Public evaluation evidence
Evidence Console
A calm, inspectable view of explicitly approved evaluation artifacts. Metrics retain their dataset, run, and claim boundaries.
Approved artifacts
4
Explicit index membership
Run artifacts
3
Available for inspection
Catalog state
Allowlisted
Public Evidence Contract V1
No universal score
Headline metrics
Each value stays attached to its named run.
Hit Rate@5
1
miracl-th-mini · demo-miracl-th-mini-v1
MRR
1
miracl-th-mini · demo-miracl-th-mini-v1
nDCG@5
1
miracl-th-mini · demo-miracl-th-mini-v1
Hit Rate@5
0.6
retrieval-fixture · demo-retrieval-fixture-v1
MRR
0.6
retrieval-fixture · demo-retrieval-fixture-v1
nDCG@5
0.523
retrieval-fixture · demo-retrieval-fixture-v1
Hit Rate@5
0.8
retrieval-fixture · demo-retrieval-reference-v1
MRR
0.8
retrieval-fixture · demo-retrieval-reference-v1
nDCG@5
0.8
retrieval-fixture · demo-retrieval-reference-v1
Approved evidence
Run artifacts
| Artifact | Dataset | Claim dimensions | Metrics | Inspect |
|---|---|---|---|---|
| demo-miracl-th-mini-v1miracl retrieval | miracl-th-minisynthetic-fixture-v1 | Verification: VERIFIEDData: SYNTHETIC_FIXTUREClaim: INTEGRATION_ONLY | Hit Rate@5 1 MRR 1 nDCG@5 1 | Inspect |
| demo-retrieval-fixture-v1retrieval | retrieval-fixturesynthetic-v1 | Verification: VERIFIEDData: SYNTHETIC_FIXTUREClaim: INTEGRATION_ONLY | Hit Rate@5 0.6 MRR 0.6 nDCG@5 0.523 | Inspect |
| demo-retrieval-reference-v1retrieval | retrieval-fixturesynthetic-v1 | Verification: VERIFIEDData: SYNTHETIC_FIXTUREClaim: INTEGRATION_ONLY | Hit Rate@5 0.8 MRR 0.8 nDCG@5 0.8 | Inspect |
Regression evidence
Regression evidence
Synthetic same-population regression demonstration; not a benchmark or model-superiority result.
Reference
demo-retrieval-reference-v1Candidate
demo-retrieval-fixture-v1Metrics
4 regression · 1 pass · MATCHEDInterpretation boundary
These artifacts are safe public evidence, not a universal model-quality ranking. Synthetic fixtures support integration claims only. A benchmark-shaped fixture is labeled as synthetic and must not be read as an official benchmark result.
Provenance and limitations
- • The index is an explicit allowlist; no directory scan is used.
- • Metrics preserve their k and run context.
- • Detail pages expose only contract-approved metadata and evidence.