Phase 4 Pattern Boundary Matrix

Run: phase4-pattern-page-scaleout-20260711
Contract: 0.1.0
Analysis snapshot: corpus-complete-20260711
Status: reviewer aid; not a library-v1 declaration

How to Read This Matrix

This is a sparse pairwise matrix across all 23 qualifying pattern pages. A row exists for every pair judged materially adjacent after reading both pages. Any pair absent from the matrix is classified not materially adjacent: the patterns may coexist in a system, but the corpus does not present a meaningful boundary ambiguity between their core responsibilities. The matrix records the distinction for review; it does not mechanically adjudicate whether the boundary is correct.

Relationship terms:

Pairwise Boundary Inventory

Pattern A Pattern B Distinct responsibilities Overlap or coexistence Deciding signal Relationship Review status
tool-use-orchestration prompt-chaining Runtime selection/execution of external capabilities vs a fixed sequence of model transformations. A chain may include tool calls. Does the model choose an external capability, or does application code advance to the next known model step? contains/invokes Clear
tool-use-orchestration multi-agent-orchestration Capability invocation vs coordination among autonomous or specialized agent roles. Agents may be exposed as tools. Does the callee own an agent role/state/loop, or merely execute a bounded capability? contains/invokes Clear
tool-use-orchestration safety-policy-enforcement Execute a selected capability vs decide whether the identity/action is allowed. Policy wraps tool discovery and execution. Is the mechanism performing work or authorizing it? coexists Clear
tool-use-orchestration structured-output-contract Tool choice/execution vs machine-readable shape of arguments/results. Tools commonly require typed contracts. Is the failure wrong capability/lifecycle or malformed structure? coexists Clear
eval-driven-development llm-as-judge Full dataset/criteria/release/monitoring lifecycle vs one model-based evaluator implementation. Judges can score eval cases and production traces. Is the decision about the evaluation program or about who/what scores a semantic criterion? contains/invokes Clear
eval-driven-development observability-tracing Decide whether behavior meets criteria vs record what happened. Traces become eval cases; eval results annotate traces. Is the artifact telemetry or a scored judgment? feeds Clear
eval-driven-development data-quality-loop Measure behavior and gate changes vs update data/labels/knowledge from observed failures. Eval failures seed the loop; updated data changes future evals. Is the system judging a version or promoting a correction to its evidence base? feeds Clear
eval-driven-development prompt-versioning Acceptance evidence and release gates vs artifact identity/history/rollback. Every prompt version should link to eval results. Is the concern whether a version is good or which version ran? coexists Clear
eval-driven-development human-in-the-loop-gate Offline/online quality measurement vs accountable intervention in a live or release decision. Humans label evals and may gate releases. Is a human producing evaluation evidence or controlling an action? coexists Clear
human-in-the-loop-gate safety-policy-enforcement Accountable human decision point vs deterministic authorization/business rules. Policy may require human approval. Can software decide from explicit policy, or is accountable judgment required? contains/invokes Clear
human-in-the-loop-gate fallback-degradation Human approval/review vs reduced service after failure. Human handoff is one fallback. Did risk/uncertainty trigger a control gate, or did a failed path require continuity? contains/invokes Clear
multi-agent-orchestration prompt-chaining Agent-role delegation and coordination vs fixed model stages. A multi-agent workflow may contain chains. Do stages possess independent role/state/choice, or only transform the prior output? contains/invokes Clear
multi-agent-orchestration context-engineering Coordinate roles vs construct and isolate each role's information surface. Sub-agent isolation is a context technique. Is the design question who acts, or what each actor can see? coexists Clear
multi-agent-orchestration durable-execution Agent topology vs persistence/recovery of workflow progress. Durable workflows may coordinate multiple agents. Would the boundary remain if all agents completed synchronously without failure? coexists Clear
rag-retrieval agent-memory Retrieve external/domain evidence vs persist and recall interaction or learned state. Memory may use retrieval infrastructure. Is the recalled item durable agent/user history or external knowledge selected for this task? coexists Clear
rag-retrieval context-engineering Acquire relevant external evidence vs assemble all instructions, tools, state, and evidence for a call. Retrieval is one context source. Is the decision how to find evidence or how to construct the complete information surface? contains/invokes Clear
rag-retrieval fine-tuning-adaptation Supply current knowledge at runtime vs change model parameters. Both can add domain capability. Does the knowledge need source-level updates/audit, or is stable behavior being learned into weights? alternative Clear
rag-retrieval data-quality-loop Retrieve from a store vs maintain and correct the store over time. Retrieval failures trigger data cleanup. Is the mechanism serving evidence or changing its quality/freshness? feeds Clear
safety-policy-enforcement guardrails-io-filtering Authorize identities/actions/workflows vs inspect or transform boundary payloads. Both can block execution. Is the decision about permission/policy or about content crossing a boundary? coexists Clear
safety-policy-enforcement privacy-preserving-processing Decide what is allowed vs minimize exposure, movement, retention, or reuse of data. Privacy requirements often become policy rules. Is the architecture enforcing an action rule or changing the data path itself? coexists Clear
safety-policy-enforcement structured-output-contract Decide whether action/content is permitted vs enforce parseable shape. A policy engine consumes structured arguments. Can a perfectly valid schema still represent a forbidden action? If yes, policy is distinct. coexists Clear
guardrails-io-filtering structured-output-contract Check acceptable content vs check required structure/types. One validator may perform both. Is rejection caused by unsafe/disallowed meaning or malformed shape? coexists Clear
