ACI Inference matched a simple linear control on shared features while keeping tenant isolation, rollback, and tenant delete visible on the live cloud surface.
0.875
ACI task score
20 Newsgroups 4-way
0.0
Gap vs control
0.875 vs 0.875
1.0
Tenant isolation match
Tenant A unchanged after tenant B update
1.0
Rollback match
Probe update then rollback
April 2026
The ACI Inference product claim is not that it beats every model or every training stack. The claim is narrower and more useful: one shared model can support per-tenant adaptation with explicit update, inference, rollback, audit, and delete operations while holding structured-task quality at a practical level.
This pack measures that contract directly on a public text task. It uses a shared feature model, creates two tenants with memory off, compares tenant A against a simple linear control on the same features, then measures isolation, rollback, and delete on the live API surface.
The public structured-task benchmark matched a simple linear control on the same feature space instead of trading away task quality for lifecycle features.
A second tenant was created on the same shared model with remapped examples, and tenant A's predictions stayed unchanged across the full evaluation set.
Probe items were added and then rolled back on tenant A with a match rate of 1.0, and the second tenant was deleted through the control plane.
The default cloud profile kept memory off, stayed in dual mode, and held support count at 96 on the shared-service tenant path.
On this public structured-task benchmark, ACI Inference matched the simple linear control at 0.875 while keeping tenant lifecycle operations explicit on the live surface. The point is not a universal accuracy win. The point is accuracy parity plus isolation, rollback, and delete on one shared service.
| Measure | ACI Inference | Simple control | Operational meaning |
|---|---|---|---|
| Structured-task accuracy | 0.875 | 0.875 | Matched a simple linear control on the same shared features. |
| Mean update latency | 15.748 ms | — | Measured per-item tenant updates on the live inference surface. |
| Mean infer latency | 11.161 ms | — | Includes request-time feature extraction plus analytic tenant state in the public run. |
| Rollback latency | 5.361 ms | — | Measured probe removal on tenant A without turning deletion into retraining. |
| Tenant isolation match | 1.0 | — | Tenant A predictions remained unchanged after tenant B updates. |
| Rollback match | 1.0 | — | Tenant A predictions returned exactly to baseline after probe rollback. |
| Tenant delete latency | 1.79 ms | — | Explicit tenant lifecycle measured on the live control plane. |
| Memory items | 0 | — | The cloud default stayed memory-off throughout the public run. |
What the result says
These results put ACI Inference on a stronger public footing. The product can now carry a structured-task cloud story with a strong frozen shared model, memory still off, accuracy parity to a simple control, and explicit tenant isolation, rollback, and delete behavior on the live service surface.
The result matters anywhere a shared model has to serve distinct tenants without hiding lifecycle behavior inside full retraining cycles or per-customer model copies.
The result fits serving stacks that need one shared model, isolated tenant state, and explicit update, rollback, audit, and delete operations.
The benchmark is relevant where domain updates, rollback, and customer offboarding must remain visible system operations rather than background retraining work.
The pack matters where isolation, deletion, and signed operational evidence have to coexist with acceptable structured-task quality on a strong public shared model.
Explore where ACI Inference fits in model-serving platforms, enterprise AI systems, and multi-tenant cloud deployments.