a mind cannot be rebuilt from a single kind of evidence about it. Further angles of that kind are in view and are not described here.

Further reading. A read from the cultural cognition profiler, on a composed subject, in the Lab · An X-ray for the Mind

07 · The stackMeasure, upgrade, defend. On an AI that is tested first.

Further reading. Agent verification, on the Work page · Videos on manipulation and defence

08 · How it worksIntake, map, program, defend.

Intake

A guided conversation, text or voice, since spoken answers keep tone and hesitation that typed ones drop.

Map

Extraction against the belief canon with the downward trace. A separate writing-based profiler reads a person's bio and writing into a ranked set of programs.

Program

Behavioural experiments and evidence logs against a chosen belief, on a daily schedule, drawn from a consented, anonymised library.

Defend

Exercises against an AI running a named tactic. Scoring is detection. Whether that transfers to unannounced attempts is the outcome question a second study asks.

Update

Program results and defence outcomes are written back as evidence against the map, so there is no re-mapping session.

Further reading. Try the working tools

09 · Scope, safety, and dataWhat it is for, what it will not do, and how it is judged.

Is this treatment?

No. There is no clinician relationship, no disorder is targeted, and no treatment claim is made. The intended use is self-directed assessment and training for adults. A per-jurisdiction memo on device classification, written by someone qualified to write it, comes before any clinic pilot enrols a person, and the clinic door needs research ethics approval first.

What happens if someone is in crisis?

Four things do the work, and none of them is a classifier. People are screened against written exclusion criteria before they enrol. The conversation carries plain escalation rules on stated triggers that end the session and show the person their route out, which is also shown at every step. The partner's clinician is on call during the cohort window. And every session is reviewed by a person afterwards. Exclusion criteria, the triggers, a stopping rule, and an adverse event definition are written with that clinician before a cohort starts.

Who holds the data?

Where MindTech runs the assessment, MindTech is the controller, the person can see, export, and delete their map on request, and retention is fixed in the protocol. Where a partner platform runs it, the partner is the controller inside its own regulation and MindTech processes under the person's consent. In both cases no individual map, score, or tactic list leaves to anyone else, and no aggregate readout is produced below a minimum cohort size fixed in the protocol. Inference runs on the device where the platform allows it, and where it does not, that is stated to the person.

How do you know it works?

Not by asking the map. The endpoint is independent of the instrument. For defence, a held-out set of attempts delivered without warning, before and after, against a comparison group. For the map itself, test-retest over a fixed interval and agreement with an established schema instrument. The beta is that study.

What is the failure you design against?

AI companions tend to fail in one of three ways. The person hands over their thinking and gets weaker, the model flatters instead of correcting, or a system that has learned a person is used to steer them. Each is treated as a product failure, and the independent endpoint above is how the first is caught.

Further reading. Videos on manipulation, coherence, and defence

10 · Where the work standsWhat exists.

The belief canon

Structured by depth, anchored to the literature, in three views. Written and reviewed.

The intake and assessment harness

The guided conversation and the extraction against the canon, running end to end.

The defence engine

Exercises against an AI running a named tactic, running.

Agent verification

In production with two organisations through a partner platform, under NDA.

The cultural cognition profiler

A writing-based read of the programs a person runs, with a world map of program prevalence by country. Open to try in the Lab.

11 · The askThree beta partners this autumn. The price of the beta is the first reliability study.

The first product to run with people is the defensive reading of the map. A person is guided through their own susceptibility to social manipulation and handed a plan to shore it up. The beta is small and reviewed by hand, on a bounded set of manipulation tactics, with the evidence behind each stated plainly, including where it is thin.

The design. Each cohort takes the assessment twice, a fixed interval apart, and once alongside an established schema questionnaire. Cohort size, the interval, and the two pass thresholds are set with a contracted psychometrician and written into the protocol before anyone enrols. The beta stops if test-retest or instrument agreement falls below its threshold, or on an unresolved adverse event, and the partner gets the report on why. Whether the defence exercises transfer to unannounced attempts is a second study and is not promised in this one.

Further reading. Write to Josh

12 · After the betaWhat a partner build looks like. Four stages, one gate that is about people.

Each stage is about four weeks. We supply the engine, the spec, and the measurement. The partner's team builds the surface.

Further reading. The essays behind the method

13 · How to work with usOne question decides the next step.

Do you have a cohort of adults who would take the assessment this autumn, and someone who can be the named counterpart. If yes, the next step is a call and a one-page beta agreement with the stop rule in it. If you are a clinic or a research group, the same, plus your ethics route. If you want to build with the engine in your field, that conversation starts after the beta.

Joshua reads every message himself.

Further reading. Contact form · mindwarelabs.ai, the product's front door · Overview from January 2026 (PDF)

FilesThe same briefing as slides.

Download the slides (PDF) · Overview, January 2026 (PDF) · Write to Joshua · joshua@joshuatanner.ai