AI Engineer first-party retrieval
Incomplete metadata-heavy preview of retrieval, graph expansion, provenance, and refusal over AI Engineer program evidence.
Read the paper
Working papers · reproducible systems research
Benchmarks deserve
the whole paper.
Questions, source evidence, experimental design, measured results, charts, failure cases, and limitations—published together so an engineering claim can be inspected and reproduced.
Read the current reports ↓Current working papers
Research reports
Each paper has a checked receipt, a source-linked figure, a reproducible protocol, and an explicit evidence boundary.
Editorial policy
A result without its caveats is not a result.
Source versions, model revisions, checksums, and measurement conditions accompany the claim.
Charts show observed comparable measurements; missing human gold is never replaced by invented accuracy.
Incomplete datasets, provisional labels, missing transcripts, and unsupported conclusions appear beside the results.