{ "generated_at": "2026-09-17T06:17:24Z", "mode": "DRAFT_EXPLORATION", "publish_eligible": false, "registries": { "architecture": { "criteria": [ { "allowedAnswers": [ "unknown", "yes", "no" ], "core": true, "id": "evaluation_ready", "question": "Is there a representative evaluation set and an explicit acceptance decision?" }, { "allowedAnswers": [ "unknown", "yes", "no" ], "core": true, "id": "fixed_steps", "question": "Can the work be described as a stable sequence of steps?" }, { "allowedAnswers": [ "unknown", "yes", "no" ], "core": true, "id": "external_knowledge", "question": "Does the task need a controlled external knowledge source?" }, { "allowedAnswers": [ "unknown", "yes", "no" ], "core": true, "id": "tool_actions", "question": "Must the system perform external tool actions?" }, { "allowedAnswers": [ "unknown", "yes", "no" ], "core": true, "id": "independent_subtasks", "question": "Can subtasks be isolated with explicit input and output contracts?" }, { "allowedAnswers": [ "unknown", "yes", "no" ], "core": true, "id": "parallel_justified", "question": "Does a measured baseline justify the added coordination cost?" }, { "allowedAnswers": [ "unknown", "yes", "no" ], "core": false, "id": "approval_boundary", "question": "For tool actions, are authorization and human approval boundaries explicit?" }, { "allowedAnswers": [ "unknown", "yes", "no" ], "core": false, "id": "recovery_tested", "question": "For tool actions, has failure recovery been exercised?" } ], "limitations": [ "Original engineering proposal, not an empirical benchmark or a released canonical source.", "Patterns may be combined after evaluation; the first match is a starting point, not proof of fitness.", "The 2024 pattern reference is conceptual; current framework features must be verified separately." ], "method": "Ascending priority, first matching rule; all AND any AND none; unspecified answers are unknown.", "references": [ { "checked_at": "2026-09-17", "scope": "Conceptual distinction between fixed workflows and dynamic agents, and common composition patterns; not validation of these proposed rules.", "url": "https://www.anthropic.com/engineering/building-effective-agents" } ], "rules": [ { "id": "ARCH-NO-EVALUATION", "next_steps": [ "Define representative cases, failure criteria and an accountable acceptance owner before selecting an autonomous path." ], "outcome": "EVALUATION_REQUIRED", "priority": 10, "rejected_options": [ "AUTONOMOUS_RELEASE" ], "when": { "all": [ { "criterionId": "evaluation_ready", "operator": "equals", "value": "no" } ], "any": [], "none": [] } }, { "id": "ARCH-UNKNOWN-CORE", "next_steps": [ "Resolve the unknown core questions; unknown is not equivalent to no." ], "outcome": "INSUFFICIENT_EVIDENCE", "priority": 20, "rejected_options": [], "when": { "all": [], "any": [ { "criterionId": "evaluation_ready", "operator": "equals", "value": "unknown" }, { "criterionId": "fixed_steps", "operator": "equals", "value": "unknown" }, { "criterionId": "external_knowledge", "operator": "equals", "value": "unknown" }, { "criterionId": "tool_actions", "operator": "equals", "value": "unknown" }, { "criterionId": "independent_subtasks", "operator": "equals", "value": "unknown" }, { "criterionId": "parallel_justified", "operator": "equals", "value": "unknown" } ], "none": [] } }, { "id": "ARCH-ACTION-BOUNDARY", "next_steps": [ "Define scoped authorization, approval, containment and recovery for external effects." ], "outcome": "ACTION_BOUNDARY_REQUIRED", "priority": 30, "rejected_options": [ "AUTONOMOUS_TOOL_ACTIONS" ], "when": { "all": [ { "criterionId": "tool_actions", "operator": "equals", "value": "yes" } ], "any": [ { "criterionId": "approval_boundary", "operator": "not_equals", "value": "yes" }, { "criterionId": "recovery_tested", "operator": "not_equals", "value": "yes" } ], "none": [] } }, { "id": "ARCH-FIXED-WORKFLOW", "next_steps": [ "Implement explicit step transitions and validation gates; attach retrieval or approved tools only where the step requires them." ], "outcome": "DETERMINISTIC_WORKFLOW", "priority": 40, "rejected_options": [], "when": { "all": [ { "criterionId": "fixed_steps", "operator": "equals", "value": "yes" } ], "any": [], "none": [] } }, { "id": "ARCH-RETRIEVAL", "next_steps": [ "Validate source access, retrieval coverage and answer grounding against the evaluation set." ], "outcome": "RETRIEVAL_PIPELINE", "priority": 50, "rejected_options": [], "when": { "all": [ { "criterionId": "tool_actions", "operator": "equals", "value": "no" }, { "criterionId": "external_knowledge", "operator": "equals", "value": "yes" } ], "any": [], "none": [] } }, { "id": "ARCH-ORCHESTRATION", "next_steps": [ "Define worker boundaries, a coordinator merge contract, cancellation and duplicate-effect handling before an evaluation." ], "outcome": "ORCHESTRATOR_WORKERS", "priority": 60, "rejected_options": [], "when": { "all": [ { "criterionId": "independent_subtasks", "operator": "equals", "value": "yes" }, { "criterionId": "parallel_justified", "operator": "equals", "value": "yes" } ], "any": [], "none": [] } }, { "id": "ARCH-BOUNDED-AGENT", "next_steps": [ "Start with one agent and scoped tools, iteration limits and a tested stop path; compare it with the simplest working baseline." ], "outcome": "BOUNDED_TOOL_AGENT", "priority": 70, "rejected_options": [], "when": { "all": [ { "criterionId": "tool_actions", "operator": "equals", "value": "yes" } ], "any": [], "none": [] } }, { "id": "ARCH-BASELINE", "next_steps": [ "Evaluate a single-call baseline before adding dynamic coordination." ], "outcome": "SINGLE_CALL_BASELINE", "priority": 80, "rejected_options": [], "when": { "all": [ { "criterionId": "evaluation_ready", "operator": "equals", "value": "yes" } ], "any": [], "none": [] } } ], "version": "0.1.0-draft.1" }, "build_buy": { "claims": { "claims": [ { "boundary": "This is an authored model contract, not an externally calibrated probability, forecast or benchmark.", "id": "CLM-METHOD-001", "sourceIds": [ "SRC-PHAROS-BUILD-BUY" ], "statement": "The scorecard uses 12 criteria with total weight 31, a signed decision index from -100 to +100 and transparent thresholds at -30 and +30." }, { "boundary": "NIST guidance is voluntary and does not transfer legal or operational accountability.", "id": "CLM-RISK-001", "sourceIds": [ "SRC-NIST-AI-RMF", "SRC-NIST-GAI-PROFILE" ], "statement": "AI sourcing decisions still require full-lifecycle risk ownership across design, development, use, evaluation and acquisition." }, { "boundary": "The categories are broad; each organization must supply its own amounts and measurable workload units.", "id": "CLM-TCO-001", "sourceIds": [ "SRC-FINOPS-TCO" ], "statement": "Technology TCO should include acquisition, management, support, labor, downtime, training and other productivity costs across time." }, { "boundary": "The cited consultancy framework is a relevance source, not independent proof of which path will succeed.", "id": "CLM-READINESS-001", "sourceIds": [ "SRC-KPMG-BUILD-BUY-BORROW" ], "statement": "Workforce capability, infrastructure, context and data, security, regulatory requirements and ability to scale are relevant to an agent sourcing decision." }, { "boundary": "The paper is conceptual and does not empirically validate this scorecard's weights or thresholds.", "id": "CLM-FACTOR-001", "sourceIds": [ "SRC-ARXIV-MAKE-BUY" ], "statement": "Cost, strategic differentiation, asset specificity, vendor lock-in, time to market, quality and compliance, and organizational capability are recurring make-or-buy determinants in the cited conceptual analysis." } ], "version": "1.0.0" }, "criteria": { "criteria": [ { "gateCritical": false, "id": "strategic_differentiation", "label": "Strategic differentiation", "options": [ { "blocks": [], "evidencePrompt": "Name the available products and the exact capability each already covers.", "id": "buy", "interpretation": "The workflow is available from several credible products and is not central to competitive advantage.", "label": "Useful but commodity", "value": -1 }, { "blocks": [], "evidencePrompt": "Separate commodity platform functions from logic that must remain yours.", "id": "hybrid", "interpretation": "The shell is common, but domain rules or user experience contain differentiated work.", "label": "Some differentiated logic", "value": 0 }, { "blocks": [], "evidencePrompt": "Describe what a competitor could not reproduce by buying the same platform.", "id": "build", "interpretation": "The agent's behavior, feedback loop or workflow is part of the product moat.", "label": "Core proprietary capability", "value": 1 }, { "blocks": [], "evidencePrompt": "Get a product-owner decision before treating the score as ready.", "id": "unknown", "interpretation": "The strategic role has not been agreed.", "label": "Unknown", "value": null } ], "order": 1, "question": "How much of the agent's behavior creates a durable product or operating advantage?", "weight": 3, "whyItMatters": "Commodity capability is expensive to recreate. Differentiating decision logic can be risky to surrender to a vendor roadmap." }, { "gateCritical": false, "id": "workflow_uniqueness", "label": "Workflow and policy uniqueness", "options": [ { "blocks": [], "evidencePrompt": "List the required steps and map them to available vendor features.", "id": "buy", "interpretation": "The process follows common product patterns with limited exceptions.", "label": "Standard workflow", "value": -1 }, { "blocks": [], "evidencePrompt": "Mark each step as configure, extend or own.", "id": "hybrid", "interpretation": "Most steps are standard while a few rules or actions need owned extensions.", "label": "Standard core, custom edges", "value": 0 }, { "blocks": [], "evidencePrompt": "Document the exceptions that generic products cannot express safely.", "id": "build", "interpretation": "The process depends on organization-specific rules, exceptions or sequencing.", "label": "Proprietary policy engine", "value": 1 }, { "blocks": [], "evidencePrompt": "Run a workflow discovery session with operators and control owners.", "id": "unknown", "interpretation": "The real workflow has not been mapped.", "label": "Unknown", "value": null } ], "order": 2, "question": "Are the agent's rules, tools and exception paths standard or organization-specific?", "weight": 3, "whyItMatters": "A standard workflow favors configuration. Deep policy logic and unusual exceptions increase the value of owned orchestration." }, { "gateCritical": true, "id": "data_control", "label": "Data control and residency", "options": [ { "blocks": [], "evidencePrompt": "Attach the approved data-flow, retention terms and deletion test.", "id": "buy", "interpretation": "A shortlisted vendor can meet documented handling and residency requirements.", "label": "Approved vendor path exists", "value": -1 }, { "blocks": [], "evidencePrompt": "Draw the trust boundary and identify which fields may cross it.", "id": "hybrid", "interpretation": "Commodity services are acceptable outside a controlled data and policy boundary.", "label": "Sensitive boundary can be isolated", "value": 0 }, { "blocks": [ "BUY" ], "evidencePrompt": "Cite the binding policy, regulation or security decision and the failed vendor requirement.", "id": "build", "interpretation": "No approved vendor route currently meets a mandatory data-control constraint.", "label": "Controlled custom path required", "value": 1 }, { "blocks": [], "evidencePrompt": "Complete data classification and a vendor data-flow review.", "id": "unknown", "interpretation": "Data classification or vendor handling is unresolved.", "label": "Unknown", "value": null } ], "order": 3, "question": "Can an approved vendor path meet data location, retention, deletion and training-use constraints?", "weight": 3, "whyItMatters": "A mandatory control that the shortlist cannot meet disqualifies a pure buy decision regardless of feature fit." }, { "gateCritical": false, "id": "integration_depth", "label": "Integration depth", "options": [ { "blocks": [], "evidencePrompt": "Verify the exact objects, permissions and rate limits in the vendor connector.", "id": "buy", "interpretation": "The use case relies on supported connectors and limited read-oriented access.", "label": "Standard connectors and reads", "value": -1 }, { "blocks": [], "evidencePrompt": "Identify the crown-jewel integrations and keep their contracts portable.", "id": "hybrid", "interpretation": "Common systems are covered, but one or more integrations require owned adapters.", "label": "Mixed connector depth", "value": 0 }, { "blocks": [], "evidencePrompt": "Map identity, authorization, idempotency, rollback and reconciliation for each action.", "id": "build", "interpretation": "The agent needs legacy access, custom APIs, transactional writes or organization-specific recovery.", "label": "Deep custom transaction paths", "value": 1 }, { "blocks": [], "evidencePrompt": "Inventory systems, actions, permissions, error modes and data owners.", "id": "unknown", "interpretation": "The integration inventory is incomplete.", "label": "Unknown", "value": null } ], "order": 4, "question": "How deeply must the agent interact with systems, permissions and failure recovery?", "weight": 3, "whyItMatters": "Connector count alone is weak evidence. Permission fidelity, write actions and recovery semantics drive integration ownership." }, { "gateCritical": false, "id": "time_to_value", "label": "Time to validated value", "options": [ { "blocks": [], "evidencePrompt": "Name the decision deadline and the KPI that must be measured by then.", "id": "buy", "interpretation": "The business needs a production-like baseline before a custom platform could be staffed and built.", "label": "Evidence needed immediately", "value": -1 }, { "blocks": [], "evidencePrompt": "Define the first reversible slice and the later migration gate.", "id": "hybrid", "interpretation": "A vendor or open component can establish a baseline while custom boundaries are validated.", "label": "Phased evidence window", "value": 0 }, { "blocks": [], "evidencePrompt": "Confirm staffing dates, release gates and the cost of delayed learning.", "id": "build", "interpretation": "The roadmap can absorb iterative architecture, evaluation and operating-model work.", "label": "Longer validation horizon", "value": 1 }, { "blocks": [], "evidencePrompt": "Set a business deadline and a production-like validation target.", "id": "unknown", "interpretation": "No decision deadline or measurable outcome is agreed.", "label": "Unknown", "value": null } ], "order": 5, "question": "How soon must a production-like workflow produce measurable evidence?", "weight": 3, "whyItMatters": "Urgency favors a configurable baseline, while a longer decision window can support owned architecture and repeated evaluation." }, { "gateCritical": true, "id": "team_capability", "label": "Engineering and product capability", "options": [ { "blocks": [ "BUILD" ], "evidencePrompt": "Record the unfilled roles and the opportunity cost of reallocating staff.", "id": "buy", "interpretation": "The organization cannot assign a durable cross-functional owner team.", "label": "No dedicated build team", "value": -1 }, { "blocks": [], "evidencePrompt": "Assign owners for product, platform, data, security and evaluation.", "id": "hybrid", "interpretation": "Internal engineers can own integration and domain logic while buying missing platform capability.", "label": "Strong software team, selective gaps", "value": 0 }, { "blocks": [], "evidencePrompt": "Name the people, allocation, prior systems and on-call plan.", "id": "build", "interpretation": "A durable team has relevant delivery and operating experience.", "label": "Experienced agent delivery team", "value": 1 }, { "blocks": [], "evidencePrompt": "Replace role names with named owners and committed capacity.", "id": "unknown", "interpretation": "Staffing is hypothetical.", "label": "Unknown", "value": null } ], "order": 6, "question": "Can a named internal team design, test and evolve the agent beyond a prototype?", "weight": 3, "whyItMatters": "Build capacity includes product ownership, data engineering, security and evaluation, not only prompt or application code." }, { "gateCritical": true, "id": "lifecycle_capacity", "label": "Lifecycle ownership", "options": [ { "blocks": [ "BUILD" ], "evidencePrompt": "Confirm what the vendor operates and what remains your responsibility.", "id": "buy", "interpretation": "The organization needs a vendor to carry much of the platform lifecycle.", "label": "Need managed lifecycle coverage", "value": -1 }, { "blocks": [], "evidencePrompt": "Create a RACI for monitoring, changes, incidents, data and access.", "id": "hybrid", "interpretation": "The vendor runs commodity layers while the organization owns domain quality and incidents.", "label": "Shared operating model", "value": 0 }, { "blocks": [], "evidencePrompt": "Attach the evaluation cadence, on-call rota, rollback path and change policy.", "id": "build", "interpretation": "The organization can own releases, regressions, infrastructure and incident response.", "label": "Full internal lifecycle team", "value": 1 }, { "blocks": [], "evidencePrompt": "Assign durable owners before approving a build.", "id": "unknown", "interpretation": "No operating owner has accepted responsibility.", "label": "Unknown", "value": null } ], "order": 7, "question": "Who will operate evaluation, monitoring, incidents and model or vendor changes?", "weight": 3, "whyItMatters": "The sourcing choice changes the work, but it never removes responsibility for outcomes and controls." }, { "gateCritical": false, "id": "capital_shape", "label": "Budget and capital shape", "options": [ { "blocks": [], "evidencePrompt": "Record approved setup spend, recurring spend and procurement limits.", "id": "buy", "interpretation": "Subscription or usage spend is easier to approve than a larger internal build.", "label": "Lower initial commitment", "value": -1 }, { "blocks": [], "evidencePrompt": "Separate vendor, integration, internal labor and contingency budgets.", "id": "hybrid", "interpretation": "The budget can fund a platform and a bounded custom layer.", "label": "Balanced spend profile", "value": 0 }, { "blocks": [], "evidencePrompt": "Verify capital, headcount and contingency approvals.", "id": "build", "interpretation": "The organization can fund discovery, engineering and operating capability before benefits stabilize.", "label": "Front-loaded investment is approved", "value": 1 }, { "blocks": [], "evidencePrompt": "Get separate approval ranges for one-time and recurring spend.", "id": "unknown", "interpretation": "Budget shape is not approved.", "label": "Unknown", "value": null } ], "order": 8, "question": "Does the organization prefer lower initial commitment or controlled long-term investment?", "weight": 2, "whyItMatters": "The same nominal budget can produce different decisions depending on cash timing and approved headcount." }, { "gateCritical": false, "id": "scale_economics", "label": "Scale and unit economics", "options": [ { "blocks": [], "evidencePrompt": "Define the unit of work and a bounded adoption range.", "id": "buy", "interpretation": "The organization benefits from avoiding fixed platform investment while demand is unproven.", "label": "Demand is uncertain or modest", "value": -1 }, { "blocks": [], "evidencePrompt": "Track cost per successful task, active user, action and exception.", "id": "hybrid", "interpretation": "Usage can be measured, but the economic crossover requires a live baseline.", "label": "Scale is measurable, crossover unclear", "value": 0 }, { "blocks": [], "evidencePrompt": "Show demand, peak shape, growth assumptions and cost per successful outcome.", "id": "build", "interpretation": "Stable high volume can justify owned cost controls and reusable platform work.", "label": "High predictable scale", "value": 1 }, { "blocks": [], "evidencePrompt": "Define a measurable unit