guardrails-io-filtering privacy-preserving-processing Detect/transform payload content vs design where sensitive data can flow and persist. Redaction/anonymization can be a filter. Does the control act on one boundary payload or govern the end-to-end data lifecycle? contains/invokes Clear
agent-memory context-engineering Persist and retrieve state vs decide what information enters the current call. Memory is a context source. Is the question retention/recall or current-call selection/assembly? contains/invokes Clear
agent-memory context-compaction Durable retention and recall vs reduce active accumulated context. Compaction may summarize memory. Is information being stored for later or removed/reduced from the active window? coexists Clear
agent-memory durable-execution Model-visible remembered state vs workflow progress and recovery state. Durable state may reference memory artifacts. Must the model recall it, or must the workflow resume it? coexists Clear
context-engineering context-compaction Entire selection/assembly discipline vs one technique for reducing accumulated state. Compaction is part of context engineering. Is the change deciding what context belongs or specifically compressing/removing existing context? contains/invokes Clear
context-engineering prompt-versioning Choose and assemble information for a call vs govern identity/history of prompt artifacts. Versioned prompts are context inputs. Is the concern composition at runtime or lifecycle of an instruction artifact? contains/invokes Clear
context-engineering fine-tuning-adaptation Supply behavior and knowledge at inference vs encode stable behavior in parameters. Both shape model behavior. Does the requirement need quick visible updates or stable learned adaptation? alternative Clear
context-compaction latency-optimization Reduce active state while preserving needed information vs reduce end-to-end elapsed/perceived time. Smaller context can lower inference latency. Is information loss/retention the central tradeoff, or critical-path time? specializes Clear
durable-execution observability-tracing Persist progress/retries across failure vs explain semantic execution behavior. Workflow history is one telemetry source. Would the artifact be used to resume work or diagnose why it behaved that way? feeds Clear
durable-execution fallback-degradation Resume/retry progress vs choose a reduced-capability path after failure. Durable workflows execute fallbacks. Is the response preserving state or changing the service promise? contains/invokes Clear
durable-execution prompt-chaining Persist lifecycle across waits/failures vs define fixed model-stage order. A durable workflow can host a chain. If the process is short and retryable end-to-end, is durability still needed? contains/invokes Clear
fallback-degradation model-routing React to failure/unmet service conditions vs select the preferred normal path. Both may switch models. Did the route change because of failure/degradation or ordinary task fit? specializes Clear
fallback-degradation cost-aware-routing Preserve reduced service after failure vs optimize ordinary unit economics. A cost ceiling can trigger degradation. Is the system choosing its preferred affordable path or recovering from an unavailable/unacceptable one? coexists Clear
fallback-degradation latency-optimization Change capability after failure/timeout vs improve the preferred path's speed. A timeout may trigger fallback. Is slower behavior being optimized or abandoned for a reduced path? coexists Clear
observability-tracing llm-as-judge Capture execution evidence vs apply semantic criteria to it. Judges consume traces and attach scores. Is the output an event history or a quality judgment? feeds Clear
model-routing cost-aware-routing General task/capability/model selection vs economics-led specialization. Cost is one routing dimension. Would the route remain if model prices were equal? specializes Clear
model-routing latency-optimization Choose a model vs optimize the full critical path. Faster-model selection is one latency technique. Is the intervention model choice or concurrency/caching/streaming/workflow design? contains/invokes Clear
model-routing fine-tuning-adaptation Choose among available models vs create/adapt a model for a specialized path. Adapted models become routing targets. Is the decision which model runs or how a model acquired its behavior? feeds Clear
cost-aware-routing latency-optimization Optimize unit economics under quality constraints vs optimize elapsed/perceived time. Smaller models, caching, and offline paths can improve both. Which metric is the binding objective when cost and speed conflict? coexists Clear
structured-output-contract prompt-chaining Define machine-readable stage interfaces vs define fixed stage sequence. Chains benefit from typed intermediate artifacts. Is the design about what crosses a boundary or which boundary runs next? contains/invokes Clear
llm-as-judge human-in-the-loop-gate Automated semantic scoring vs accountable human control/review. Humans calibrate judges and inspect uncertain cases. Can the decision be delegated to a calibrated rubric, or does it require accountable judgment? coexists Clear
llm-as-judge fine-tuning-adaptation Produce evaluation/reward signals vs update parameters. Judges can provide post-training rewards. Is the component scoring behavior or learning from the score? feeds Clear
data-quality-loop fine-tuning-adaptation Curate/correct data, labels, schemas, and knowledge vs update model parameters. The loop may produce training data. Does the promoted fix change data artifacts or model weights? feeds Clear
data-quality-loop prompt-versioning Promote corrections to evidence/data vs track instruction artifact releases. A discovered error may produce a new prompt version. Is the maintained artifact data/knowledge or prompt instructions? feeds Clear
fine-tuning-adaptation prompt-versioning Parameter change vs instruction artifact lifecycle. Both alter behavior and need evaluations. Can the change be inspected and rolled back as prompt text, or is it in weights? alternative Clear

Coverage Index

Every qualifying pattern appears below with all materially adjacent neighbors represented in the matrix. A pair not listed in a pattern's neighbor set is classified not materially adjacent for checkpoint-4 purposes.

