Architecture pattern catalog

Pattern 1: deterministic cascade

Shape: inexpensive deterministic rules or small model first; expensive fallback only on uncertainty or policy trigger.

Use when: easy cases dominate, confidence is calibratable, and fallback latency is acceptable.

Failure modes: false confidence prevents escalation, sequential latency accumulates, and first-stage bias contaminates later stages.

Pattern 2: learned router with independent specialists

Shape: a router chooses one or more behaviorally interchangeable specialist packages.

Use when: workloads contain stable niches with different cost or capability profiles.

Failure modes: router drift, starvation, correlated training data, and hidden differences in contracts.

Pattern 3: parallel independent ensemble

Shape: multiple models produce candidates independently; a deterministic or separately governed judge aggregates.

Use when: error cost is high and diversity provides measurable benefit.

Failure modes: correlated mistakes, judge bias, high cost, and latency tails.

Pattern 4: shared base with adapters

Shape: one frozen base remains loaded while small adapters provide domains or skills.

Use when: skills share a compatible foundation and memory is constrained.

Failure modes: base dependency, adapter interference, tokenizer lock-in, and difficult cross-base migration.

Pattern 5: offline champion–challenger factory

Shape: telemetry and curated failures feed an isolated candidate-generation pipeline; candidates enter ordinary release controls.

Use when: governance and reproducibility matter more than minute-by-minute structural adaptation.

Failure modes: stale datasets, slow adaptation, benchmark overfitting, and insufficient production feedback.

Pattern 6: quality-diversity archive

Shape: candidates compete within defined behavioral niches; the system retains the best artifact in each occupied niche.

Use when: future task diversity matters and a single scalar winner would collapse useful variation.

Failure modes: poorly chosen descriptors, archive bloat, expensive evaluation, and novelty without utility.

Pattern 7: federated adapter network

Shape: sites train bounded local updates against a shared base; a protected aggregator creates a global challenger.

Use when: raw data must remain local and participants can satisfy a common round contract.

Failure modes: non-IID data, client poisoning, stragglers, privacy leakage, and local regressions hidden by global means.

Pattern 8: edge–cloud hybrid

Shape: local models handle privacy- or latency-sensitive work, escalating minimized requests to a cloud specialist when permitted.

Use when: devices can support core capabilities but need a broader fallback.

Failure modes: policy mistakes during transfer, offline degradation, dual-version complexity, and unpredictable fallback cost.

Pattern 9: model-merge laboratory

Shape: compatible parents are recombined offline across a bounded coefficient or data-flow search space, then compared with both parents and an output ensemble.

Use when: parent compatibility is strong and artifact consolidation has a concrete value.

Failure modes: parameter interference, hidden regressions, invalid licenses, and mistaking loadability for compatibility.

Pattern 10: teleodynamic structural controller

Shape: fast runtime adaptation is separated from a slow controller that proposes add, split, merge, compress, retire, or no-op actions under a resource ledger.

Use when: the organization already has mature lineage, evaluation, rollback, and observability.

Failure modes: feedback loops, metric gaming, oscillation, runaway complexity, and control-plane compromise.

Composition guidance

Combine patterns only to solve measured constraints. A router over an edge–cloud cascade can be reasonable; a router over an ensemble of merged models with online structural mutation is usually an evidence and operations problem before it is an intelligence advantage.