and collect a baseline.", "id": "unknown", "interpretation": "No unit or demand model exists.", "label": "Unknown", "value": null } ], "order": 9, "question": "Are demand, workload units and marginal costs measurable enough to compare paths?", "weight": 2, "whyItMatters": "A low pilot price says little about per-user, per-task or per-action economics at production scale." }, { "gateCritical": false, "id": "portability_tolerance", "label": "Portability and vendor dependency", "options": [ { "blocks": [], "evidencePrompt": "Record renewal, export, termination assistance and deletion terms.", "id": "buy", "interpretation": "The organization accepts a managed platform dependency with bounded exit terms.", "label": "Contractual dependency is acceptable", "value": -1 }, { "blocks": [], "evidencePrompt": "Run an export and redeploy test for the assets that must remain yours.", "id": "hybrid", "interpretation": "Domain logic, data and evals remain portable while replaceable services are bought.", "label": "Portable core, managed shell", "value": 0 }, { "blocks": [], "evidencePrompt": "List every asset needed to operate after vendor account termination.", "id": "build", "interpretation": "The organization needs control over runtime boundaries and migration paths.", "label": "High portability requirement", "value": 1 }, { "blocks": [], "evidencePrompt": "Write an exit scenario before vendor selection or platform build.", "id": "unknown", "interpretation": "Exit requirements are undefined.", "label": "Unknown", "value": null } ], "order": 10, "question": "What must remain movable if the model, platform or commercial terms change?", "weight": 2, "whyItMatters": "Code ownership is incomplete when prompts, evals, state, logs or deployment cannot leave the vendor runtime." }, { "gateCritical": false, "id": "evaluation_control", "label": "Evaluation and observability control", "options": [ { "blocks": [], "evidencePrompt": "Verify each required artifact in a production-like trial.", "id": "buy", "interpretation": "The platform exposes the required traces, eval hooks, exports and release controls.", "label": "Vendor evidence is sufficient", "value": -1 }, { "blocks": [], "evidencePrompt": "Keep test data, metrics and release criteria outside the vendor-only surface.", "id": "hybrid", "interpretation": "The platform can be wrapped by an organization-owned test and observability layer.", "label": "Own evals around a bought runtime", "value": 0 }, { "blocks": [], "evidencePrompt": "Name the failure taxonomy, test set, trace fields and stop conditions.", "id": "build", "interpretation": "The use case needs custom evaluation, traces or action-level controls unavailable from products.", "label": "Domain-specific evidence is central", "value": 1 }, { "blocks": [], "evidencePrompt": "Define what must be observable before comparing vendors or architectures.", "id": "unknown", "interpretation": "The evidence contract is not defined.", "label": "Unknown", "value": null } ], "order": 11, "question": "Can a bought product expose the evidence needed to test domain quality and actions?", "weight": 2, "whyItMatters": "A system cannot be governed from a demo score. Teams need test sets, traces, failure labels and release gates tied to their own workflow." }, { "gateCritical": true, "id": "compliance_path", "label": "Compliance and third-party path", "options": [ { "blocks": [], "evidencePrompt": "Attach the approved control matrix, contracts and residual-risk owner.", "id": "buy", "interpretation": "A vendor route can satisfy documented procurement, security and compliance gates.", "label": "Approved vendor controls fit", "value": -1 }, { "blocks": [], "evidencePrompt": "Map each obligation to the vendor, internal owner or shared process.", "id": "hybrid", "interpretation": "The path needs vendor evidence plus organization-owned controls and records.", "label": "Control boundary must be shared", "value": 0 }, { "blocks": [ "BUY" ], "evidencePrompt": "Cite the failed gate and the evidence a controlled build must produce.", "id": "build", "interpretation": "A mandatory approval or control cannot be met by the current vendor shortlist.", "label": "Custom controlled path required", "value": 1 }, { "blocks": [], "evidencePrompt": "Get legal, security and risk owners to classify the use case.", "id": "unknown", "interpretation": "The applicable obligations or approval route are unresolved.", "label": "Unknown", "value": null } ], "order": 12, "question": "Which sourcing path can produce the required approvals, records and accountability?", "weight": 2, "whyItMatters": "A vendor certification does not transfer accountability. A custom build also creates documentation and control obligations." } ], "publishedOn": "2026-08-25", "thresholds": { "boundaryMargin": 7, "buildMin": 30, "buyMax": -30, "minimumKnownWeightShare": 0.75 }, "totalWeight": 31, "version": "1.0.0" }, "methodology": { "boundaries": [ "The decision index is an authored decision-support model, not a probability, benchmark, forecast or confidence score.", "Thresholds are transparent editorial defaults and have not been statistically calibrated against project outcomes.", "TCO values are editable scenario assumptions, not Pharos pricing and not market averages.", "A low-cost path can still fail a mandatory data, security, legal or operating constraint.", "The tool does not replace architecture discovery, procurement, security review, legal advice or financial analysis." ], "exportSchemaVersion": "1.0.0", "formula": "round-half-away-from-zero(100 * signed known points / known weight)", "lastVerifiedOn": "2026-08-25", "outcomeIds": [ "BUY", "HYBRID", "BUILD", "UNRESOLVED" ], "procedure": [ "Choose one evidence-backed answer for each criterion.", "Multiply each known answer value by its criterion weight and sum signed points.", "Normalize signed points only across known weight so unknown answers never act as neutral evidence.", "Apply BUY at -30 or lower, BUILD at +30 or higher and HYBRID between those thresholds.", "Withhold a decision when known-weight coverage is below 75 percent or any gate-critical criterion is unknown.", "Show a decision-boundary warning when the index is within seven points of either outcome threshold.", "If a selected constraint blocks the numeric outcome, require constraint review and return HYBRID as the working architecture boundary.", "Compare the result with editable three-year TCO assumptions and record both artifacts together." ], "publishedOn": "2026-08-25", "registryVersions": { "claims": "1.0.0", "criteria": "1.0.0", "sources": "1.0.0", "tco": "1.0.0" }, "scoringScale": { "build": 1, "buy": -1, "hybrid": 0, "unknown": null }, "statusIds": [ "DECISION_READY", "INSUFFICIENT_DECISION_EVIDENCE", "CONSTRAINT_REVIEW_REQUIRED" ], "version": "1.0.0" }, "profiles": { "profiles": [ { "answers": { "capital_shape": "unknown", "compliance_path": "unknown", "data_control": "unknown", "evaluation_control": "unknown", "integration_depth": "unknown", "lifecycle_capacity": "unknown", "portability_tolerance": "unknown", "scale_economics": "unknown", "strategic_differentiation": "unknown", "team_capability": "unknown", "time_to_value": "unknown", "workflow_uniqueness": "unknown" }, "description": "Every answer remains unknown until evidence is supplied.", "id": "blank", "label": "Start blank" }, { "answers": { "capital_shape": "buy", "compliance_path": "buy", "data_control": "buy", "evaluation_control": "hybrid", "integration_depth": "buy", "lifecycle_capacity": "buy", "portability_tolerance": "buy", "scale_economics": "buy", "strategic_differentiation": "buy", "team_capability": "hybrid", "time_to_value": "buy", "workflow_uniqueness": "buy" }, "description": "Common workflow, modest scale, approved vendor route and urgent baseline.", "id": "commodity-internal", "label": "Commodity internal workflow" }, { "answers": { "capital_shape": "build", "compliance_path": "build", "data_control": "build", "evaluation_control": "build", "integration_depth": "build", "lifecycle_capacity": "build", "portability_tolerance": "build", "scale_economics": "build", "strategic_differentiation": "build", "team_capability": "build", "time_to_value": "hybrid", "workflow_uniqueness": "build" }, "description": "Proprietary workflow with deep integrations, controlled data path and durable internal ownership.", "id": "controlled-core", "label": "Differentiated controlled system" }, { "answers": { "capital_shape": "hybrid", "compliance_path": "hybrid", "data_control": "hybrid", "evaluation_control": "hybrid", "integration_depth": "build", "lifecycle_capacity": "hybrid", "portability_tolerance": "hybrid", "scale_economics": "hybrid", "strategic_differentiation": "hybrid", "team_capability": "hybrid", "time_to_value": "hybrid", "workflow_uniqueness": "hybrid" }, "description": "Buy commodity infrastructure while retaining domain logic, critical adapters and evaluation assets.", "id": "hybrid-transition", "label": "Hybrid transition" } ], "version": "1.0.0" }, "sources": { "sources": [ { "id": "SRC-PHAROS-BUILD-BUY", "limitations": "Used to seed the topic vocabulary and six-factor starting point. Its engagement examples, universal ranges and commercial conclusions are not treated as independent validation of this tool's thresholds.", "publisher": "Pharos Production", "sourceType": "first-party", "title": "Build vs Buy AI Agent: 2026 Decision Framework", "url": "https://pharosproduction.com/insights/engineering/build-vs-buy-ai-agent/", "verifiedOn": "2026-08-25" }, { "id": "SRC-NIST-AI-RMF", "limitations": "Voluntary risk-management guidance. NIST states that AI RMF 1.0 is being revised; it does not prescribe build, buy or hybrid thresholds.", "publisher": "National Institute of Standards and Technology", "sourceType": "primary-standard", "title": "AI Risk Management Framework", "url": "https://www.nist.gov/itl/ai-risk-management-framework", "verifiedOn": "2026-08-25" }, { "id": "SRC-NIST-GAI-PROFILE", "limitations": "A cross-sectoral