Pattern Materially adjacent neighbors
tool-use-orchestration prompt-chaining, multi-agent-orchestration, safety-policy-enforcement, structured-output-contract
eval-driven-development llm-as-judge, observability-tracing, data-quality-loop, prompt-versioning, human-in-the-loop-gate
human-in-the-loop-gate eval-driven-development, safety-policy-enforcement, fallback-degradation, llm-as-judge
multi-agent-orchestration tool-use-orchestration, prompt-chaining, context-engineering, durable-execution
observability-tracing eval-driven-development, durable-execution, llm-as-judge
rag-retrieval agent-memory, context-engineering, fine-tuning-adaptation, data-quality-loop
safety-policy-enforcement tool-use-orchestration, human-in-the-loop-gate, guardrails-io-filtering, privacy-preserving-processing, structured-output-contract
agent-memory rag-retrieval, context-engineering, context-compaction, durable-execution
model-routing fallback-degradation, cost-aware-routing, latency-optimization, fine-tuning-adaptation
structured-output-contract tool-use-orchestration, safety-policy-enforcement, guardrails-io-filtering, prompt-chaining
llm-as-judge eval-driven-development, observability-tracing, human-in-the-loop-gate, fine-tuning-adaptation
cost-aware-routing fallback-degradation, model-routing, latency-optimization
latency-optimization context-compaction, fallback-degradation, model-routing, cost-aware-routing
guardrails-io-filtering safety-policy-enforcement, structured-output-contract, privacy-preserving-processing
context-compaction agent-memory, context-engineering, latency-optimization
data-quality-loop eval-driven-development, rag-retrieval, fine-tuning-adaptation, prompt-versioning
durable-execution multi-agent-orchestration, agent-memory, observability-tracing, fallback-degradation, prompt-chaining
fine-tuning-adaptation rag-retrieval, context-engineering, model-routing, llm-as-judge, data-quality-loop, prompt-versioning
privacy-preserving-processing safety-policy-enforcement, guardrails-io-filtering
context-engineering multi-agent-orchestration, rag-retrieval, agent-memory, context-compaction, fine-tuning-adaptation, prompt-versioning
fallback-degradation human-in-the-loop-gate, durable-execution, model-routing, cost-aware-routing, latency-optimization
prompt-versioning eval-driven-development, context-engineering, data-quality-loop, fine-tuning-adaptation
prompt-chaining tool-use-orchestration, multi-agent-orchestration, durable-execution, structured-output-contract

Accepted Neighbor Vocabulary

Unresolved Dependencies for Checkpoint 4

These singleton or ambiguous names remain explicitly unresolved. They are boundary-review inputs, not promoted pages:

Name Current likely boundary Checkpoint-4 question
workflow-orchestration tool-use-orchestration / durable-execution / multi-agent-orchestration Which actual topology does its single usage describe?
source-citation-grounding rag-retrieval plus citation fidelity Is citation provenance a distinct reusable control or a retrieval/output contract?
fallback-strategy fallback-degradation / model-routing Is the usage failure-triggered degradation or ordinary routing?
structured-output-validation structured-output-contract Does the usage add a distinct validation loop or only enforce the same schema boundary?
latency-aware-model-routing model-routing / latency-optimization Is latency merely one route dimension or the defining architecture?
fine-tuning-feedback-loop fine-tuning-adaptation plus data-quality-loop Does it have a distinct closed-loop training boundary?

Reviewer Questions

  1. Do the specialization pairs remain useful as separate pages, especially model-routing/cost-aware-routing and context-engineering/context-compaction?
  2. Does structured-output-contract remain distinct enough from guardrails-io-filtering when one validator performs both checks?
  3. Is the data-quality-loop/fine-tuning-adaptation boundary clear enough when production feedback directly becomes training data?
  4. Does prompt-chaining sit at the correct altitude beneath workflow-decomposition and beside tool/multi-agent orchestration?
  5. Should any unresolved singleton be normalized before library v1, or remain visible as an unresolved analysis note?
  6. Are any absent pairs materially adjacent and therefore missing from this sparse matrix?

Methodology and Limitations

The matrix was hand-authored after reading all 23 pages and the corpus-complete normalization record. It describes architectural responsibilities supported by the pages; it does not compute semantic distance, prove mutual exclusivity, or imply that unrelated patterns do not coexist. Checkpoint 4 remains the blocking qualitative review that may merge, split, rename, or clarify pages before a v1 declaration.