Public research and experiments
Source-linked experiments by Stan Shih · 施博瀚.
Research artifacts are separate from released products. Each project defines its own methods, evidence and limitations.
Math Notebook
AI-assisted computational mathematics with interactive explainers, sequence computations and Lean formal checks.
The repository distinguishes small new computations, confirmations of earlier results and unfinished questions. Finite searches do not prove an open conjecture.
AIMO R01A negative study
A reproducible retrospective study of whether fixed text-structure features improve a matched-family mathematics benchmark.
The prespecified improvement was not observed. This is a negative result with replay artifacts, not a hidden-test score or a competition award.
Vesuvius HF65 alignment
Coordinate mapping and resampling tools for comparing HF65 image canvases with tifxyz surfaces, with held-out-region checks.
Reported alignment results apply to the segments and regions evaluated in the repository; they are not a guarantee for every scan.
Vesuvius layout constraints
An experiment translating two-dimensional ink layout into constraints for three-dimensional spiral fitting, with documented A/B tests and follow-up runs.
The reported GPU follow-up did not separate a small decline from seed variation. An improvement in reconstruction quality has not been established.
PHerc0139 gap investigation
A CT-based investigation of material layers and mesh geometry around a gap, with cross-sections, measurements and reproduction code.
The study covers part of the gap and leaves compressed layers unresolved. It does not claim to recover the missing winding.
Typeless Lab
A public experiment comparing speech recognition and language-model correction for Traditional Chinese dictation.
The README identifies synthetic TTS speech, a small corpus and hardware-specific timing. Reported benchmark results should not be generalized to ordinary live speech.
Vesuvius ink-depth validation
A public preregistered study of Vesuvius 3D ink labels against an independent scan.
A validation study, not a claim of a solved historical-text or mathematics problem.
Autoresearch — Index Strategies
A public experiment in hypothesis-driven automated research, with strategy evaluation, experiment records and a separate historical validation step.
The README describes a limited historical training and validation period and lists broader market-regime testing as unfinished. This directory does not independently validate its financial results or present them as investment advice.
Financial commentary factor research
A public research roadmap studying whether Chinese-language financial commentary adds predictive information beyond existing price-and-volume factors.
The June 2026 roadmap explicitly marks the core incremental-information test against Alpha101 as unfinished. Existing proof-of-concept results do not establish independent predictive value or verified investment performance.