run_2026-10-07_haiku-5-5
v3.4 · Claude Haiku 5.5 comparison
Claude Haiku 5.5 · bge-m3 · hybrid (BM25 + dense, RRF) · support-prompt v7 on support-golden-v4 v4.6
Stopped at sample 419 of 1,240
Judge quota reached: Claude Opus 5.5 returned 429 after 3 retries at sample 419. Resume continues from the last checkpoint.
Metrics below cover the 418 samples judged before the stop.
Pass rate
82.3%
Faithfulness
0.89
Answer correctness
0.83
Context precision
0.79
Context recall
0.84
Latency p95
2.3s
Failures by category
74 failed samples · largest share: unsupported
| Category | Share | vs reference |
|---|---|---|
| Unsupported claims | 28.4% | +1.2 pts |
| Incorrect factual answers | 23.0% | +2.0 pts |
| Missing context | 17.6% | −1.6 pts |
| Retrieval mismatch | 17.6% | −0.3 pts |
| Instruction-following failures | 9.5% | −1.7 pts |
| Other | 4.1% | +0.4 pts |
Run details
- Dataset
- support-golden-v4 v4.6 · 1,240 items
- Generator
- Claude Haiku 5.5
- Retriever
- bge-m3 · hybrid (BM25 + dense, RRF)
- Prompt
- support-prompt v7
- Judge
- LLM judge · Claude Opus 5.5 · rubric v3
- Started
- Oct 7, 07:40 UTC
- Duration
- 1h 41m
- Triggered by
- Sara Lindqvist
- Tokens / request
- 3,420 in · 198 out
- Compared with
- support-assistant v3.4 · release candidate
- cost
- comparison
All metrics
Against support-assistant v3.4 · release candidate
| Metric | Group | This run | Reference | Change |
|---|---|---|---|---|
| Pass rateShare of samples the judge marked as passing all rubric checks. | Outcome | 82.3% | 86.9% | −4.6 pts |
| FaithfulnessShare of answer claims supported by the retrieved context. | Generation | 0.89 | 0.91 | −0.02 |
| Answer correctnessAgreement with the gold answer on facts, not style. | Generation | 0.83 | 0.88 | −0.05 |
| Instruction adherenceRespect for tone, format, length, language and policy constraints. | Generation | 0.91 | 0.93 | −0.02 |
| Context precisionRank-weighted precision of relevant chunks passed to the generator. | Retrieval | 0.79 | 0.79 | 0.00 |
| Context recallShare of gold-answer statements supported by the retrieved context. | Retrieval | 0.84 | 0.84 | 0.00 |
| Recall@5Share of gold passages found in the top 5 retrieved chunks. | Retrieval | 0.88 | 0.88 | 0.00 |
| MRRMean reciprocal rank of the first relevant chunk. | Retrieval | 0.74 | 0.74 | 0.00 |
| NDCG@10Graded ranking quality of the top 10, normalized by the ideal order. | Retrieval | 0.81 | 0.81 | 0.00 |
| Latency p50Median end-to-end time to final token. | Operational | 1.1s | 1.9s | −0.8s |
| Latency p9595th percentile end-to-end time to final token. | Operational | 2.3s | 3.8s | −1.5s |
| Cost per requestModel, embedding and reranking cost per answered request (USD). | Operational | $0.0013 | $0.0042 | −0.0029 |
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 · 82.3%
Across datasets
v3.4 · Claude Haiku 5.5 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-v4this run | 1,240 | 82.3% | 0.89 | 0.84 | 0.88 | 2.3s | matrix |
| billing-edge-cases | 312 | 77.2% | 0.85 | 0.82 | 0.86 | 2.5s | matrix |
| roaming-and-travel | 188 | 76.6% | 0.86 | 0.79 | 0.83 | 2.4s | matrix |
| device-troubleshooting | 426 | 79.8% | 0.87 | 0.83 | 0.87 | 3.0s | matrix |
| account-security | 154 | 83.8% | 0.90 | 0.85 | 0.89 | 2.2s | matrix |
| locale-formatting | 96 | 75.0% | 0.83 | 0.80 | 0.84 | 2.6s | matrix |
| synthetic-hard-negatives | 540 | 71.9% | 0.84 | 0.73 | 0.77 | 2.5s | matrix |
Traced samples
17 samples with full traces (4 failing, 13 passing), stratified by domain. Open a row to inspect retrieval and the judge's reasoning.
What happens if I go over my fair-use limit on Unlimited at home? Plans & upgrades FailMissingjudge 0.07 |
Does the Fall 2026 upgrade offer work with a family plan discount? Promotions Passjudge 0.95 |
What is the late fee if I pay my bill after the due date? Billing & payments Passjudge 0.98 |
How long does it take to port my number to Harbor? Number porting Passjudge 0.90 |
How do I get a copy of the personal data Harbor holds about me? Data privacy requests Passjudge 0.89 |
My phone shows 4G but mobile data does not work after the iOS update. Device troubleshooting Passjudge 0.97 |
Is SIM swap protection turned on by default? Account & security Passjudge 0.97 |
What is the excess on a theft insurance claim? Returns & warranty Passjudge 0.87 |
My eSIM QR code says it has already been used. What should I do? eSIM & activation Passjudge 0.89 |
My bill says 41,20 €. Is that forty-one euros or four thousand? Billing & payments Passjudge 0.93 |