# ERPipe Jev × Odoo MCP synthetic pilot evidence

Evidence date: 2026-09-30
Article: https://erpipe.com/guides/jev-model-odoo-mcp

## Environment and method

Model: jev-1.13.0. Current ERPipe adapter with real TypeSafe API calls, synthetic Odoo transport and constructed memory candidates. 120 discovery and 80 memory cases derive from 30 discovery bases and eight memory bases through alternate phrasing. Labels are developer-authored; no human-reviewed quality holdout.

## Aggregate result

| Metric | Ordinary path | Rules/aliases | Rules/aliases + optional Jev |
|---|---:|---:|---:|
| Expected first discovery model | 39/120 | 63/120 | 114/120 |
| Expected supplied memory in top 3 | 41/80 | Not a separate branch | 72/80 |

- Provider requests: 154; validated responses: 154.
- Reported input tokens: 131773.
- Estimated Jev input cost: USD 0.00553447, using USD 0.042 per million input tokens. This is not an invoice or full-stack cost.
- Recorded adapter-attempt interval p50: 291 ms; p95: 441 ms.

## Limits

The timing interval includes local health/accounting admission and network/response handling before usage settlement. It is not pure model inference time or complete-task latency.

- The 200 phrased cases are not 200 independent business intents.
- Memory order and vector scores were constructed; this is not measured production retrieval quality.
- No production Odoo timings, real embedding recall, client LLM tokens or complete-task p95 were measured.
- Tenant/connection/expiry enforcement was checked separately with SQLite-backed tests.
- Production deployment and feature enrollment do not prove an organic authenticated inference call or improved customer-task outcomes.

This public packet contains aggregate evidence only. It does not contain credentials, customer data, request payloads or original internal ledger rows.