Announcement
Introducing ACI: What Post-Training Learning Looks Like
Most AI teams do not have a pretraining problem. They have a post-deployment change problem. New tenants, device fleets, jurisdictions, policies, and local behaviors all require the model to change after launch. The default answer today is more gradient work, more adapters, more retrieval infrastructure, and more operational risk.
ACI addresses that problem directly. Train or choose a strong backbone once, then handle frequent learning, editing, personalization, and rollback through explicit post-deployment operations: bind, adapt, constrain, unbind, and audit.
What ACI is
ACI has four core capabilities. It is a post-training learning engine, an exact unlearning engine, a governed personalization engine, and a hard-constraint safety engine. These describe what the system does after deployment: lower adaptation overhead, exact or certificate-bounded removal of specific learned contributions, per-tenant and per-device personalization without full model clones, and explicit constraints that live in the model itself rather than in prompts or sidecar middleware.
Stability, plasticity, and editability
Stability means protected outputs stay inside a declared non-regression budget when the system changes. This guarantee is defined on named protected evaluation sets. In exact mode, retained constraints hold up to floating-point error. After lossy consolidation, the guarantee becomes certificate-bounded.
Plasticity means new admissible items can be incorporated in bounded time and bounded memory. A tenant fact, a device habit, or a new policy item should not require a full retraining cycle.
Editability means specific learned contributions can be removed and the system can reconstruct the counterfactual retained state. If a tenant workflow, user preference, or subject-linked item must disappear, unbind returns the retained state to what would have existed had that item never been ingested.
Plasticity accepts new items. Stability keeps protected behavior from drifting. Editability removes items later and reconstructs the correct counterfactual state.
The three products plus the add-on
ACI Inference is the multi-tenant cloud and on-prem product: one shared service model, isolated tenant state, and operational surfaces for production serving.
ACI Personal Agents is the desktop, laptop, and on-device product: private personalization, local persistence, local erase, and explicit privacy reporting for personal agents and device software.
ACI Edge Runtime is the compiled edge surface: local adaptation under hard resource bounds for robotics, industrial, medical, defense, and automotive programs.
ACI Safety & Policy is the constraint and evidence add-on: typed constraints with deterministic precedence, hard deny and projection behavior, and signed evidence.
Where ACI fits relative to RAG and adapter stacks
Retrieval does not change the prediction function. Periodic retraining changes it, but does not provide exact removal of specific learned contributions. ACI changes the prediction function while also providing exact removal in exact mode or certificate-bounded removal in certified mode.
For safety: vector databases and adapter stacks do not preserve protected prior constraints by construction. In ACI, those constraints live in the retained control layer, which is why policy is part of the deployment surface rather than a late-stage wrapper.
What the benchmark data supports
The benchmark section establishes three points. First, supervised ACI performs strongly on reported benchmark artifacts. Second, pure analytic adaptation from scratch does not match mature baseline policies on hard continuous control. Third, ACI still delivers valuable stability and editability properties in robotics, which is why the edge position is hybrid refinement and safety rather than replacing established policies.
If the problem starts after deployment rather than before it, that is where ACI fits — a layer for post-training change, not a replacement for pretraining.