Reasoning Independent AI research · Japan
Japan's workforce is shrinking while the work that keeps society running is not. We research AI that expands what a smaller workforce can do — reasoning systems, physical robotics and autonomous agents — and we measure whether it actually does.
Ask a clear question.
→Make it testable.
→Show the evidence.
01 / The problem
The work does not disappear when the workers do. Roads still need upkeep, patients still need care, and small firms still need to close their books.
Projected shortfall as a share of labor demand, by occupation
| Occupation | Shortfall (thousand workers) | Share of demand (%) |
|---|---|---|
| Nursing care | 580 | 25.3 |
| Sales | 1,089 | 24.8 |
| Transport & driving | 998 | 24.2 |
| Construction | 657 | 22.0 |
| Healthcare professionals | 816 | 17.5 |
| Food service | 566 | 15.1 |
| Production | 1,124 | 13.3 |
| Office, technical & professional | 1,566 | 6.8 |
Office, technical and professional work has the largest absolute shortfall but the smallest share of demand; the report describes supply and demand there as roughly balanced.
02 / Our working thesis
Benchmark scores are a means, not the goal.
For every project we ask: how much real work,
work that would otherwise go undone,
can this system reliably take on — and how would we know?
AI Labor Offset Ratio = AI-enabled labor-equivalent gain ÷ projected labor loss
03 / Four research directions
Each direction starts with a practical question
and states what would count as evidence
before we claim anything.
Society & Systems · Infrastructure
Where will the work go undone, and how much of it can AI realistically cover?
Mapping Japan's labor gap sector by sector, from care to infrastructure upkeep to regional services, and modeling how far AI-driven productivity could close it.
The AI Labor Offset Ratio by sector.
Dated public statistics, transparent scenarios and sensitivity analysis.
Society & Systems · Physical AI
Which physical tasks can robots take on, and when is a robot worth it?
Building on Japan's strengths in robotics and precision engineering to find which tasks are deployable, where robot foundation models beat fixed automation, and where a robot pays for itself.
A Physical AI Deployability Index and return-on-investment thresholds by sector.
A task taxonomy, repeatable tests in bounded conditions, and total-cost analysis against vacancy and service-failure costs.
Society & Systems · Autonomous organizations
Can a small firm keep every function running without hiring for each one?
AI agents as an extra layer of operating capacity for small and mid-sized firms — accounting, HR, procurement, support and compliance — with five levels of authority from read-only to autonomous.
Workflow completion, human review burden and error recovery.
Traceable decisions, clear permission boundaries and a comparison with existing workflows.
Science · Cellular aging
Can AI help us understand how cells age, so people stay healthy and independent for longer?
A computational survey of cellular aging — senescence, iPS cells, partial reprogramming and protein design — using public data and published research only.
Reproducibility on public aging datasets and robustness on held-out data.
Dataset provenance, leakage controls and independent laboratory validation.
Computational and exploratory. No treatment or longevity benefit is claimed.
Foundation · AI safety
Can tool-using AI agents be kept within approved limits, and can that be shown with evidence rather than assumed?
A research track that supports every direction above: adversarial testing, prompt-injection resistance, least-privilege tool use and independent monitoring of AI agents before they take on operational work.
Attack success under adaptive red-teaming, monitor recall at a fixed false-positive rate, and the share of consequential agent actions covered by independent checks.
Fixed scenarios and baselines, checks against evaluation gaming, and a traceable record linking each claim to its test.
Active research track. No safety claims are made until a dated evaluation report is published.
Read the research direction04 / Our stance
Care homes, regional services and small firms will not run frontier-scale models. We study how much capability small models deliver with limited data, compute and staff, and exactly where they fail.
Evidence-led incident triage, runbook-guided decisions and disciplined failure analysis, studied in controlled scenarios.
Operational work cannot absorb confident mistakes. We build reasoning that cites its sources, states its uncertainty and stops short of conclusions the evidence does not support.
05 / How we'll know
A higher score is not proof. Each direction moves
through the same four stages, and we report
what fails as well as what works.
Define the task, the constraints and what success would mean.
Fix the test, the data version and the resources, so every result is compared like with like.
Report regressions and failure modes alongside gains.
Publish dated results with their uncertainty and limitations intact.
06 / Where we are
Founded by Arnel Perez Posadas
Reasoning Frontier is an independent research practice spanning AI reasoning, evaluation, operational reliability, industrial transformation and AI-assisted science. Today that means public research directions, prototypes and controlled tests. Dated, reproducible results will be published here as they are completed.
Until then, nothing on this site is evidence of production, security, medical, safety or investment performance.
Contact details will be added once a dedicated public channel is verified.