Temporal retrieval benchmark

Problem

A reference that resembles the current image can still describe the wrong interval. The benchmark measures how often retrieved records mismatch the query interval.

Contribution

Relationship to research: Independent implementation. A new benchmark inspired by the interval-conditioning idea in a manuscript shared with the account owner; not a reproduction of that manuscript's benchmark.

Account owner’s role: Commissioned, reviewed and published the implementation. The code was produced with AI-assisted generation on commission and reviewed before release; it is not the original code of any cited publication.

What was built

• Seeded corpus generator with ordered studies, interval buckets and a reference bank that excludes test patients
• Lexical, hashed-vector, metadata-conditioned and oracle systems as separate named classes
• recall@k, graded nDCG@k, interval-mismatch rate and hit rate, pooled and per-query

Method

Relevance is defined explicitly from finding, region, interval bucket and source patient; the ablation flag removes the bucket filter only.

Data and access

Synthetic patients and references, fixture v1, seed 42.

Evaluation protocol

Compare table computed by the demo command; numbers differ across systems and no system is claimed better.

Results

No measured result is published for this project. Demonstration outputs from the repository’s own synthetic fixtures are labelled as such in the repository and are not presented here as results.

Limitations

• Synthetic relevance labels encode the benchmark's own definition of relevance.

Reproducibility

Offline; the corpus regenerates from the seed.

Repository metadata: Python, MIT, last push 2026-09-13, 0 stars.

Source publication and links

No publication is attached to this project.