Skip to content

run_2026-08-28_nightly

Nightly · v3.3 · 2026-08-28

Claude Sonnet 5.5 · bge-m3 · hybrid (BM25 + dense, RRF) · support-prompt v6 on support-golden-v4 v4.6

CompletedNightlyIn productionsupport-assistant v3.3v2 · 512 tokens / 64 overlap
Compare

Stratified slice of support-golden-v4 (430 items).

Pass rate

81.9%

0.0 pts(no change)vs previous nightly352 of 430 samples passed

Faithfulness

0.87

+0.01(improvement)vs previous nightly

Answer correctness

0.83

−0.01(regression)vs previous nightly

Context precision

0.75

+0.02(improvement)vs previous nightly

Context recall

0.78

−0.02(regression)vs previous nightly

Latency p95

3.6s

−0.2s(improvement)vs previous nightly

Failures by category

78 failed samples · largest share: unsupported

Share of failed samples by category, compared with Nightly · v3.3 · 2026-08-27
CategorySharevs reference
Unsupported claims25.6%0.0 pts
Incorrect factual answers20.5%0.0 pts
Missing context21.8%0.0 pts
Retrieval mismatch20.5%0.0 pts
Instruction-following failures9.0%0.0 pts
Other2.6%0.0 pts

Run details

Dataset
support-golden-v4 v4.6 · 430 items
Generator
Claude Sonnet 5.5
Retriever
bge-m3 · hybrid (BM25 + dense, RRF)
Prompt
support-prompt v6
Judge
LLM judge · Claude Opus 5.5 · rubric v3
Started
Aug 28, 02:00 UTC
Duration
38m 13s
Triggered by
Scheduler
Tokens / request
3,980 in · 221 out
  • nightly
  • golden-slice

All metrics

Against Nightly · v3.3 · 2026-08-27

All metrics for Nightly · v3.3 · 2026-08-28
MetricThis runReferenceChange
Pass rate81.9%81.9%0.0 pts
Faithfulness0.870.86+0.01
Answer correctness0.830.84−0.01
Instruction adherence0.890.890.00
Context precision0.750.73+0.02
Context recall0.780.80−0.02
Recall@50.830.830.00
MRR0.700.68+0.02
NDCG@100.760.75+0.01
Latency p501.9s1.8s+0.1s
Latency p953.6s3.8s−0.2s
Cost per request$0.0047$0.0046+0.0001

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 v3.3 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
81.9%
0.870.780.833.6sThis run
billing-edge-cases312
76.0%
0.830.770.813.9sRun · Aug 20
roaming-and-travel188
75.5%
0.840.740.783.8smatrix
device-troubleshooting426
78.9%
0.850.780.824.4smatrix
account-security154
82.5%
0.880.800.843.6smatrix
locale-formatting96
75.0%
0.820.750.794.0smatrix
synthetic-hard-negatives540
70.7%
0.820.680.723.9smatrix

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 · v3.3 · 2026-08-28

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

Roaming & travel

FailMismatchjudge 0.18

Is there a restocking fee if I return an opened phone?

Returns & warranty

FailMissingjudge 0.13

Can I keep my old number if I move from prepaid to a contract plan?

Plans & upgrades

FailIncorrectjudge 0.48

My eSIM stopped working after I reset my phone. Do I need to pay for a new one?

eSIM & activation

FailUnsupportedjudge 0.12

What is the excess on a theft insurance claim?

Returns & warranty

Passjudge 0.90

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

eSIM & activation

Passjudge 0.92

Is SIM swap protection turned on by default?

Account & security

Passjudge 0.90

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

Returns & warranty

Passjudge 0.90

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

Billing & payments

Passjudge 0.94

Why was I charged twice for my September bill?

Billing & payments

Passjudge 0.94