Research notes
Thinking in public, with rigor.
These notes document observations, hypotheses, and research directions from our work in learning, cognition, and applied modelling. They are written for partners, funders, collaborators, and serious hires who want to see how we reason before they work with us.
Notes are not marketing. They are an early record of what we believe is worth measuring, building, and debating.
A perfect braid, an ordinary classifier
Exact algebra survived every tested braid rewrite, yet lost on MNIST. A guarantee helps learning when the variation it removes is irrelevant to the task.
Our braid model preserved algebraic equivalence by construction, then lost to a smaller conventional image model. We report the measured results, the failed conversion, and the matched controls: a case study in when a mathematical guarantee becomes a useful learning bias.
- Learning & cognition
- Applied modelling
- Institutions & evaluation
After Eureka
Scientific credit in the age of industrialized reasoning — when ten million tries is not a eureka moment.
As machine reasoning becomes scalable, institutions must distinguish cognitive achievement from computational expenditure. We treat large-scale search as scarce scientific infrastructure — allocated by knowledge gain, proportionality, reproducibility, and opportunity cost — with the Clay problems as case study, not target.
- Institutions & evaluation
- Infrastructure & energy
- Human–AI interaction
Metrics that still discriminate
What funders and partners can still measure when agents saturate hiring, review, and drafting — and polish stops signaling depth.
Linguistic homogenization, verifier scarcity, and energy-limited review converge in the committee room. We argue for lineage, verification budgets, and distributional reporting — metrics that remain informative when production is cheap and form converges.
- Institutions & evaluation
- Human–AI interaction
- Learning & cognition
A latent lingua franca
What multilingual models may share internally — and why that representation becomes an institutional question once it mediates public language.
Multilingual models can align semantically equivalent expressions in shared latent spaces. We distinguish that model-internal finding from human homogenization, then ask what institutions must measure when those representations mediate translation, reasoning, and public language.
- Language
- Learning & cognition
- Human–AI interaction
Delegation, lock-in, and the shrinking verifier class
Delegation may be rational; institutions still need a workforce able to verify what it delegates.
Delegation can aggregate into sociotechnical lock-in while entry routes to verification work weaken. We separate formal mechanisms, measured junior-employment patterns, and disputed cognitive claims — then ask what institutions must preserve before the verdict is in.
- Labor & institutions
- Human–AI interaction
- Learning & cognition
Linguistic homogenization
When people, agents, and published language begin to sound the same — without necessarily meaning the same thing.
Our earlier work on learning and cognition pointed toward expressive convergence across people and models. That convergence is now visible in the open — shared jargon, tone, and rhetorical form spreading through agents and published materials, often creating an appearance of agreement without proving one.
- Learning & cognition
- Language
- Human–AI interaction
The power wall
Compute-heavy AI is physical infrastructure: its viability depends on workload depth, power, and a budget for verification.
As reasoning workloads deepen, their viability depends on context growth, hardware, power, and verification capacity. We argue that civilian AI programs need explicit depth-aware operating budgets — not just benchmark claims.
- Infrastructure & energy
- Applied modelling
- Human–AI interaction
Seconds before the wave
Earthquake early warning is a governed latency problem: measure who receives actionable seconds, not only whether an alert was sent.
Early warning is not prediction. It is governed detection, latency budgeting, and equitable last-mile delivery under mortal stakes. ML can accelerate phase picking, but it cannot substitute for sensors, drills, or accountable alert systems.
- Disaster risk reduction
- Applied modelling
- Human–AI interaction