Stability, Plasticity, and
Editability in ACI
ACI delivers three capabilities together: strong continual-learning accuracy that nearly doubles baselines, near-exact editability for precise deletion and rollback, and a governed hybrid edge runtime with over 12x lower forgetting than standard approaches.
0.4750 +/- 0.0248
Streaming text final accuracy
1.76x the strongest baseline
1.7e-9
Editability error
Near-exact item removal
46.75 vs 572.51
Continual forgetting
12x lower than baseline
The capability contract
ACI guarantees three properties at every deployment surface. Together, they make post-deployment change a governed operation rather than uncontrolled drift.
Stability
When the system learns something new, previously protected outputs remain within a declared non-regression budget. Production behavior stays bounded.
Plasticity
New items are incorporated in bounded time and bounded memory, without requiring a full retraining cycle. The system adapts in place.
Editability
Individual learned contributions can be precisely removed and the resulting state reconstructed, enabling deletion, rollback, and subject-level data removal.
Supervised Evidence
Supervised continual-learning results
Across streaming text and domain-incremental benchmarks, ACI demonstrates materially stronger final accuracy, over 20x lower forgetting, and near-zero editability error.
Streaming text final accuracy
ACI 0.4750 +/- 0.0248 (best), 0.4663 +/- 0.0228 (second)
Best non-ACI baselines: 0.2625 -- 0.2700.
ACI nearly doubles the final accuracy of the strongest non-ACI baseline on continual streaming-text learning.
Forgetting
ACI 0.0087 +/- 0.0049 (best), 0.0100 +/- 0.0049 (second)
Baseline forgetting: 0.2000+.
Over 20x lower forgetting than baselines. Protected behavior stays bounded as the system continues to learn.
Editability
ACI ~1.7e-9
Replay baseline: about 0.8.
Near-exact removal of individual learned contributions, enabling precise deletion and rollback at the item level.
Domain-incremental accuracy
ACI 0.2626 +/- 0.0019
Best baseline (replay): 0.1881 +/- 0.0073.
ACI outperforms the strongest baseline by 40% on DomainNet, demonstrating effective knowledge binding under structured domain shift.
Edge Runtime
Edge and robotics results
On continuous-control benchmarks, ACI achieves dramatically lower forgetting than standard RL and provides governed online adaptation with explicit rollback, editability, and typed safety constraints at the edge.
12x lower forgetting
On a 20-task continual-control suite, ACI achieves forgetting of 46.75 versus 572.51 for the established baseline. On a 10-task suite, 54.56 versus 407.43. The system retains what it learned while continuing to adapt.
Hybrid edge runtime
ACI operates as a governed online layer: it handles adaptation, editability, typed constraints, and rollback on the edge while deep RL handles raw policy optimization.
Governed adaptation
Every change at the edge goes through explicit bind, constrain, and audit operations. The system adapts within declared safety bounds, not unchecked gradient updates.
Three products plus the add-on
ACI Inference — multi-tenant cloud adaptation with per-tenant binding, isolation, and deletion.
ACI Personal Agents — desktop, laptop, and on-device agent product with local memory, reset, and erase at the user level.
ACI Edge Runtime — compiled edge runtime for bounded local adaptation, rollback, and typed control.
ACI Safety & Policy — typed constraints, hard denial boundaries, and auditable evidence when those controls are part of the deployment boundary.
Read the full benchmark report
The full report includes the capability contract definitions, complete benchmark tables with baselines, the edge and robotics analysis, and product interpretation across the three products plus the safety and policy add-on.
Results shown are from the current benchmark artifacts. See the full report for methodology, baselines, and scope.