AttriByte

Transparent lab notebook

We would rather show the model than ask you to trust the score.

Attribyte publishes assumptions, limitations, validation plans, and method status.

RESEARCH PREVIEW · HIERARCHY-PRESERVING ATTRIBUTIONRead the technical overview

Current research

Hierarchy-preserving multi-touch attribution

Status: Research preview. Patent status is shown as research preview until a filing receipt exists.

The research combines a typed marketing graph, structure-aware negative-curvature representation, explicit temporal coordinate, event contribution, credit-conserving rollup, and incremental graph updates/versioning.

Can a marketing-measurement system preserve the account hierarchy, keep journey time explicit, calculate event contribution, and roll credit through the hierarchy without double-counting the outcome?

Validation plan

Narrow questions, inspectable work.

Representation

Distortion, dimensional efficiency, update stability.

Attribution

Compare with first, last, linear, time decay, Markov, and Shapley; test predictive/holdout performance where appropriate.

Conservation

Event credit sums to the defined total; ancestor rollups create no additional conversion mass.

Operations

Source coverage, versioning, reproducible runs.

Human usefulness

Whether operators can explain and act on results.

Limitations

The model is not a shortcut around evidence.

  • Attribution is not causation.
  • Missing identity/revenue data limits every model.
  • Structure-aware models may not outperform simpler models on shallow or sparse hierarchies.
  • Results depend on graph design, windows, and contribution method.
  • Research outputs should not automatically control spend without guardrails.

Research principles

Rules we intend to keep.

  • No “true attribution” claim.
  • No “most accurate” claim without benchmark.
  • Show model disagreement.
  • Do not duplicate credit.
  • Publish exact patent status and date.
  • Link results to journeys, data, and configuration.

Changelog

Publish dated model versions, schema changes, sources, validation results, limitations, and patent/publication status.

Have a dataset that can challenge the model?

We want complicated journeys, real revenue, and operators willing to tell us when the answer is wrong.