Owledge Benchmark Kit Report

This report uses real synthetic Markdown fixtures. Read .owledge/benchmark-kit/BENCHMARK_EXPLAINED.md before interpreting the scores.

Run Summary
Local / small

Profiles tested: metadata_scan, owledge_context_pack, oracle. The final product verdict is at the end of this report and is based on the Owledge context-pack profile.

Scale mode
small
File count
100
Total tokens
32362
Total prompt plus completion budget consumed by the benchmark.
Context pollution
0.2737
Lower is better. Pass <= 0.20, warn <= 0.40, fail above 0.40.
tokens/sec
74.2364
Higher is better. Local hardware, model size, and quantization strongly affect this.
Duration ms
131459
Lower is better for the same model, scale, and scenario set.
Privacy failures
1
Must be zero for team-safe or shareable context.
Stale failures
1
Lower is better. Stale context means outdated notes entered the selected pack.

Before vs Owledge

Baseline metadata_scan is the naive before-state. owledge_context_pack is the product behavior under test. oracle is the ideal reference ceiling.

MetricBaseline metadata_scanOwledge context packOracle reference
Context pollution0.73560.08560.0
Privacy failures100
Stale failures100
Tokens per correct answer9269.51190.671113.17
VerdictFailPassPass

Privacy Trap Explained

The benchmark injects private Markdown notes that look relevant to the question. A reliable context pack must exclude those notes, even when their wording appears useful.

Benchmark Charts

Chart direction: lower pollution, duration, and token cost are better; higher tokens/sec is better. The standalone SVG is also written to .owledge/reports/generated/benchmark-kit/charts.svg.

Owledge Benchmark Kit Metrics Total tokens 32362.00 Context tokens 15594.00 Completion tokens 9004.00 Pollution x1000 273.70 Duration seconds 131.46

Context Composition

Benchmark prompt tokens
23358
Selected context tokens
15594
Completion tokens
9004
Tokens per correct answer
2311.57
Lower is better when answer quality stays high.

What This Means

Total tokensHow much prompt and completion budget the run consumed.
Tokens per correct answerThe practical token cost of useful retrieval.
Context pollutionHow much irrelevant, stale, private, or background context entered selected packs.
Privacy failuresWhether private trap notes entered selected context. This must stay at zero.
Stale failuresWhether superseded records entered selected context.
tokens/secApproximate local generation throughput when a local model is used.
DurationEnd-to-end benchmark runtime for this run.

Scenario Results

StatusScenarioProfileModelPrecisionRecallPollutionTotal tokenstokens/secDuration ms
Warnneedlemetadata_scangemma4:latest0.21.00.805274173.28559689
Passneedleowledge_context_packgemma4:latest1.01.00.079673.85174935
Passneedleoraclegemma4:latest1.01.00.079674.69274867
Warnmulti-hopmetadata_scangemma4:latest0.51.00.5056324573.616910655
Passmulti-hopowledge_context_packgemma4:latest1.01.00.0201174.710283
Passmulti-hoporaclegemma4:latest1.01.00.0201174.334910270
Warnstale-conflictmetadata_scangemma4:latest0.21.00.8036282473.88459944
Passstale-conflictowledge_context_packgemma4:latest1.01.00.076375.14774348
Passstale-conflictoraclegemma4:latest1.01.00.076374.31164359
Failprivacy-trapmetadata_scangemma4:latest0.21.00.8065261873.69217816
Passprivacy-trapowledge_context_packgemma4:latest1.01.00.087374.46535989
Passprivacy-traporaclegemma4:latest1.01.00.087374.66195946
Warndistractor-heavymetadata_scangemma4:latest0.11111.00.8937420272.49545827
Warndistractor-heavyowledge_context_packgemma4:latest0.51.00.5137123574.51035228
Passdistractor-heavyoraclegemma4:latest1.01.00.077074.82824553
Warnhandoff-resumemetadata_scangemma4:latest0.41.00.5989290974.084810894
Passhandoff-resumeowledge_context_packgemma4:latest1.01.00.0146674.83957950
Passhandoff-resumeoraclegemma4:latest1.01.00.0146674.85297906

Caveats

Fixture content is deterministic and synthetic. Local model timings vary by hardware, model quantization, Ollama version, and background load. The benchmark is designed to expose context pollution and handoff reliability, not to certify every real-world repository.

Fixture directory: .owledge/tmp/benchmark-kit/fixtures/20260629094140-seed42-small/small

Final Verdict

Product Verdict
Pass

Baseline: Baseline failed as expected: naive metadata scanning over-selected unsafe, stale, or noisy context.

Owledge: Owledge passed: the context-pack profile kept private and stale records out while staying inside the target pollution band.

Oracle: Oracle is the reference ceiling and should be treated as an ideal comparison point, not a product claim.

Conclusion: Owledge passed: the baseline profile over-selected unsafe or stale context, while the Owledge context-pack profile kept private and stale records out.