MIKOTO / AI SAFETY EVALUATION
MIKOTO is a research engineering company focused on AI safety evaluation reliability. We examine inputs, tests and scoring assumptions, then deliver reproducible evidence and proposed fixes.
Research grants · Collaborations · Corporate contracts
01 / AI SAFETY EVALUATION
Four public investigations across ControlArena, FACE-Eval and Debate-style AI control schemes. Code, observations and limits are available for inspection.
CONTROLARENA / INPUT VALIDATION
Published audit · response from PR author
Compared 24 stored backdoor inputs across 10 sampled problems with their stated constraints. Identified an unresolved specification/data conflict.
Read the findings & scope ↗PR author’s scope guidance ↗CONTROLARENA / SIDE-TEST PROBES
Offline package · published
Probed two Django tasks locally. Identified a condition that accepted both escaped and unescaped output, and compared baseline, gold and mutated behavior.
Read the findings & scope ↗FACE-EVAL / INSPECT SCORER
Issue #1 & offline reproducer · published
Reproduced a mismatch between missing-reasoning scoring and the integration README using a local source-level probe. Asked upstream to clarify missing-data and parse-failure handling.
Read the findings & scope ↗DEBATE CONTROL / SENSITIVITY BOUNDS
Manifund methods note · published
Separated two averaging schemes in archived Phase 2 results and derived assumption-dependent outer bounds on unobserved conditions.
Read the findings & scope ↗These are offline and archived-output investigations by our representative; full evaluation pipelines and live model API calls/attacks were not run. Publication does not imply upstream endorsement.
02 / SERVICES
Start with a scoped investigation, typically one to two weeks. Scope, deliverables and fees are agreed for each engagement.
AUDIT
Examine dataset constraints, test logic and scoring assumptions. Separate observed defects from unresolved specifications.
Findings / minimal cases / evidence map
VERIFY & FIX
Build runnable probes and scoped fixes. Compare behavior before and after changes, including failure paths.
Reproduction harness / proposed patch / tests
COLLABORATE
Develop validators and verification tools against agreed milestones through grants, research collaborations or corporate contracts.
Research plan / tooling / validation reports
03 / ENGINEERING BACKGROUND
Public contributions in numerical optimization and Ethereum infrastructure provide engineering context for our evaluation work.
A maintainer confirmed the optimality bug we reported with a reproducible model and an exact rational checker. The upstream-authored fix #3408 was merged on 16 October 2026.
Evidence & status ↗A maintainer independently reproduced the historical-proof bug and commented on our proposed fix. Formal approval and merge remain pending.
Evidence & status ↗04 / WORKING TOGETHER
AI tools assist our research and development. We agree how confidential code, data and tools may be used before starting.
Define the evaluation, environment and question.
Set deliverables, acceptance criteria and data handling.
Run probes and record versions, results and limits.
Deliver code, findings and the questions that remain.
05 / COMPANY
Previous company work: event-management systems, social-media analysis and research-data preparation for a university. Past company work (JA) ↗
TOKYO, JAPAN
EST. MAY 2005
LET’S BUILD THE EVIDENCE
Share the benchmark, repository or evaluation question you are working on. We welcome research collaborations, grant-funded work and scoped corporate engagements.
Discuss an evaluation ↗Opens our inquiry form.