Evidence and publications

Evidence before claims

DotrixAI intends to publish enough evidence for its technical claims to be independently evaluated.

Publications

  • No public papers yet

    Public research materials will be released as programs mature. They will be listed here.

    Research in progress
Publication philosophy

Public enough to evaluate. Protected where it matters.

DotrixAI expects to operate partly through licensable technology. Evidence should be public enough to test claims. It need not expose every commercially sensitive detail.

What we intend to publish

  • Research reports and papers
  • Methodology
  • Benchmark methodology
  • Controlled comparisons
  • Scaling results
  • Limitations
  • Selected reproducibility material

What may remain proprietary

  • Optimized implementations
  • Specific architecture details, where appropriate
  • Runtime technology and kernels
  • Compiler techniques
  • Training recipes
  • Internal research infrastructure
  • Commercially sensitive know how
Evidence standard

Three kinds of numbers, never mixed.

Measured

Observed in an experiment

Produced by a run we performed, under stated conditions and scale.

Estimated

Derived from measurements

Calculated from measured data with a stated method and assumptions.

Projected

Expected, not yet observed

An extrapolation to scales or settings we have not yet tested.

  • Baselines are strong and fairly tuned

  • One benchmark is never treated as universal proof

  • Unexpected failures are reported, not hidden

  • Limitations are stated next to results

The first program working toward published evidence is CIR.

Explore CIR Research principles