companion profile for generative AI risk management. It informs lifecycle and acquisition questions but does not validate economic weights or sourcing outcomes.", "publisher": "National Institute of Standards and Technology", "sourceType": "primary-standard", "title": "Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile", "url": "https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.600-1.pdf", "verifiedOn": "2026-08-25" }, { "id": "SRC-FINOPS-TCO", "limitations": "Defines broad TCO and unit-economics concepts for technology costs. It supplies no AI-agent price ranges and does not validate this scenario's illustrative amounts.", "publisher": "FinOps Foundation", "sourceType": "primary-framework", "title": "FinOps Terminology: Total Cost of Ownership and Unit Economics", "url": "https://framework.finops.org/assets/terminology/", "verifiedOn": "2026-08-25" }, { "id": "SRC-KPMG-BUILD-BUY-BORROW", "limitations": "A consultancy-authored framework with commercial interests. It supports the relevance of readiness, infrastructure, data, security, compliance and scale dimensions, not this tool's score formula.", "publisher": "KPMG Ireland", "sourceType": "first-party-research", "title": "Agentic AI Untangled: Navigating the Build, Buy, or Borrow Decision", "url": "https://assets.kpmg.com/content/dam/kpmgsites/ie/pdf/insights/ai-artificial-intelligence/ie-agentic-ai-build-buy-borrow.pdf.coredownload.inline.pdf", "verifiedOn": "2026-08-25" }, { "id": "SRC-ARXIV-MAKE-BUY", "limitations": "A conceptual preprint rather than an empirical calibration study. Its factor taxonomy informs coverage but cannot establish universal sourcing thresholds or expected outcomes.", "publisher": "arXiv", "sourceType": "conceptual-preprint", "title": "The Buy-or-Build Decision, Revisited: How Agentic AI Changes the Economics of Enterprise Software", "url": "https://arxiv.org/abs/2604.26482", "verifiedOn": "2026-08-25" } ], "verifiedOn": "2026-08-25", "version": "1.0.0" }, "tco": { "categories": [ { "cadence": "one_time", "description": "Workflow, data, risk, product and architecture discovery before implementation.", "id": "discovery_design", "label": "Discovery and architecture" }, { "cadence": "one_time", "description": "Configuration or development, integrations, environments, testing and release preparation.", "id": "implementation_integration", "label": "Implementation and integration" }, { "cadence": "one_time", "description": "Data cleanup, permissions, migration, indexing and initial evaluation-set preparation.", "id": "data_migration", "label": "Data preparation and migration" }, { "cadence": "annual", "description": "Platform subscription, model usage, cloud resources, storage, networking and observability services.", "id": "licenses_model_infrastructure", "label": "Licenses, models and infrastructure" }, { "cadence": "annual", "description": "Product ownership, engineering, data, platform operations and incident coverage retained internally.", "id": "internal_operations", "label": "Internal product and operations team" }, { "cadence": "annual", "description": "Test-set maintenance, red teaming, security review, audit evidence and control operation.", "id": "evaluation_security_compliance", "label": "Evaluation, security and compliance" }, { "cadence": "annual", "description": "User enablement, support, workflow change, model or vendor updates and business-process maintenance.", "id": "change_support", "label": "Change, training and support" }, { "cadence": "exit", "description": "Export, migration, contract exit, redeployment, retraining and parallel-run allowance at the horizon.", "id": "exit_switching_reserve", "label": "Exit and switching reserve" } ], "currency": "USD", "defaultScenario": { "boundary": "These figures are not Pharos pricing, vendor quotes or market benchmarks. Replace every amount with organization-specific evidence.", "label": "Editable illustrative scenario", "values": { "build": { "change_support": 60000, "data_migration": 80000, "discovery_design": 90000, "evaluation_security_compliance": 90000, "exit_switching_reserve": 40000, "implementation_integration": 300000, "internal_operations": 260000, "licenses_model_infrastructure": 120000 }, "buy": { "change_support": 50000, "data_migration": 50000, "discovery_design": 40000, "evaluation_security_compliance": 70000, "exit_switching_reserve": 220000, "implementation_integration": 120000, "internal_operations": 100000, "licenses_model_infrastructure": 300000 }, "hybrid": { "change_support": 60000, "data_migration": 60000, "discovery_design": 70000, "evaluation_security_compliance": 80000, "exit_switching_reserve": 130000, "implementation_integration": 210000, "internal_operations": 190000, "licenses_model_infrastructure": 160000 } } }, "horizonYears": 3, "paths": [ "build", "buy", "hybrid" ], "version": "1.0.0" } }, "framework": { "claims": { "claimEvidence": "Named official documentation, repositories, and release records in data/sources.json", "claims": [ { "boundary": "Inclusion is an editorial shortlist boundary, not a universal market ranking.", "disposition": "KEEP", "id": "CLM-001", "role": "PRIMARY", "sourceIds": [ "LG-REL", "MAF-REL", "CR-REL", "OA-REL", "GA-REL" ], "statement": "The comparison contains five maintained, permissively licensed code-first frameworks with documented paths for state, approval, observability, recovery, and deployment.", "status": "OBSERVED" }, { "boundary": "Production durability depends on a persistent checkpointer and replay-safe node design.", "disposition": "KEEP", "id": "CLM-002", "role": "PRIMARY", "sourceIds": [ "LG-PERSIST", "LG-HITL", "LG-SERVER" ], "statement": "LangGraph provides native checkpointed state, interrupts, and durable execution primitives for explicit state-graph workflows.", "status": "OBSERVED" }, { "boundary": "Language and UI integration parity must be verified for the selected client surface.", "disposition": "KEEP", "id": "CLM-003", "role": "PRIMARY", "sourceIds": [ "MAF-CHECKPOINT", "MAF-HITL" ], "statement": "Microsoft Agent Framework workflows provide checkpoint managers and request-based human approval with pending requests preserved across restore.", "status": "OBSERVED" }, { "boundary": "Managed AMP capabilities are distinct from the open-source Python runtime.", "disposition": "KEEP", "id": "CLM-004", "role": "PRIMARY", "sourceIds": [ "CR-FLOWS", "CR-AMP" ], "statement": "CrewAI Flows provide persisted state and human-feedback routing, while CrewAI AMP adds a managed deployment and observability path.", "status": "OBSERVED" }, { "boundary": "The Python SDK does not itself operate an external durable workflow runtime.", "disposition": "KEEP", "id": "CLM-005", "role": "PRIMARY", "sourceIds": [ "OA-RUN", "OA-HITL", "OA-TRACE" ], "statement": "OpenAI Agents SDK provides sessions, serialized approval state, and built-in tracing; process-restart durability uses documented external workflow integrations.", "status": "OBSERVED" }, { "boundary": "In-memory sessions are non-durable and cross-language feature parity is not assumed.", "disposition": "KEEP", "id": "CLM-006", "role": "PRIMARY", "sourceIds": [ "GA-STATE", "GA-HITL", "GA-RESUME", "GA-OBS", "GA-DEPLOY" ], "statement": "Google ADK persists state through durable session services, supports human input and resume in workflows, exports telemetry, and documents several Google Cloud deployment targets.", "status": "OBSERVED" }, { "boundary": "This is a method conclusion drawn from the documented mechanism boundaries, not a vendor performance claim.", "disposition": "KEEP", "id": "CLM-007", "role": "SUPPORTING", "sourceIds": [ "LG-HITL", "MAF-HITL", "CR-FLOWS", "OA-HITL", "GA-HITL" ], "statement": "A framework primitive is not by itself an enterprise control: identity, authorization, policy, audit retention, side-effect idempotency, and incident procedures remain application and operations work.", "status": "OBSERVED" } ], "evidenceCutoff": "2026-08-28", "schemaVersion": "1.0.0" }, "frameworks": { "comparisonBoundary": "Maintained, permissively licensed, code-first agent frameworks with documented state, human-approval, observability, recovery, and deployment paths for enterprise workflows.", "evidenceCutoff": "2026-08-28", "frameworks": [ { "bestFor": "Long-running or failure-sensitive workflows where teams need explicit control over state transitions, pause/resume behavior, and replay boundaries.", "capabilities": { "approval": { "detail": "The interrupt primitive pauses execution, exposes approval context, and resumes the same thread after an external decision.", "sourceIds": [ "LG-HITL" ], "status": "native" }, "deployment": { "detail": "The open-source runtime can run in application infrastructure, while Agent Server packages graphs with persistence and a durable task queue for self-hosted or managed deployment.", "sourceIds": [ "LG-SERVER" ], "status": "self-host-or-managed" }, "observability": { "detail": "Execution paths and state transitions can be traced through LangSmith; tracing can be disabled or conditioned for data-sensitive runs.", "sourceIds": [ "LG-OVERVIEW", "LG-PRIVACY" ], "status": "integrated" }, "recovery": { "detail": "Durable execution resumes from checkpoints after interruption or worker failure; node boundaries determine how much work may repeat.", "sourceIds": [ "LG-PERSIST", "LG-SERVER" ], "status": "native" }, "state": { "detail": "A checkpointer stores graph state as snapshots organized into threads at graph steps; persistent backends replace InMemorySaver for production.", "sourceIds": [ "LG-PERSIST" ], "status": "native" } }, "constraints": "The graph model exposes control but increases design and testing work. Production persistence needs a durable checkpointer or Agent Server rather than the in-memory example backend.", "deploymentControl": [ "self-host", "managed" ], "disqualifiers": [ "A short, deterministic tool call sequence is simpler in ordinary application code", "The team cannot own explicit graph topology and