The Elora Taurus Project · Explore our work
Controlled experiments in how selected memory is represented, admitted and submitted to a model, with revisioned treatments and retained evidence.
Format comparison
Natural language, mixed, JSON-only and positional-with-legend representations use the same selected evidence. Deterministic rendering is distinct from memory selection and model inference.
Understand memory and recallExperimental harness
Registered comparison arms retain separate histories and revisioned treatments. Short validation, boundary and full studies support controlled investigation without treating one successful answer as general effectiveness.
Explore Observer researchEvidence and budgets
Hash bindings connect rendered memory to retained requests. Protected context fitting records budget denials and integrity failures rather than silently discarding the treatment. Reports separate preparation, submission and response observations.
Read about execution evidenceInspect the treatment revision, selected source evidence, admitted and submitted request, budget outcomes and scoring dimensions. Format/reference scores and exact-field content assessment answer different questions. A historical run retains its original protocol; a newer source revision needs its own validation.
This is an experimental comparison mechanism, separate from production learning and from the Adaptive Working Knowledge selector. The implementation and its bounded local validation are documented as E111 in Engineering. The separately published reports carry their own results and limitations.
