Sample report. Made by the real Wallras engine from staged traffic: a demo AI answering support questions, which then changes silently. No customer data. Back to Wallras
AI monitoring report
Period: 2026-10-13 13:54 UTC to 2026-11-12 13:54 UTC
Generated: 2026-11-12 13:54 UTC
System
- Provider: unknown-runtime
- Observation mode: hybrid
- Model
drift-support-assistant served as: drift-support-assistant-v1
Data handling and record integrity
- All records are stored only on the operator's own infrastructure.
- Leaves the infrastructure: License key only. No prompts, answers or logs.
- Tamper check: intact (288 records verified)
- Receipt: record #288, fingerprint
2d5d29e43b37ee9f8cab55f4c956e3f8f46927f126f1fc289bf50a07b305d072
- Keep this report outside the server. Checking the records against this receipt later shows whether anything up to this record was changed since.
Activity
- Requests: 201
- Errors: 0 (0.00%)
Drift monitoring
- Overall now: drifting
drift-support-assistant: drifting
- high: Answer length changed sharply
- high: The way the AI words its answers changed a lot
- high: AI is refusing more requests than its baseline
Incidents in period:
- 2026-11-12 13:54 UTC
drift-support-assistant drifting: Answer length changed sharply; The way the AI words its answers changed a lot; AI is refusing more requests than its baseline
Verification
Hidden tests:
- What is 2+2?: 100% passed of 2
Consistency checks: 10, contradiction rate 0.0% (changes with wording 0.0%, random even when asked identically 0.0%)
Accuracy and confidence
- Feedback received: 52
- Accuracy: 76.9%
- Calibration error: 0.088
Blind spots
- account (topic): wrong 31% of the time vs 0% elsewhere
Human oversight
- 2026-11-12 13:53 UTC: Hidden test added (model: drift-support-assistant)
- 2026-11-12 13:53 UTC: Admin action (actor: admin, method: POST, path: /wallras/canaries, status: 200)
- 2026-11-12 13:53 UTC: Admin action (actor: admin, method: POST, path: /wallras/canaries/run, status: 200)
- 2026-11-12 13:53 UTC: Regression set created (check: similar, name: support-golden)
- 2026-11-12 13:53 UTC: Admin action (actor: admin, method: POST, path: /wallras/datasets, status: 200)
- 2026-11-12 13:53 UTC: dataset_items_added (count: 6, name: support-golden)
- 2026-11-12 13:53 UTC: Admin action (actor: admin, method: POST, path: /wallras/datasets/support-golden/items, status: 200)
- 2026-11-12 13:53 UTC: Admin action (actor: admin, method: POST, path: /wallras/datasets/support-golden/run, status: 200)
- 2026-11-12 13:53 UTC: Admin action (actor: admin, method: POST, path: /wallras/canaries/run, status: 200)
- 2026-11-12 13:54 UTC: Admin action (actor: admin, method: POST, path: /wallras/datasets/support-golden/run, status: 200)
- 2026-11-12 13:54 UTC: Admin action (actor: admin, method: POST, path: /wallras/snapshot, status: 200)
Limitations
- Wallras observes behaviour and reports signals for human review. It cannot detect every problem.
- Accuracy and calibration figures cover only answers that received feedback.
- Consistency checks catch unstable answers, not answers that are consistently wrong.
- The tamper check proves the records match receipts kept elsewhere. Without an external receipt, it proves internal consistency only.
- This report documents monitoring evidence. It is not a certification of legal or regulatory compliance.