Brain-computer interface01
62 words / min
Speech, given back
An invasive speech neuroprosthesis decoded attempted speech at 62 words per minute with a 23.8% word error rate in one participant with ALS.
Phroneme 90 · evidence-first wellness
Build a more reliable mind and body.
Build steadier routines through small, measurable behavior experiments. Every recommendation shows its evidence and limits, while Phroneme's cognitive-autonomy research remains open to inspect.
- 90 days
- four phases, one behavior at a time
- 2 minutes
- for the daily reflection loop
- 1 experiment
- to review and adjust each week
Phroneme · wellness and research
structural wireframe
continuous scan
Phroneme is developing a research instrument for a specific unresolved question: what happens to unaided reasoning, retention, transfer, and calibration as AI use changes?
VO₂ max
aerobic fitness
Credit score
financial trust
Recovery score
physical readiness
Cognitive Autonomy Index
proposed cognitive-autonomy construct
The challenge
Early evidence makes the question testable, not settled.
External studies suggest that assisted performance and later unaided performance may differ. Phroneme has not yet collected the participant data needed to estimate that difference or validate the Cognitive Autonomy Index.
The research opportunity is to build a versioned instrument, test its failure modes, and publish null as well as positive findings. The evidence chain is the product.
External preprint · Kosmyna et al. 2025
not Phroneme data0
participants, recorded across
three tool conditions
three tool conditions
Schematic ordering reported by the preprint
Brain-only
Search-assisted
AI-assisted
Bar lengths are a qualitative interface schematic, not reported effect sizes.
In a 54-person preprint, the LLM-assisted group showed the weakest EEG connectivity of three tool conditions and struggled to quote its own work. The study does not establish long-term causal harm or population-wide effects.
Review the evidence record →Conceptual illustration, not a brain image, participant trajectory, or observed effect. The direction and magnitude remain research questions.
Cognitive acceleration
Do not outsource the capacity you want to protect.
The Cognitive Acceleration Lab is a local-first experiment prototype. It records short-form accuracy and confidence, while retention and transfer remain outcomes that a delayed, fresh-form study would need to establish.
Explore the learning lab →The evidence
External research that defines the scientific neighborhood.
These examples are external evidence, not Phroneme-generated results. Each retained claim has a public evidence record with its classification and limitations.
Non-invasive decoding02
61% word accuracy
Reading type from the skull
Brain2Qwerty v2 reports 61% mean and 78% best-participant word accuracy on briefly memorized typed sentences decoded from MEG recordings.
Cognitive measurement03
54 participants
A testable warning
In a 54-person preprint, the LLM-assisted group showed the weakest EEG connectivity of three tool conditions and struggled to quote its own work. The study does not establish long-term causal harm or population-wide effects.
Brain2Qwerty
Reads the brain's output, memorized sentences you imagine typing, at roughly one word in four wrong, from a $2M half-ton MEG scanner you sit motionless inside. A landmark result, and still a lab instrument.
Phroneme
Proposes a behavioral construct focused on unaided reasoning and related domains. The current browser demo is not a validated instrument and has no population norms.
Architecture
Three proposed modules. One bounded estimate.
The current demo implements illustrative adaptive selection and 2PL scoring in the browser. Field-calibrated items, equated forms, norms, and a longitudinal trajectory model remain proposed work and must not be inferred from the demo.
Input
Reasoning response
Module 1
Adaptive items
L_IRT
Select the item with maximum information at the current ability estimate.
Module 2
IRT engine
L_2PL
Two-parameter logistic model with Bayesian ability estimation and honest SEM.
Module 3
Trajectory layer
L_long
Proposed: repeated equated forms, practice-effect controls, and uncertainty-aware change estimates.
Output
demoillustrative estimate
No validated trajectory
The browser demo currently implements only adaptive selection and illustrative 2PL scoring. Inspect the current demo; the longitudinal layer remains proposed.
Explore results
See what a trajectory looks like.
The intended product is repeated measurement with uncertainty, not one permanent label. Browse synthetic demonstration scenarios, then inspect the current assessment demo.
Synthetic demonstration · not participant data
Synthetic scenario A
Generated example of a decline larger than the assumed measurement-noise threshold.
Baseline measurement
108± 7
Synthetic baseline
Follow-up measurement
eroding91± 6-17
Synthetic follow-up
MDC₉₅ threshold
±12 pts
Observed change
-17 pts · outside assumed noise
Under these assumed values, the change exceeds the demonstration threshold. This does not show that any person declined.
The displayed trajectories are synthetic demonstration scenarios used to explain measurement uncertainty. They are not participant observations, pilot results, norms, or evidence of change.
Run the assessment demo →What makes it different
Four deliberate moves, in one screen.
Brain-training apps optimize for entertainment. Standardized tests optimize for gatekeeping. Phroneme does neither.
Eliminate01
The gamification arms race. This is the opposite of a dopamine loop.
Reduce02
Scope. The proposed instrument targets specific cognitive-autonomy domains rather than claiming to measure the whole mind.
Raise03
Transparency. Evidence class, uncertainty, version, and limitations must travel with each material claim.
Create04
A versioned research platform that can test whether longitudinal cognitive-autonomy measurement is possible.
Design comparison
What each product category is designed to do.
This is a design comparison, not a performance benchmark. It shows intended product choices and the evidence still required before those choices can become claims.
| Design dimension | Brain-training apps | Standardized tests | Phroneme CAI |
|---|---|---|---|
| Primary optimization target | Engagement and repeat use | Achievement or selection | Proposed cognitive-autonomy construct |
| Longitudinal interpretation | Varies by product | Usually secondary | Proposed; not validated |
| Uncertainty shown to users | Product-specific | Method-specific | Present in demo; parameters illustrative |
| Current evidence status | Must be evaluated product by product | Established for specific instruments and uses | Awaiting pilot data |
Qualitative design comparison only. No numeric product-performance ranking is claimed, and competing products must be evaluated individually against their own evidence.
One instrument. Two jobs to be done.
Remaining challenges
What we have not solved yet.
We publish limits openly, because honesty is the moat incumbents cannot copy without indicting their own marketing.
01
Population norms do not exist
The demo runs real item-response math on illustrative item parameters. A production CAI needs thousands of respondents, drift monitoring, and differential-item-functioning screens before scores are comparable across people.
Validation architecture →02
One point is nearly meaningless
The browser demo shows an uncertainty band derived from illustrative parameters. Phroneme has not established that repeat measurements are comparable or that change can be separated from practice and noise.
See the demo →03
Category creation is slow and expensive
Institutions can substitute grades, existing tests, or doing nothing at zero marginal cost. Phroneme has not yet established willingness to pay or a validated institutional deliverable.
Five Forces analysis →04
Neural anchor is bounded, not magic
The CAI is proposed as a behavioral instrument with an optional neural convergent-validity study. No neural validation layer currently exists, and the product is not a brain-reading system.
Reverse inference paper →Open science
Everything currently available to inspect.
Draft papers, a proposed method, product demonstrations, sources, and an explicit public evidence-status ledger.
One measurement asset. Several uses.
If the construct survives simulation, expert review, feasibility work, psychometric validation, and independent replication, a versioned longitudinal measurement system could support individuals, researchers, and institutions. That outcome is not yet established.