Skip to content

run_2026-07-27_nightly

Nightly · v3.2 · 2026-07-27

Claude Sonnet 5.5 · bge-m3 · dense · support-prompt v5 on support-golden-v4 v4.6

CompletedNightlysupport-assistant v3.2v2 · 512 tokens / 64 overlap
Compare

Stratified slice of support-golden-v4 (409 items).

Pass rate

78.2%

+0.8 pts(improvement)vs previous nightly320 of 409 samples passed

Faithfulness

0.86

0.00(no change)vs previous nightly

Answer correctness

0.81

−0.01(regression)vs previous nightly

Context precision

0.70

+0.01(improvement)vs previous nightly

Context recall

0.74

0.00(no change)vs previous nightly

Latency p95

3.7s

+0.2s(regression)vs previous nightly

Failures by category

89 failed samples · largest share: unsupported

Share of failed samples by category, compared with Nightly · v3.2 · 2026-07-26
CategorySharevs reference
Unsupported claims23.6%−0.1 pts
Incorrect factual answers19.1%−0.5 pts
Missing context23.6%+0.9 pts
Retrieval mismatch23.6%−0.1 pts
Instruction-following failures6.7%−0.5 pts
Other3.4%+0.3 pts

Run details

Dataset
support-golden-v4 v4.6 · 409 items
Generator
Claude Sonnet 5.5
Retriever
bge-m3 · dense
Prompt
support-prompt v5
Judge
LLM judge · Claude Opus 5.5 · rubric v3
Started
Jul 27, 02:00 UTC
Duration
38m 10s
Triggered by
Scheduler
Tokens / request
3,870 in · 230 out
  • nightly
  • golden-slice

All metrics

Against Nightly · v3.2 · 2026-07-26

All metrics for Nightly · v3.2 · 2026-07-27
MetricThis runReferenceChange
Pass rate78.2%77.4%+0.8 pts
Faithfulness0.860.850.00
Answer correctness0.810.82−0.01
Instruction adherence0.870.86+0.01
Context precision0.700.69+0.01
Context recall0.740.750.00
Recall@50.770.780.00
MRR0.630.65−0.02
NDCG@100.720.710.00
Latency p501.7s1.7s+0.1s
Latency p953.7s3.5s+0.2s
Cost per request$0.0045$0.0043+0.0002

Nightly trend

Nightly runs on the support-golden-v4 slice, Jul 11 to Oct 8. This run is one point in the series.

Pass rate moved from 78.5% to 87.2% across 90 nightly runs. Configuration changes: v3.3 on Aug 4, v3.4 on Sep 8.

Across datasets

support-assistant v3.2 on every dataset in the project. Rows marked “matrix” come from the comparison matrix rather than a dedicated run.

Configuration results per dataset
DatasetItemsPass rateFaith.Ctx recallRecall@5p95Source
support-golden-v4this run1,240
78.2%
0.860.740.773.7sThis run
billing-edge-cases312Not evaluated with this configuration
roaming-and-travel188Not evaluated with this configuration
device-troubleshooting426Not evaluated with this configuration
account-security154Not evaluated with this configuration
locale-formatting96Not evaluated with this configuration
synthetic-hard-negatives540Not evaluated with this configuration

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.

Traced samples for Nightly · v3.2 · 2026-07-27

How much data can I use in Germany on Unlimited Plus before I get charged?

Roaming & travel

FailMismatchjudge 0.18

My phone shows 4G but mobile data does not work after the iOS update.

Device troubleshooting

Passjudge 0.96

How do I get a copy of the personal data Harbor holds about me?

Data privacy requests

Passjudge 0.96

What is the late fee if I pay my bill after the due date?

Billing & payments

Passjudge 0.89

What is the excess on a theft insurance claim?

Returns & warranty

Passjudge 0.95

How long does it take to port my number to Harbor?

Number porting

Passjudge 0.90

Does the Fall 2026 upgrade offer work with a family plan discount?

Promotions

Passjudge 0.95

Can I return a phone I bought online if I opened the box?

Returns & warranty

Passjudge 0.95

My bill says 41,20 €. Is that forty-one euros or four thousand?

Billing & payments

Passjudge 0.91

My eSIM QR code says it has already been used. What should I do?

eSIM & activation

Passjudge 0.86