checkpoint migration" ], "id": "langgraph", "languages": [ "python", "typescript" ], "license": "MIT", "modelStrategy": [ "agnostic" ], "name": "LangGraph", "orchestrationModel": "Explicit state graph with nodes, edges, conditional routing, cycles, and checkpointed execution.", "priorities": [ "durable-recovery", "approval", "observability", "explicit-control" ], "release": { "publishedAt": "2026-08-11", "sourceId": "LG-REL", "track": "Python core", "version": "1.2.11" }, "tradeoffs": [ "High control and inspectable state", "More orchestration code and a steeper graph mental model", "LangSmith is an integrated product, not a requirement of the open-source runtime" ] }, { "bestFor": "Enterprises standardizing on .NET, Python, Go, Azure identity, or Microsoft operations while retaining explicit multi-agent workflow control.", "capabilities": { "approval": { "detail": "RequestPort and tool-approval requests pause workflows; pending requests are included in checkpoints and re-emitted after restore.", "sourceIds": [ "MAF-HITL" ], "status": "native" }, "deployment": { "detail": "Agents and workflows run in-process or behind application services and integrate with Microsoft and Azure agent services; deployment architecture remains an application decision.", "sourceIds": [ "MAF-OVERVIEW" ], "status": "self-host-or-managed" }, "observability": { "detail": "The agent capability surface exports traces, metrics, and logs and supports shared interception hooks for operational controls.", "sourceIds": [ "MAF-CAPABILITIES" ], "status": "native" }, "recovery": { "detail": "Workflows resume or rehydrate from a checkpoint, including pending request state; durable storage must replace in-memory checkpoint storage.", "sourceIds": [ "MAF-CHECKPOINT", "MAF-HITL" ], "status": "native" }, "state": { "detail": "Checkpoint managers capture executor state, pending messages, requests, responses, and shared state at superstep boundaries.", "sourceIds": [ "MAF-CHECKPOINT" ], "status": "native" } }, "constraints": "Feature parity varies by language and integration. Teams must verify the exact client surface because some UI workflow lifecycle events and resume controls are not mapped in the .NET AG-UI integration.", "deploymentControl": [ "self-host", "managed" ], "disqualifiers": [ "The team requires identical workflow lifecycle behavior across every language client", "The deployment platform must avoid Microsoft-specific operational dependencies" ], "id": "microsoft-agent-framework", "languages": [ "python", "dotnet", "go" ], "license": "MIT", "modelStrategy": [ "agnostic", "microsoft" ], "name": "Microsoft Agent Framework", "orchestrationModel": "Agent abstractions plus graph workflows, orchestrations, request ports, middleware, and checkpoint managers.", "priorities": [ "durable-recovery", "approval", "observability", "microsoft-stack", "multi-language" ], "release": { "publishedAt": "2026-08-28", "sourceId": "MAF-REL", "track": "Python", "version": "1.16.0" }, "tradeoffs": [ "Typed enterprise integration and multi-language workflow support", "Fast-moving APIs require pinned versions and migration tests", "Azure alignment is strongest, although the framework supports multiple providers" ] }, { "bestFor": "Python teams composing role-oriented multi-agent business processes that benefit from a concise crew abstraction and Flow-based approval routing.", "capabilities": { "approval": { "detail": "The human_feedback decorator pauses a Flow for review and can route approved, rejected, or revision outcomes to distinct listeners.", "sourceIds": [ "CR-FLOWS" ], "status": "native" }, "deployment": { "detail": "Python services can self-host crews and Flows; AMP offers GitHub-, CLI-, or Studio-driven managed deployments with generated API access.", "sourceIds": [ "CR-AMP", "CR-PROD" ], "status": "self-host-or-managed" }, "observability": { "detail": "CrewAI exposes execution tracing, while the managed AMP platform adds detailed traces, logs, team controls, and production monitoring.", "sourceIds": [ "CR-DOCS", "CR-AMP" ], "status": "integrated" }, "recovery": { "detail": "Persisted Flow state survives restarts and can resume by ID, but the deployment must supply durable storage and idempotent side-effect handling.", "sourceIds": [ "CR-FLOWS", "CR-PROD" ], "status": "configured" }, "state": { "detail": "The persist decorator stores Flow state, uses SQLite by default, and supports resuming or forking a persisted flow by state identifier.", "sourceIds": [ "CR-FLOWS" ], "status": "native" } }, "constraints": "Crews and Flows are different orchestration layers. Durable production behavior depends on configuring persistence and the surrounding runtime rather than assuming a crew definition is automatically recoverable.", "deploymentControl": [ "self-host", "managed" ], "disqualifiers": [ "The core application stack is TypeScript or .NET with no Python service boundary", "The workflow requires graph-level checkpoint semantics to be the primary programming model" ], "id": "crewai", "languages": [ "python" ], "license": "MIT", "modelStrategy": [ "agnostic" ], "name": "CrewAI", "orchestrationModel": "Role-based crews for agent collaboration plus event-driven Flows for routing, persistence, and approval points.", "priorities": [ "multi-agent", "approval", "rapid-composition", "managed-cloud" ], "release": { "publishedAt": "2026-08-27", "sourceId": "CR-REL", "track": "Python", "version": "1.15.18" }, "tradeoffs": [ "Concise role and task composition", "Flows add operational control but create a second mental model", "Managed deployment and full observability are concentrated in CrewAI AMP" ] }, { "bestFor": "Python products centered on OpenAI models that need a small orchestration surface, built-in tracing, sessions, guardrails, and tool approval.", "capabilities": { "approval": { "detail": "Tools can require approval, paused RunState can be serialized, and unresolved interruptions remain pending until approved or rejected.", "sourceIds": [ "OA-HITL" ], "status": "native" }, "deployment": { "detail": "The SDK runs inside a Python application; teams choose their hosting and add a durable orchestrator when workflows span long waits or process restarts.", "sourceIds": [ "OA-RUN" ], "status": "application-owned" }, "observability": { "detail": "Built-in tracing records model generations, tool calls, handoffs, guardrails, and custom events with controls for sensitive trace content.", "sourceIds": [ "OA-TRACE" ], "status": "native" }, "recovery": { "detail": "Serialized run state supports approval pause/resume; process-restart durability and automatic recovery use documented Dapr, Temporal, Restate, or DBOS integrations.", "sourceIds": [ "OA-HITL", "OA-RUN" ], "status": "external-runtime" }, "state": { "detail": "Sessions store conversation history through SQLite, Redis, SQLAlchemy, Dapr, encrypted, file-backed, or OpenAI-managed implementations.", "sourceIds": [ "OA-RUN" ], "status": "native" } }, "constraints": "The SDK manages the agent loop, but durable long-running recovery comes through Dapr, Temporal, Restate, DBOS, or application infrastructure rather than a built-in workflow server.", "deploymentControl": [ "self-host", "managed" ], "disqualifiers": [ "Provider neutrality is a hard procurement requirement", "The system needs a built-in graph runtime and checkpoint manager without external workflow infrastructure" ], "id": "openai-agents-sdk", "languages": [ "python" ], "license": "MIT", "modelStrategy": [ "openai", "agnostic-with-adapters" ], "name": "OpenAI Agents SDK", "orchestrationModel": "A compact runner with agents, tools, handoffs, guardrails, sessions, interruptions, and tracing.", "priorities": [ "approval", "observability", "minimal-sdk", "openai-stack" ], "release": { "publishedAt": "2026-08-19", "sourceId": "OA-REL", "track": "Python", "version": "0.22.0" }, "tradeoffs": [ "Small API surface and first-party OpenAI integrations", "Provider adapters exist but the default path is OpenAI-first", "Durability adds an external orchestration dependency" ] }, { "bestFor": "Multi-language teams using Gemini or Google Cloud that want one agent toolkit across local development, Agent Runtime, Cloud Run, and GKE.", "capabilities": { "approval": { "detail": "Graph workflows can request external human input and resume through the workflow runtime rather than hiding approval inside prompt text.", "sourceIds": [ "GA-HITL", "GA-RESUME" ], "status": "native" }, "deployment": { "detail": "Documented targets include Agent Runtime, Cloud Run, and GKE, with CLI support for CI/CD, service identity, secrets, scaling, and observability.", "sourceIds": [ "GA-DEPLOY" ], "status": "self-host-or-managed" }, "observability": { "detail": "OpenTelemetry instrumentation exports agent and tool traces; Google tooling connects traces to Cloud Trace and optional BigQuery analytics.", "sourceIds": [ "GA-OBS" ], "status": "integrated" }, "recovery": { "detail": "Resumable workflow sessions persist execution progress and continue graph work after interruption when backed by a persistent session service.", "sourceIds": [ "GA-RESUME", "GA-STATE" ], "status": "native" }, "state": { "detail": "Serializable session state is persisted by DatabaseSessionService or VertexAiSessionService; InMemorySessionService loses state on restart.", "sourceIds": [ "GA-STATE" ], "status": "native" } }, "constraints": "Session persistence depends on the selected SessionService, and operational features differ across SDK languages and deployment targets. In-memory sessions do not survive a restart.", "deploymentControl": [ "self-host", "managed" ], "disqualifiers": [ "The organization requires a cloud-neutral operations plane", "The chosen language needs a feature that exists only in another ADK SDK" ], "id": "google-adk", "languages": [ "python", "typescript", "go", "java", "kotlin" ], "license": "Apache-2.0", "modelStrategy": [ "google", "agnostic" ], "name": "Google ADK", "orchestrationModel": "LLM agents, deterministic