Dual-time-scale adaptive architecture for reliable online learning

1 . A computing system comprising one or more processors and one or more non-transitory computer-readable storage media storing instructions that, when executed by the one or more processors, cause the computing system to:

receive a request and determine a routing certificate by applying a scope function to at least one of the request and a context associated with the request, wherein the routing certificate is a discrete identifier representing a context partition;

generate an output using a machine learning model comprising base parameters and adapter parameters, wherein an adapter state keyed to the routing certificate is applied as an additive low-rank delta to at least one transformation of the machine learning model, and wherein adapter states keyed to routing certificates other than the determined routing certificate are not applied to the at least one transformation during generation of the output;

receive a supervision signal associated with the request and the output, and compute a candidate update to the adapter state keyed to the routing certificate;

evaluate a gating policy to determine whether to authorize committing the candidate update, wherein the gating policy authorizes committing the candidate update only when each of the following conditions is satisfied: (i) a confidence measure meets a confidence threshold, (ii) a risk measure satisfies a risk threshold, (iii) a scope eligibility check confirms that the routing certificate matches the adapter state and is eligible for update, and (iv) a budget check confirms that a cumulative drift budget for the adapter state has not been exceeded;

when the gating policy does not authorize committing the candidate update, refrain from modifying the adapter state;

when the gating policy authorizes committing the candidate update, apply a projection operator to the candidate update to produce a constrained update, commit the constrained update to the adapter state, and increment a version index associated with the adapter state;

append, to a versioned transaction log, a transaction record comprising at least the routing certificate, a delta update corresponding to the constrained update, provenance metadata comprising a timestamp, a supervision source identifier, the confidence measure, the risk measure, and a context hash, and the version index; and

responsive to a rollback instruction identifying a target version, restore the adapter state to the target version by computing a prefix sum of delta updates from a baseline adapter state through the target version using the versioned transaction log, without modifying the base parameters of the machine learning model.

2 . The computing system of claim 1 , wherein the additive low-rank delta for each adapted transformation W having dimensions d_out by d_in is parameterized as a product of a first factor matrix A having dimensions d_out by r and a second factor matrix B having dimensions r by d_in, where r is less than each of d_out and d_in.

3 . The computing system of claim 1 , wherein applying the adapter state during generation of the output comprises computing, for each adapted transformation W, a sum of a product of W and an input vector u and a product of A, B, and the input vector u, without materializing an adapted weight matrix.

4 . The computing system of claim 1 , wherein the scope function is implemented as one of: a deterministic rules engine with explicit versioning, a classifier model producing discrete labels, or a hybrid system in which rules override classifier outputs, and wherein the instructions further cause the computing system to log at least inputs to the scope function and the routing certificate output for each request.

5 . The computing system of claim 1 , wherein the adapter state is keyed by a composite key comprising the routing certificate and at least one additional identifier selected from: a tenant identifier, a user identifier, an application identifier, and a session identifier.

6 . The computing system of claim 1 , wherein the confidence measure is derived from at least one of: a calibrated probability of the output, an agreement metric between the output and a verification module outcome, a consistency metric across multiple stochastic forward passes of the machine learning model, and a trust level associated with the supervision source.

7 . The computing system of claim 1 , wherein the risk measure is derived from at least one of: a safety classifier applied to the request and the candidate update, a policy engine evaluation, and a certificate-based default risk threshold associated with the routing certificate.

8 . The computing system of claim 1 , wherein the budget check evaluates at least one of: a maximum number of committed updates within a session for the routing certificate, a maximum cumulative norm of committed delta updates, a maximum cumulative deviation from the baseline adapter state, and a maximum per-layer delta magnitude.

9 . The computing system of claim 1 , wherein the instructions further cause the computing system to, when the gating policy does not authorize committing the candidate update, log a denied-update record comprising at least a summary of the candidate update, the confidence measure, the risk measure, a scope eligibility result, a budget check result, and one or more reason codes for denial.

10 . The computing system of claim 1 , wherein restoring the adapter state to the target version further comprises:

loading a stored snapshot of the adapter state at a snapshot version that is at or prior to the target version; and

applying delta updates from the versioned transaction log from the snapshot version through the target version to reconstruct the adapter state at the target version.

11 . A computer-implemented method comprising:

receiving, by a computing system comprising one or more processors operating a machine learning model having base parameters and adapter parameters applied as additive low-rank deltas, a candidate delta update to the adapter parameters, wherein the candidate delta update is computed from a supervision signal received during runtime operation of the machine learning model;

evaluating a gating policy that authorizes committing the candidate delta update only when: (i) a confidence measure satisfies a configured confidence threshold, (ii) a risk measure satisfies a configured risk threshold, (iii) a scope eligibility check associated with a routing certificate confirms update eligibility, and (iv) a budget check confirms that cumulative drift of the adapter parameters from a session baseline remains within a configured limit;

when the gating policy authorizes committing the candidate delta update, applying a projection operator to the candidate delta update to produce a projected delta update, wherein the projection operator enforces at least one constraint selected from:

a norm constraint that limits a norm of the additive low-rank delta for each adapted transformation to at most a configured norm bound,

a structured sparsity constraint that restricts the candidate delta update to a subset of layers or parameter blocks specified in a configuration associated with the routing certificate, and

a trust region constraint that limits deviation of the adapter parameters from the session baseline to at most a configured trust region bound;

committing the projected delta update to the adapter parameters without modifying the base parameters; and

appending to a versioned transaction log a transaction record comprising the routing certificate, the projected delta update, provenance metadata comprising at least a timestamp, a supervision source identifier, the confidence measure, and the risk measure, and a version index, wherein the versioned transaction log and a baseline adapter state together enable deterministic restoration of the adapter parameters to any previously committed version by computing a prefix sum of delta updates from the baseline adapter state.

12 . The computer-implemented method of claim 11 , wherein the norm constraint is enforced by computing a Frobenius norm of the additive low-rank delta for each adapted transformation and, when the Frobenius norm exceeds the configured norm bound, rescaling factor matrices of the additive low-rank delta by a ratio of the configured norm bound to the computed Frobenius norm.

13 . The computer-implemented method of claim 11 , wherein applying the projection operator comprises applying the trust region constraint prior to applying the norm constraint.

14 . The computer-implemented method of claim 11 , wherein the configured confidence threshold, the configured risk threshold, the configured norm bound, and the configured trust region bound are each configurable per routing certificate.