The operations platform for enterprise AI
Panewright shows your operations team which version is serving each request and the tests it passed before it took traffic.
Rollout
support-triagev3.1.4 → v3.2.0
Steps, 3 of 5 donestarted 14:01 by m.kerr
Traffic on v3.2.0
25%
Next step at 50% in 06:12
| Region | p95 | State |
|---|---|---|
| eu-west-2 | 212 ms | Canary |
| eu-central-1 | – | Queued |
| eu-north-1 | – | Queued |
Sample workspace
p95 latency, eu-west-2
212ms
Requests per minute
18.4k
Sample workspace for Aldwerne Mutual, a fictional insurer.
Teams running Panewright in production
Placeholder marks for fictional companies
Inside the console
prod-eu, last hour
Endpoints 6
Requests/min
18.4k
6.2%
p95 latency
212ms
9 ms
Error rate
0.02%
0.01 pt
Tokens/s
2,802
4.1%
GPU pool, eu-west-2
62% of 48 GPUs
| Endpoint | Version | Region | p95 | Tok/s | State |
|---|---|---|---|---|---|
| support-triage | v3.1.4 | eu-west-2 | 212 ms | 1,840 | Serving |
| claims-summary | v2.8.0 | eu-west-2 | 348 ms | 962 | Serving |
| fraud-signals | v1.4.2 | eu-central-1 | – | – | Deploying |
| policy-qa | v0.9.1 | eu-central-1 | – | – | Queued |
| email-draft | v1.1.0 | eu-north-1 | – | – | Paused |
| doc-classify | v4.0.3 | eu-north-1 | – | – | Paused |
fraud-signals v1.4.2 deploying to eu-central-1, started 14:06 by j.osei
claims-summary suite
Evaluation run_418
v2.8.0 Serving
94.8% pass
v2.9.0 Candidate
97.1% pass
2,944 of 3,200 cases
| Suite | v2.8.0 | v2.9.0 | Change |
|---|---|---|---|
| Faithfulness | 96.2% | 97.8% | +1.6 |
| Coverage | 91.0% | 95.4% | +4.4 |
| Refusal accuracy | 98.9% | 99.1% | +0.2 |
| PII leakage | 100% | 100% | 0.0 |
| Tone | 93.7% | 94.9% | +1.2 |
| Latency budget | 97.5% | 96.2% | −1.3 Regressed |
Opens when the run finishes
Access policy
pol_claims v14
Roles bind to data scopes. Edits land in v15 draft
Applies to claims-summarysupport-triagepolicy-qa
| Role | Claims PII | Policy docs | Payments | Transcripts |
|---|---|---|---|---|
| Claims handler | ||||
| Underwriter | ||||
| Fraud analyst | ||||
| Support agent | ||||
| Contractor | ||||
| Auditor |
No changes in the draft
v14 deployed 2026-09-22 09:14 by j.osei, 41,208 requests checked since
All endpoints
Monitoring
p95 latency per interval
Current interval
Monthly budget
41% of $60,000
| Serving endpoint | p95 | Errors | Spend |
|---|---|---|---|
| support-triage | 212 ms | 0.02% | $31 |
| claims-summary | 348 ms | 0.01% | $51 |
Governance
Panewright keeps a trace of every request: who called, the policy version that allowed it, the model that answered and each field it redacted before the model saw it.
How tracing works
Request trace
req_01J9Q3F7ZK4MAllowed
POST/v1/endpoints/claims-summary/invoke2026-09-24 14:02:11.382 UTC
Latency
228ms
Redacted before the model
policyholder.namepayment.iban
Data scopes read
Claims PII, maskedPolicy docsTranscripts, denied
Grounding score
0.94Threshold 0.80
Sample workspace
Traced this hour
18,412
The platform
3,200cases
Test cases in a typical release run, sample
Each candidate runs your test sets before it takes traffic, and the run names every case that regressed.
38s
Median time to roll back a failing canary, sample
Releases take traffic in steps you set, and roll back on their own when latency slips.
7years
Trace retention chosen by one customer, sample
Traces stay searchable by caller, model or policy, and export as signed files for your reviewers.
Deployment targets
Any major cloud, in your own account and the regions you pick. Nothing runs on our side.
Kubernetes 1.28 or later on GPU nodes, installed from one Helm chart.
A signed offline bundle that installs and updates from removable media.
Independent assurance
Invented schemes and marks, shown as a sample
Halvane
Release signing, audited yearly
Tarnholt
Access control review
Vesperline
Data kept in your chosen region
Aurlen
Offline install attestation
From platform teams
Fictional customers. Quotes written for this sample page.
We used to learn about a regressed model from a support ticket. Since we moved releases onto Panewright, a bad candidate stops at evaluation with the broken cases listed.
Release cycle 41 days → 6 days, sample
Our auditors now read the same request trace as our engineers. That ended a quarterly argument about who was allowed to see claims data.
Audit preparation 3 weeks → 2 days, sample
First release
Bring one model and its test set, and a Panewright engineer will see it through evaluation to live traffic.