sequential/parallel/loop agents, and graph workflows backed by sessions and events.", "priorities": [ "multi-language", "managed-cloud", "observability", "google-stack", "multi-agent" ], "release": { "publishedAt": "2026-08-26", "sourceId": "GA-REL", "track": "Python", "version": "2.8.0" }, "tradeoffs": [ "Broad language coverage and Google Cloud deployment paths", "Strongest operational path assumes Google Cloud", "Teams must verify feature parity for their chosen language" ] } ], "methodVersion": "1.0.0", "schemaVersion": "1.0.0" }, "sources": { "evidenceCutoff": "2026-08-28", "schemaVersion": "1.0.0", "sources": [ { "framework": "langgraph", "id": "LG-REL", "limitations": "The repository publishes several packages; this record identifies the Python langgraph core release rather than the separately versioned SDK or checkpoint packages.", "publisher": "LangChain", "sourceType": "official-release", "supports": [ "release", "license" ], "title": "LangGraph 1.2.11 release", "url": "https://github.com/langchain-ai/langgraph/releases/tag/1.2.11", "verifiedOn": "2026-08-28" }, { "framework": "langgraph", "id": "LG-OVERVIEW", "limitations": "Product overview language is used only for documented mechanisms; it is not evidence of comparative superiority, adoption, or outcome quality.", "publisher": "LangChain", "sourceType": "official-docs", "supports": [ "orchestration", "observability", "deployment" ], "title": "LangGraph overview", "url": "https://docs.langchain.com/oss/python/langgraph/overview", "verifiedOn": "2026-08-28" }, { "framework": "langgraph", "id": "LG-PERSIST", "limitations": "InMemorySaver examples are development fixtures. Production durability depends on selecting and operating a persistent checkpointer backend.", "publisher": "LangChain", "sourceType": "official-docs", "supports": [ "state", "recovery" ], "title": "LangGraph persistence", "url": "https://docs.langchain.com/oss/python/langgraph/persistence", "verifiedOn": "2026-08-28" }, { "framework": "langgraph", "id": "LG-HITL", "limitations": "An interrupt supplies a pause/resume primitive. Approval policy, identity, authorization, timeout, and escalation remain application responsibilities.", "publisher": "LangChain", "sourceType": "official-docs", "supports": [ "approval", "recovery" ], "title": "LangGraph interrupts", "url": "https://docs.langchain.com/oss/python/langgraph/interrupts", "verifiedOn": "2026-08-28" }, { "framework": "langgraph", "id": "LG-SERVER", "limitations": "Agent Server is a deployment runtime with infrastructure and licensing considerations separate from the open-source graph library.", "publisher": "LangChain", "sourceType": "official-docs", "supports": [ "recovery", "deployment" ], "title": "LangGraph Agent Server", "url": "https://docs.langchain.com/langsmith/agent-server", "verifiedOn": "2026-08-28" }, { "framework": "langgraph", "id": "LG-PRIVACY", "limitations": "Storage and trace behavior depends on runtime configuration. Teams must verify telemetry, retention, encryption, and regional requirements for their deployment.", "publisher": "LangChain", "sourceType": "official-docs", "supports": [ "observability", "state" ], "title": "LangSmith data storage and privacy", "url": "https://docs.langchain.com/langsmith/data-storage-and-privacy", "verifiedOn": "2026-08-28" }, { "framework": "microsoft-agent-framework", "id": "MAF-REL", "limitations": "The repository versions language packages independently. This comparison stamps the Python track and does not imply identical version numbers across .NET and Go.", "publisher": "Microsoft", "sourceType": "official-release", "supports": [ "release", "license" ], "title": "Microsoft Agent Framework Python 1.16.0 release", "url": "https://github.com/microsoft/agent-framework/releases/tag/python-1.16.0", "verifiedOn": "2026-08-28" }, { "framework": "microsoft-agent-framework", "id": "MAF-OVERVIEW", "limitations": "The overview identifies product scope and succession from AutoGen and Semantic Kernel; it does not prove compatibility with every predecessor application.", "publisher": "Microsoft Learn", "sourceType": "official-docs", "supports": [ "orchestration", "deployment", "positioning" ], "title": "Microsoft Agent Framework overview", "url": "https://learn.microsoft.com/en-us/agent-framework/overview/", "verifiedOn": "2026-08-28" }, { "framework": "microsoft-agent-framework", "id": "MAF-CHECKPOINT", "limitations": "Checkpoint availability does not make external side effects idempotent. Durable storage, schema migration, and replay-safe tool design remain deployment work.", "publisher": "Microsoft Learn", "sourceType": "official-docs", "supports": [ "state", "recovery" ], "title": "Agent Framework workflow checkpoints", "url": "https://learn.microsoft.com/en-us/agent-framework/workflows/checkpoints", "verifiedOn": "2026-08-28" }, { "framework": "microsoft-agent-framework", "id": "MAF-HITL", "limitations": "Request and approval primitives do not define the enterprise approver identity, authorization policy, response deadline, or separation-of-duties rules.", "publisher": "Microsoft Learn", "sourceType": "official-docs", "supports": [ "approval", "recovery" ], "title": "Agent Framework workflow human-in-the-loop", "url": "https://learn.microsoft.com/en-us/agent-framework/workflows/human-in-the-loop", "verifiedOn": "2026-08-28" }, { "framework": "microsoft-agent-framework", "id": "MAF-CAPABILITIES", "limitations": "A capability catalog confirms supported surfaces, not their readiness for a particular workload, language binding, compliance regime, or operations team.", "publisher": "Microsoft Learn", "sourceType": "official-docs", "supports": [ "observability", "governance" ], "title": "Agent capabilities", "url": "https://learn.microsoft.com/en-us/agent-framework/agents/", "verifiedOn": "2026-08-28" }, { "framework": "crewai", "id": "CR-REL", "limitations": "The release tag stamps the Python package. Managed AMP capabilities have a separate service lifecycle and are not versioned by this tag.", "publisher": "CrewAI", "sourceType": "official-release", "supports": [ "release", "license" ], "title": "CrewAI 1.15.18 release", "url": "https://github.com/crewAIInc/crewAI/releases/tag/1.15.18", "verifiedOn": "2026-08-28" }, { "framework": "crewai", "id": "CR-DOCS", "limitations": "The documentation landing page describes the combined open-source and enterprise journey; individual deployment entitlements require separate verification.", "publisher": "CrewAI", "sourceType": "official-docs", "supports": [ "orchestration", "observability" ], "title": "CrewAI documentation", "url": "https://docs.crewai.com/", "verifiedOn": "2026-08-28" }, { "framework": "crewai", "id": "CR-FLOWS", "limitations": "SQLite persistence is a default mechanism, not a production architecture. The application must select durable storage and define idempotent effects and approval delivery.", "publisher": "CrewAI", "sourceType": "official-docs", "supports": [ "state", "approval", "recovery" ], "title": "CrewAI Flows", "url": "https://docs.crewai.com/en/concepts/flows", "verifiedOn": "2026-08-28" }, { "framework": "crewai", "id": "CR-AMP", "limitations": "AMP is a managed commercial platform. Its presence does not mean the open-source package includes every managed monitoring, access, or scaling function.", "publisher": "CrewAI", "sourceType": "official-docs", "supports": [ "observability", "deployment" ], "title": "CrewAI AMP introduction", "url": "https://docs.crewai.com/enterprise/introduction", "verifiedOn": "2026-08-28" }, { "framework": "crewai", "id": "CR-PROD", "limitations": "This versioned repository document predates the comparison cutoff release. It supports the architectural boundary only; current commands must be checked against live docs.", "publisher": "CrewAI", "sourceType": "official-repository-docs", "supports": [ "recovery", "deployment" ], "title": "CrewAI production architecture", "url": "https://github.com/crewAIInc/crewAI/blob/main/docs/v1.12.2/en/concepts/production-architecture.mdx", "verifiedOn": "2026-08-28" }, { "framework": "openai-agents-sdk", "id": "OA-REL", "limitations": "This tag covers the Python SDK. Other OpenAI agent products and language SDKs have distinct release surfaces and are outside this version stamp.", "publisher": "OpenAI", "sourceType": "official-release", "supports": [ "release", "license" ], "title": "OpenAI Agents SDK Python 0.22.0 release", "url": "https://github.com/openai/openai-agents-python/releases/tag/v0.22.0", "verifiedOn": "2026-08-28" }, { "framework": "openai-agents-sdk", "id": "OA-RUN", "limitations": "Durable execution integrations add separate runtimes and operating models. The SDK alone does not operate their queues, storage, retries, or worker lifecycle.", "publisher": "OpenAI", "sourceType": "official-docs", "supports": [ "state", "recovery", "deployment" ], "title": "Running agents", "url": "https://openai.github.io/openai-agents-python/running_agents/", "verifiedOn": "2026-08-28" }, { "framework": "openai-agents-sdk", "id": "OA-HITL", "limitations": "Tool approval is a mechanism, not a complete governance policy. Identity, authorization, separation of duties, escalation, and audit retention remain application concerns.", "publisher": "OpenAI", "sourceType": "official-docs", "supports": [ "approval", "state" ], "title": "Human-in-the-loop", "url": "https://openai.github.io/openai-agents-python/human_in_the_loop/", "verifiedOn": "2026-08-28" }, { "framework": "openai-agents-sdk", "id": "OA-TRACE", "limitations": "Trace content can include sensitive data depending on configuration. Enterprises must set capture, export, retention, and access policies before production use.", "publisher": "OpenAI", "sourceType": "official-docs", "supports": [ "observability" ], "title": "Tracing", "url": "https://openai.github.io/openai-agents-python/tracing/", "verifiedOn": "2026-08-28" }, { "framework": "google-adk", "id": "GA-REL", "limitations": "This tag stamps the Python SDK. TypeScript, Go, Java, and