WUBBERY PUBLIC PROOF NOTES
Published: 2026-07-25

WHAT IS PUBLIC
- The aggregate benchmark results in results.json and benchmark-results.csv.
- Dataset names, metric, run date, seed, baseline scores, module scores and candidate counts for the memory evaluation.
- Aggregate request, token, session and cost totals for the private-log cost replay.

MEMORY EVALUATION
- Harness: BEIR datasets.
- Metric: nDCG@10.
- Datasets: scifact, nfcorpus and fiqa.
- Baseline: dense semantic retrieval with a full scan.
- Module result: session-regime candidate selection followed by the same scoring comparison.
- Run date: 2026-07-10.
- Seed: 7.
- Falsification boundary: with the state signal disabled, the module must return to the baseline score.

COST REPLAY
- Input: the founder's complete private Claude Code usage history used for this reported run.
- Aggregate size: 34,734 requests, 9.68 billion tokens and 3,079 sessions.
- Comparison: actual logged spend versus a shadow routing decision over the same request history.
- Cache pricing was held constant across comparisons.
- Raw prompts and session content are not public because they contain private material.

IMPORTANT LIMITS
- These files make the currently published outputs accessible; they do not make the sealed module implementation public.
- The numbers are not a guarantee for a different workload.
- A prospective user should request a shadow evaluation using their own traffic and receive their own result before making a commercial decision.
- Any claim that cannot be supported by an accessible aggregate output should be treated as provisional, not as a product promise.