AI Evaluation, Safety, Cost & Observability
Continuous proof that AI features are useful, safe and affordable: golden-dataset eval runs vs baseline, safety incidents, token cost per use case, latency percentiles, drift alerts, budgets and kill switches per environment/tenant/use case.
AI-08
§12.3
July spend / budget
AED 9,420
of AED 12,000 cap · 78.5% used
Eval gate pass rate
94%
16 of 17 gates this cycle
Safety incidents open
2
1 medium · 1 low
P95 latency (generation)
2.8s
SLO ≤ 4.0s · within SLO
Drift alerts
3
1 requires re-evaluation
Evaluation runs vs baseline — golden datasets
Cycle: Jul 2026 · thresholds versioned in registry
| Run | Use case · dataset | Metric | Baseline | Current | Δ | Threshold | Gate |
|---|---|---|---|---|---|---|---|
| EV-0712 14 Jul · rec-nba@1.4.2 |
Lead scoring · golden-leads-v6 (1,200 cases) |
Ranking AUC | 0.81 | 0.83 | +0.02 | ≥ 0.78 | Pass |
| EV-0711 14 Jul · doc-extract@3.0.1 |
Invoice extraction · golden-extract-inv-v4 |
Field accuracy | 97.5% | 96.8% | −0.7 | ≥ 97.0% | Fail — promotion blocked (AI-02) |
| EV-0710 13 Jul · rag-kb@2.3.0 |
Knowledge RAG · golden-kb-qa-v9 |
Groundedness / citation | 0.90 | 0.92 | +0.02 | ≥ 0.88 | Pass |
| EV-0709 13 Jul · anom-fin@2.1.0 |
Anomaly detection · golden-anom-v3 |
False-positive rate | 10.3% | 8.2% | −2.1 | ≤ 10.0% | Pass |
| EV-0708 12 Jul · all chat surfaces |
Adversarial suite · redteam-inject-v11 (injection/jailbreak) |
Attack block rate | 99.1% | 98.2% | −0.9 | ≥ 99.0% | Conditional — mitigation due 21 Jul |
| EV-0707 11 Jul · val-comp@1.2.0 |
Valuation · golden-comps-v5 vs approved comparables |
MAPE | 7.1% | 6.4% | −0.7 | ≤ 8.0% | Pass |
| EV-0706 10 Jul · doc-extract@3.0.1 |
Arabic KYC extraction · golden-kyc-ar-v2 |
Field accuracy | 94.9% | 95.6% | +0.7 | ≥ 94.0% | Pass |
| EV-0705 08 Jul · rec-nba@1.4.2 |
Lead scoring bias review · protected-attribute proxy scan | Disparity ratio | 1.06 | 1.04 | −0.02 | ≤ 1.10 | Pass — Q3 review scheduled |
8 of 17 gates · failed gate blocks production activation per AI-02
Safety incidents
2 open| Incident | Type | Severity | Status |
|---|---|---|---|
| SI-2026-014 15 Jul 13:47 · copilot |
Prompt injection in uploaded vendor PDF — instruction text quarantined; retrieved text treated as data (AI-03) | High | Contained — kill switch on copilot tools · INC-2026-0231 |
| SI-2026-013 14 Jul 09:12 · transcription |
PII redaction miss — Emirates ID fragment in call summary (masked on detection: 784-19●●-●●●●●●●-●) | Medium | Open — privacy review · privacy workspace |
| SI-2026-012 11 Jul 16:03 · RAG |
Hallucinated citation blocked by citation-check; answer declined and qualified | Low | Open — dataset case added to golden-kb-qa-v10 |
| SI-2026-011 08 Jul 20:40 · tenant portal |
Jailbreak attempt via portal chat — refused; session flagged | Low | Closed — no data exposure · audit trail |
| SI-2026-010 05 Jul 11:26 · extraction |
Invoice total vs line-item conflict — posting blocked, shown for review (AI-04 criterion) | Low | Closed — worked as designed · doc intelligence |
Latency & errors by capability
SLOs met (3 of 4)| Capability | P50 | P95 | P99 | Error rate | SLO |
|---|---|---|---|---|---|
| Generation | 0.9s | 2.8s | 5.1s | 0.6% | Met (≤4.0s P95) |
| Extraction | 1.4s | 3.9s | 7.2s | 1.1% | Met (≤4.5s P95) |
| Embeddings | 0.2s | 0.5s | 0.9s | 0.2% | Met (≤1.0s P95) |
| Tool calls (copilot) | 1.8s | 6.4s | 11.0s | 3.4% | Breached (≤5.0s P95) — killed use case skews sample |
| Route | P95 | Timeouts | Circuit | Fallback used |
|---|---|---|---|---|
| Provider A · region AE · gen | 2.8s | 0.4% | Closed (healthy) | — |
| Provider B · region AE · gen (fallback) | 3.6s | 0.9% | Closed (healthy) | 41 advisory requests on 12 Jul outage |
| Provider A · extraction | 3.9s | 1.2% | Half-open — probing | Manual keying queue offered |
| Provider C · embeddings | 0.5s | 0.1% | Closed (healthy) | — |
Routing, retries and circuit breakers are governed by the AI gateway (AI-01); provider eligibility & DLQ in provider health.
