run_2026-10-08_synthetic-hn
v3.4 · synthetic-hard-negatives v1.3
Claude Sonnet 5.5 · bge-m3 · hybrid (BM25 + dense, RRF) · support-prompt v7 on synthetic-hard-negatives v1.3
Judging samples · 340 of 540
63% complete · about 50m remainingMetrics update as results arrive. Partial figures are provisional until the run completes.
Pass rate
76.5%
Faithfulness
0.86
Answer correctness
0.83
Context precision
0.68
Context recall
0.73
Latency p95
4.0s
Failures by category
80 failed samples · largest share: unsupported
| Category | Share |
|---|---|
| Unsupported claims | 27.5% |
| Incorrect factual answers | 21.3% |
| Missing context | 18.8% |
| Retrieval mismatch | 17.5% |
| Instruction-following failures | 11.3% |
| Other | 3.8% |
Run details
- Dataset
- synthetic-hard-negatives v1.3 · 540 items
- Generator
- Claude Sonnet 5.5
- Retriever
- bge-m3 · hybrid (BM25 + dense, RRF)
- Prompt
- support-prompt v7
- Judge
- LLM judge · Claude Opus 5.5 · rubric v3
- Started
- Oct 8, 08:05 UTC
- Duration
- In progress
- Triggered by
- Sara Lindqvist
- Tokens / request
- 3,420 in · 214 out
- adversarial
- ranking
All metrics
Retrieval, generation and operational metrics
| Metric | Group | This run |
|---|---|---|
| Pass rateShare of samples the judge marked as passing all rubric checks. | Outcome | 76.5% |
| FaithfulnessShare of answer claims supported by the retrieved context. | Generation | 0.86 |
| Answer correctnessAgreement with the gold answer on facts, not style. | Generation | 0.83 |
| Instruction adherenceRespect for tone, format, length, language and policy constraints. | Generation | 0.91 |
| Context precisionRank-weighted precision of relevant chunks passed to the generator. | Retrieval | 0.68 |
| Context recallShare of gold-answer statements supported by the retrieved context. | Retrieval | 0.73 |
| Recall@5Share of gold passages found in the top 5 retrieved chunks. | Retrieval | 0.77 |
| MRRMean reciprocal rank of the first relevant chunk. | Retrieval | 0.63 |
| NDCG@10Graded ranking quality of the top 10, normalized by the ideal order. | Retrieval | 0.70 |
| Latency p50Median end-to-end time to final token. | Operational | 2.0s |
| Latency p9595th percentile end-to-end time to final token. | Operational | 4.0s |
| Cost per requestModel, embedding and reranking cost per answered request (USD). | Operational | $0.0043 |
Nightly trend
Nightly runs on the support-golden-v4 slice, Jul 11 to Oct 8. Dashed line marks this run.
- Pass rate · nightly
- This run · 76.5%
Across datasets
support-assistant v3.4 on every dataset in the project. Rows marked “matrix” come from the comparison matrix rather than a dedicated run.
| Dataset | Items | Pass rate | Faith. | Ctx recall | Recall@5 | p95 | Source |
|---|---|---|---|---|---|---|---|
| support-golden-v4 | 1,240 | 87.2% | 0.92 | 0.85 | 0.88 | 4.3s | Run · Oct 8 |
| billing-edge-cases | 312 | 81.7% | 0.87 | 0.82 | 0.86 | 4.0s | Run · Oct 2 |
| roaming-and-travel | 188 | 80.9% | 0.88 | 0.79 | 0.83 | 3.9s | Run · Sep 28 |
| device-troubleshooting | 426 | 84.5% | 0.89 | 0.83 | 0.87 | 4.5s | Run · Sep 30 |
| account-security | 154 | 88.3% | 0.92 | 0.85 | 0.89 | 3.7s | Run · Sep 25 |
| locale-formatting | 96 | 80.2% | 0.86 | 0.80 | 0.84 | 4.1s | Run · Oct 4 |
| synthetic-hard-negativesthis run | 540 | 76.5% | 0.86 | 0.73 | 0.77 | 4.0s | matrix |
Traced samples
4 of 5 traced samples judged so far. Open a row to see retrieval and the judge's reasoning.
How much roaming data is included on Unlimited (not Unlimited Plus) in the EU? Roaming & travel Passjudge 0.97 |
Can I use my Unlimited hotspot allowance abroad? Roaming & travel FailMismatchjudge 0.16 |
Can I combine the refer-a-friend credit with the Fall 2026 upgrade offer? Promotions Passjudge 0.98 |
Which plan includes Disney+? Plans & upgrades FailMissingjudge 0.05 |
4 rows