Token cost per use case — July 2026 (prod · Al Manara)
| Use case | Requests | Tokens in/out (M) | Cost (AED) | Budget | State |
|---|---|---|---|---|---|
| Document extraction | 18,204 | 41.2 / 6.8 | 3,120.00 | 4,000.00 | 78% used |
| RAG assistant | 9,861 | 28.7 / 9.1 | 2,040.00 | 3,000.00 | 68% used |
| Recommendations | 22,410 | 18.9 / 4.2 | 1,610.00 | 2,200.00 | 73% used |
| Listing copy | 1,318 | 6.4 / 5.7 | 1,500.00 | 1,500.00 | Cap reached — fallback/defer |
| Anomaly intelligence | 4,092 | 12.1 / 1.9 | 890.00 | 1,300.00 | 68% used |
| Call transcription | 311 | 3.8 / 3.1 | 260.00 | 1,000.00 | 26% used |
| Total | 56,196 | 111.1 / 30.8 | 9,420.00 | 12,000.00 | 78.5% of tenant cap |
Budgets, rate limits & kill switches — per environment / tenant / use case
| Scope | Monthly cap | Spend | Rate limit | Kill switch |
|---|---|---|---|---|
| prod · Al Manara · extraction critical — finance close | AED 4,000 | AED 3,120 | 120 rpm | Armed · off |
| prod · Al Manara · listing copy noncritical | AED 1,500 | AED 1,500 | 30 rpm | Budget stop → fallback gen-lite@0.8 |
| prod · Al Manara · copilot tools | AED 800 | AED 112 | 20 rpm | Active — INC-2026-0231 |
| prod · Abu Dhabi Branch · all AI | AED 2,500 | AED 1,140 | 60 rpm | Armed · off |
| staging · all tenants | AED 600 | AED 415 | 15 rpm | Armed · off |
Drift alerts
1 action required| Alert | Signal | Value / limit | Action |
|---|---|---|---|
| DR-2026-009 16 Jul · lead scoring |
Feature drift — source mix shift after new portal connector | PSI 0.28 / 0.20 | Re-evaluation required |
| DR-2026-008 14 Jul · valuation |
Comparable staleness — 22% of comps older than 180 days | 22% / 15% | Data refresh task → Elena Petrova |
| DR-2026-007 12 Jul · RAG |
Embedding distribution shift after KB reindex | cos-shift 0.11 / 0.15 | Monitor — within limit |
Run log sample — full telemetry per run (privacy-policy scoped)
Inputs/references stored redacted| Run · correlation | Use case | Model · prompt | Conf. | Edited | Approval | Cost | Latency |
|---|---|---|---|---|---|---|---|
run-8842f1corr-7c19e4b0 | Journal draft proposal | anom-fin@2.1.0 · v5 | 0.94 | No | Pending — Daniel Chen (PRP-0459) | AED 0.42 | 2.1s |
run-8839c7corr-51b0aa92 | Invoice extraction | doc-extract@3.0.1 · v12 | 0.99 | 2 fields corrected | Confirmed — Yusuf Karim | AED 0.18 | 3.4s |
run-8835d0corr-9e77c3d1 | RAG answer (KB) | rag-kb@2.3.0 · v7 | 0.88 | No | n/a — advisory, 3 citations returned | AED 0.09 | 1.6s |
run-8831b8corr-c02d1f44 | Listing copy (AR) | gen-lite@0.8 · v3 (fallback) | 0.72 | Edited by Layla Hassan | Listing publish gate pending | AED 0.03 | 0.8s |
run-8828a4corr-33f8be02 | Valuation advice | val-comp@1.2.0 · v4 | 0.58 | No | Reviewer required — tolerance breach (PR-B-3204) | AED 0.11 | 2.9s |
run-8824e9corr-add1907b | Listing copy (EN) | — deferred | — | — | Deferred with explanation — budget cap (noncritical) | AED 0.00 | — |
6 of 56,196 runs · full trail in audit explorer
Static render mockup — no live data. Source: ERP Modern Development Specification v2.0 · Covers: AI-08 (run telemetry: provider/model, prompt, inputs/references, output, confidence, edits, approval, cost, latency, errors under privacy policy; golden-dataset, groundedness/citation, extraction-accuracy, false-positive and adversarial tests; budgets, rate limits and kill switches per environment/tenant/use case; budget-cap fallback/defer behavior).