Kotlin bindings are separate packages and may expose different features or release numbers.", "publisher": "Google", "sourceType": "official-release", "supports": [ "release", "license" ], "title": "Google ADK Python 2.8.0 release", "url": "https://github.com/google/adk-python/releases/tag/v2.8.0", "verifiedOn": "2026-08-28" }, { "framework": "google-adk", "id": "GA-STATE", "limitations": "Persistence depends on SessionService. In-memory state is explicitly non-durable and cannot support process-restart recovery.", "publisher": "Google", "sourceType": "official-docs", "supports": [ "state", "recovery" ], "title": "ADK session state", "url": "https://google.github.io/adk-docs/sessions/state/", "verifiedOn": "2026-08-28" }, { "framework": "google-adk", "id": "GA-HITL", "limitations": "Human input primitives need application-specific identity, authorization, UI, timeout, and escalation handling to become an enterprise approval control.", "publisher": "Google", "sourceType": "official-docs", "supports": [ "approval" ], "title": "ADK graph human input", "url": "https://google.github.io/adk-docs/graphs/human-input/", "verifiedOn": "2026-08-28" }, { "framework": "google-adk", "id": "GA-RESUME", "limitations": "Resume behavior depends on a resumable workflow and persistent session implementation. Teams must test replay behavior around external side effects.", "publisher": "Google", "sourceType": "official-docs", "supports": [ "approval", "recovery" ], "title": "ADK runtime resume", "url": "https://google.github.io/adk-docs/runtime/resume/", "verifiedOn": "2026-08-28" }, { "framework": "google-adk", "id": "GA-OBS", "limitations": "Cloud Trace and analytics behavior depends on deployment and capture configuration. Full prompt-response logging has additional privacy and storage implications.", "publisher": "Google", "sourceType": "official-docs", "supports": [ "observability" ], "title": "ADK observability", "url": "https://google.github.io/adk-docs/observability/", "verifiedOn": "2026-08-28" }, { "framework": "google-adk", "id": "GA-DEPLOY", "limitations": "Agent Runtime, Cloud Run, and GKE have different cost, identity, networking, scaling, and operational boundaries; listing them is not a recommendation.", "publisher": "Google", "sourceType": "official-docs", "supports": [ "deployment" ], "title": "ADK deployment options", "url": "https://google.github.io/adk-docs/deploy/", "verifiedOn": "2026-08-28" } ] } }, "rag": { "claims": { "claims": [ { "id": "CLAIM-RAG-RUNTIME-RETRIEVAL", "sourceIds": [ "SOURCE-RAG-PAPER", "SOURCE-MICROSOFT-RAG-VS-FT" ], "statement": "RAG supplies retrieved external information to a generator at inference time instead of encoding every changing fact in model parameters.", "status": "VERIFIED" }, { "id": "CLAIM-RAG-FRESHNESS-FIT", "sourceIds": [ "SOURCE-MICROSOFT-RAG-VS-FT", "SOURCE-PHAROS-GUIDE" ], "statement": "A retrieval path is a suitable candidate when the required knowledge changes independently of the model and the corpus can be maintained.", "status": "BOUNDED" }, { "id": "CLAIM-RAG-PROVENANCE", "sourceIds": [ "SOURCE-RAG-PAPER", "SOURCE-ANTHROPIC-CONTEXTUAL-RETRIEVAL" ], "statement": "Retrieved passages can provide provenance for an answer, but the application must still test whether each citation supports the generated claim.", "status": "BOUNDED" }, { "id": "CLAIM-FINE-TUNING-BEHAVIOR", "sourceIds": [ "SOURCE-MICROSOFT-RAG-VS-FT", "SOURCE-OPENAI-MODEL-OPTIMIZATION" ], "statement": "Fine-tuning is an optimization path for task behavior, format, or specialization; it is not a dependable mechanism for continuously changing facts.", "status": "VERIFIED" }, { "id": "CLAIM-EXAMPLES-REQUIRED", "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION", "SOURCE-MICROSOFT-RAG-VS-FT" ], "statement": "A fine-tuning decision needs representative examples and an evaluation set that reflects the intended task.", "status": "VERIFIED" }, { "id": "CLAIM-EVAL-FIRST", "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION", "SOURCE-PHAROS-GUIDE" ], "statement": "Prompting, retrieval, and training changes should be compared against a defined evaluation baseline rather than selected from an architecture label alone.", "status": "BOUNDED" }, { "id": "CLAIM-LEXICAL-SEMANTIC", "sourceIds": [ "SOURCE-ANTHROPIC-CONTEXTUAL-RETRIEVAL" ], "statement": "Lexical retrieval such as BM25 and embedding retrieval recover different signals and can be combined when exact terms and semantic similarity both matter.", "status": "VERIFIED" }, { "id": "CLAIM-RERANKING", "sourceIds": [ "SOURCE-ANTHROPIC-CONTEXTUAL-RETRIEVAL" ], "statement": "A reranker can reorder an initial retrieval candidate set before context is passed to the model.", "status": "VERIFIED" }, { "id": "CLAIM-CORPUS-QUALITY", "sourceIds": [ "SOURCE-MICROSOFT-RAG-VS-FT", "SOURCE-PHAROS-GUIDE" ], "statement": "A RAG recommendation presumes an authoritative, permissioned corpus and an operating path for indexing and updates.", "status": "BOUNDED" }, { "id": "CLAIM-LORA-METHOD", "sourceIds": [ "SOURCE-LORA-PAPER" ], "statement": "LoRA is a parameter-efficient adaptation method that freezes pretrained weights and learns low-rank update matrices.", "status": "VERIFIED" }, { "id": "CLAIM-HYBRID-PATTERN", "sourceIds": [ "SOURCE-RAFT-PAPER", "SOURCE-MICROSOFT-RAG-VS-FT" ], "statement": "Retrieval and parameter adaptation can be combined when a workload separately requires external domain evidence and specialized model behavior.", "status": "BOUNDED" }, { "id": "CLAIM-RAFT-SCOPE", "sourceIds": [ "SOURCE-RAFT-PAPER" ], "statement": "RAFT is one retrieval-aware fine-tuning method that trains with relevant documents and distractor documents for domain-specific RAG settings.", "status": "VERIFIED" }, { "id": "CLAIM-DUAL-OPERATIONS", "sourceIds": [ "SOURCE-RAFT-PAPER", "SOURCE-PHAROS-GUIDE" ], "statement": "A hybrid design creates both retrieval-index and model-version lifecycle work, so both operating capabilities must be assessed.", "status": "BOUNDED" }, { "id": "CLAIM-PRIVACY-TOPOLOGY", "sourceIds": [ "SOURCE-OPENAI-DATA-CONTROLS", "SOURCE-MICROSOFT-RAG-VS-FT" ], "statement": "Privacy and retention properties depend on the selected provider, endpoint, storage path, deployment topology, and configuration; neither architecture has one universal data boundary.", "status": "BOUNDED" }, { "id": "CLAIM-RUNTIME-MEASUREMENT", "sourceIds": [ "SOURCE-MICROSOFT-RAG-VS-FT", "SOURCE-OPENAI-MODEL-OPTIMIZATION" ], "statement": "Latency, quality, and cost tradeoffs must be measured on the target workload and deployment rather than inferred from RAG or fine-tuning labels.", "status": "BOUNDED" } ], "version": "2026-08-18.1" }, "criteria": { "answerSemantics": { "no": "Evidence contradicts the condition for the target workload.", "unknown": "The condition has not been verified; the engine will not infer an answer.", "yes": "Evidence supports the condition for the target workload." }, "criteria": [ { "allowedAnswers": [ "yes", "no", "unknown" ], "claimIds": [ "CLAIM-RAG-FRESHNESS-FIT" ], "core": true, "dimension": "knowledge", "id": "fresh_facts", "label": "Changing knowledge", "question": "Must answers use facts that change independently of the model release?", "sourceIds": [ "SOURCE-MICROSOFT-RAG-VS-FT" ] }, { "allowedAnswers": [ "yes", "no", "unknown" ], "claimIds": [ "CLAIM-RAG-PROVENANCE" ], "core": true, "dimension": "knowledge", "id": "source_citations", "label": "Passage-level evidence", "question": "Must users inspect source passages behind material answer claims?", "sourceIds": [ "SOURCE-RAG-PAPER" ] }, { "allowedAnswers": [ "yes", "no", "unknown" ], "claimIds": [ "CLAIM-FINE-TUNING-BEHAVIOR" ], "core": true, "dimension": "behavior", "id": "behavior_change", "label": "Behavior specialization", "question": "Is the primary gap consistent task behavior, style, format, classification, or extraction?", "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION" ] }, { "allowedAnswers": [ "yes", "no", "unknown" ], "claimIds": [ "CLAIM-EXAMPLES-REQUIRED" ], "core": true, "dimension": "data", "id": "representative_examples", "label": "Representative examples", "question": "Are representative, permitted examples available for training and held-out evaluation?", "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION" ] }, { "allowedAnswers": [ "yes", "no", "unknown" ], "claimIds": [ "CLAIM-EVAL-FIRST" ], "core": true, "dimension": "evaluation", "id": "evaluation_set", "label": "Evaluation set", "question": "Is there a versioned evaluation set with acceptance criteria for this workload?", "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION" ] }, { "allowedAnswers": [ "yes", "no", "unknown" ], "claimIds": [ "CLAIM-CORPUS-QUALITY" ], "core": true, "dimension": "data", "id": "authoritative_corpus", "label": "Authoritative corpus", "question": "Is an authoritative, permissioned, maintainable corpus available for retrieval?", "sourceIds": [ "SOURCE-MICROSOFT-RAG-VS-FT" ] }, { "allowedAnswers": [ "yes", "no", "unknown" ], "claimIds": [ "CLAIM-LEXICAL-SEMANTIC" ], "core": false, "dimension": "retrieval", "id": "exact_term_retrieval", "label": "Exact-term retrieval", "question": "Do identifiers, codes, names, or exact phrases materially affect retrieval quality?", "sourceIds": [ "SOURCE-ANTHROPIC-CONTEXTUAL-RETRIEVAL" ] }, { "allowedAnswers": [ "yes", "no", "unknown" ], "claimIds": [ "CLAIM-RUNTIME-MEASUREMENT" ], "core": false, "dimension": "operations", "id": "strict_runtime_budget", "label": "Strict runtime budget", "question": "Does the workload have a strict measured latency or per-request cost budget?", "sourceIds": [ "SOURCE-MICROSOFT-RAG-VS-FT" ] }, { "allowedAnswers": [ "yes", "no", "unknown" ], "claimIds": [ "CLAIM-PRIVACY-TOPOLOGY" ], "core": false, "dimension": "governance", "id": "privacy_topology", "label": "Verified privacy topology", "question": "Have storage, retrieval, training, retention, and provider data paths been mapped?", "sourceIds": [ "SOURCE-OPENAI-DATA-CONTROLS" ] }, { "allowedAnswers": [ "yes", "no", "unknown" ], "claimIds": [ "CLAIM-CORPUS-QUALITY" ], "core": true, "dimension": "operations", "id": "retrieval_operations", "label": "Retrieval operations", "question": "Can the team own corpus ingestion, index updates, retrieval evaluation, and rollback?", "sourceIds": [ "SOURCE-MICROSOFT-RAG-VS-FT" ] }, { "allowedAnswers": [ "yes", "no", "unknown" ], "claimIds": [ "CLAIM-EXAMPLES-REQUIRED", "CLAIM-EVAL-FIRST" ], "core": true, "dimension": "operations", "id": "training_operations", "label": "Training operations", "question": "Can the team own training data, model versions, regression evaluation, and rollback?", "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION" ] }, { "allowedAnswers": [ "yes", "no", "unknown" ], "claimIds": [ "CLAIM-FINE-TUNING-BEHAVIOR" ], "core": true, "dimension": "behavior", "id": "stable_narrow_task", "label": "Stable narrow task", "question": "Is the target behavior narrow enough and stable enough to maintain as a trained model contract?", "sourceIds": [ "SOURCE-MICROSOFT-RAG-VS-FT" ] } ], "version": "2026-08-18.1" }, "methodology": { "allowedAnswers": [ "yes", "no", "unknown" ], "decisionProcedure": [ "Validate that every registered criterion has exactly one allowed answer and reject unknown criterion IDs.", "Evaluate decision rules in ascending unique priority and select the first matching outcome.", "Evaluate advisory rules independently and attach their tests, risks, and provenance without changing the outcome.", "Export the answers, winning rule, triggered advisories, blockers, unresolved inputs, source IDs, claim IDs, and registry versions." ], "disclosure": "Research synthesis and implementation were AI-assisted; claims are bounded to the listed sources and verification date.", "evaluationContract": { "minimumJobs": [ "Define representative tasks and failure classes.", "Record a prompt-only or current-system baseline.", "For retrieval, test retrieval relevance and claim-to-passage support separately from final-answer quality.", "For tuning, test target behavior and regression on capabilities that must be preserved.", "Measure latency, cost, privacy paths, and operational recovery in the intended topology." ], "requiredBeforeWinner": true }, "limitations": [ "The outcome is a candidate for a bounded experiment, not an implementation guarantee.", "The method contains no universal dataset-size, accuracy, latency, document-count, or cost threshold.", "Provider features and data controls can change after the dated source verification.", "Independent human technical review of this microsite is not claimed." ], "name": "Evidence-bound RAG versus fine-tuning decision method", "outcomes": [ { "id": "RAG", "label": "RAG candidate", "meaning": "Test a retrieval-grounded baseline because changing or traceable knowledge is required and the corpus and retrieval operating path exist." }, { "id": "FINE_TUNING", "label": "Fine-tuning candidate", "meaning": "Test parameter adaptation because the primary need is stable task behavior and representative examples and training operations exist." }, { "id": "HYBRID", "label": "Hybrid candidate", "meaning": "Test retrieval plus behavior adaptation because both needs and both operational paths are independently evidenced." }, { "id": "BASELINE_FIRST", "label": "Baseline first", "meaning": "Establish or improve a prompting/evaluation baseline before adding retrieval or training complexity." }, { "id": "BLOCKED", "label": "Blocked", "meaning": "A required data or operating dependency is absent; a plausible architecture label cannot remove that blocker." }, { "id": "INSUFFICIENT_EVIDENCE", "label": "Insufficient evidence", "meaning": "A core input or evaluation contract is unresolved, so the method declines to select a winner." } ], "priorityDirection": "ascending", "publishedOn": "2026-08-18", "publisher": "Pharos Production", "purpose": "Classify the next architecture experiment without converting qualitative evidence into a probability or universal score.", "version": "2026-08-18.1" }, "rules": { "rules": [ { "claimIds": [ "CLAIM-RAG-FRESHNESS-FIT", "CLAIM-CORPUS-QUALITY" ], "disqualifiers": [ "NO_AUTHORITATIVE_CORPUS" ], "id": "RULE-BLOCKED-NO-CORPUS", "kind": "decision", "nextSteps": [ "Identify an authoritative, permissioned corpus and an owner for updates before selecting a knowledge architecture." ], "outcomeId": "BLOCKED", "priority": 10, "rejectedOutcomes": [ "RAG", "HYBRID", "FINE_TUNING" ], "sourceIds": [ "SOURCE-MICROSOFT-RAG-VS-FT", "SOURCE-RAG-PAPER" ], "when": { "all": [ { "criterionId": "authoritative_corpus", "operator": "equals", "value": "no" } ], "any": [ { "criterionId": "fresh_facts", "operator": "equals", "value": "yes" }, { "criterionId": "source_citations", "operator": "equals", "value": "yes" } ] } }, { "claimIds": [ "CLAIM-EVAL-FIRST" ], "id": "RULE-INSUFFICIENT-UNKNOWN-CORE", "kind": "decision", "nextSteps": [ "Resolve every unknown core input with a named evidence owner before forcing an architecture choice." ], "outcomeId": "INSUFFICIENT_EVIDENCE", "priority": 20, "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION" ], "when": { "any": [ { "criterionId": "fresh_facts", "operator": "equals", "value": "unknown" }, { "criterionId": "source_citations", "operator": "equals", "value": "unknown" }, { "criterionId": "behavior_change", "operator": "equals", "value": "unknown" }, { "criterionId": "representative_examples", "operator": "equals", "value": "unknown" }, { "criterionId": "evaluation_set", "operator": "equals", "value": "unknown" }, { "criterionId": "authoritative_corpus", "operator": "equals", "value": "unknown" }, { "criterionId": "retrieval_operations", "operator": "equals", "value": "unknown" }, { "criterionId": "training_operations", "operator": "equals", "value": "unknown" }, { "criterionId": "stable_narrow_task", "operator": "equals", "value": "unknown" } ] } }, { "claimIds": [ "CLAIM-EVAL-FIRST", "CLAIM-EXAMPLES-REQUIRED" ], "disqualifiers": [ "NO_EVALUATION_SET" ], "id": "RULE-INSUFFICIENT-NO-EVAL", "kind": "decision", "nextSteps": [ "Create a versioned task evaluation set and acceptance criteria, then compare candidates against the same baseline." ], "outcomeId": "INSUFFICIENT_EVIDENCE", "priority": 30, "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION" ], "when": { "all": [ { "criterionId": "evaluation_set", "operator": "equals", "value": "no" } ] } }, { "claimIds": [ "CLAIM-EXAMPLES-REQUIRED", "CLAIM-EVAL-FIRST" ], "id": "RULE-BASELINE-NO-EXAMPLES", "kind": "decision", "nextSteps": [ "Establish a prompt baseline and collect permitted, representative failure and success examples before training." ], "outcomeId": "BASELINE_FIRST", "priority": 40, "rejectedOutcomes": [ "FINE_TUNING", "HYBRID" ], "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION" ], "when": { "all": [ { "criterionId": "behavior_change", "operator": "equals", "value": "yes" }, { "criterionId": "representative_examples", "operator": "equals", "value": "no" } ] } }, { "claimIds": [ "CLAIM-CORPUS-QUALITY" ], "disqualifiers": [ "NO_RETRIEVAL_OPERATING_PATH" ], "id": "RULE-BLOCKED-NO-RETRIEVAL-OPS", "kind": "decision", "nextSteps": [ "Assign ownership for ingestion, index freshness, retrieval evaluation, rollback, and access control." ], "outcomeId": "BLOCKED", "priority": 50, "rejectedOutcomes": [ "RAG", "HYBRID" ], "sourceIds": [ "SOURCE-MICROSOFT-RAG-VS-FT" ], "when": { "all": [ { "criterionId": "authoritative_corpus", "operator": "equals", "value": "yes" }, { "criterionId": "retrieval_operations", "operator": "equals", "value": "no" } ], "any": [ { "criterionId": "fresh_facts", "operator": "equals", "value": "yes" }, { "criterionId": "source_citations", "operator": "equals", "value": "yes" } ] } }, { "claimIds": [ "CLAIM-EXAMPLES-REQUIRED", "CLAIM-EVAL-FIRST" ], "disqualifiers": [ "NO_TRAINING_OPERATING_PATH" ], "id": "RULE-BLOCKED-NO-TRAINING-OPS", "kind": "decision", "nextSteps": [ "Assign ownership for datasets, model versions, regression evaluation, rollback, and provider lifecycle." ], "outcomeId": "BLOCKED", "priority": 60, "rejectedOutcomes": [ "FINE_TUNING", "HYBRID" ], "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION" ], "when": { "all": [ { "criterionId": "behavior_change", "operator": "equals", "value": "yes" }, { "criterionId": "representative_examples", "operator": "equals", "value": "yes" }, { "criterionId": "training_operations", "operator": "equals", "value": "no" } ] } }, { "claimIds": [ "CLAIM-HYBRID-PATTERN", "CLAIM-DUAL-OPERATIONS", "CLAIM-RAFT-SCOPE" ], "id": "RULE-HYBRID-CANDIDATE", "kind": "decision", "nextSteps": [ "Test retrieval and behavior adaptation separately before testing the combined system; retain both rollback paths." ], "outcomeId": "HYBRID", "priority": 70, "sourceIds": [ "SOURCE-RAFT-PAPER", "SOURCE-MICROSOFT-RAG-VS-FT" ], "when": { "all": [ { "criterionId": "behavior_change", "operator": "equals", "value": "yes" }, { "criterionId": "representative_examples", "operator": "equals", "value": "yes" }, { "criterionId": "authoritative_corpus", "operator": "equals", "value": "yes" }, { "criterionId": "retrieval_operations", "operator": "equals", "value": "yes" }, { "criterionId": "training_operations", "operator": "equals", "value": "yes" }, { "criterionId": "stable_narrow_task", "operator": "equals", "value": "yes" } ], "any": [ { "criterionId": "fresh_facts", "operator": "equals", "value": "yes" }, { "criterionId": "source_citations", "operator": "equals", "value": "yes" } ] } }, { "claimIds": [ "CLAIM-RAG-RUNTIME-RETRIEVAL", "CLAIM-RAG-FRESHNESS-FIT", "CLAIM-RAG-PROVENANCE" ], "id": "RULE-RAG-CANDIDATE", "kind": "decision", "nextSteps": [ "Build a retrieval baseline and evaluate passage relevance, claim support, answer quality, freshness, and recovery from stale indexes." ], "outcomeId": "RAG", "priority": 80, "sourceIds": [ "SOURCE-RAG-PAPER", "SOURCE-MICROSOFT-RAG-VS-FT" ], "when": { "all": [ { "criterionId": "authoritative_corpus", "operator": "equals", "value": "yes" }, { "criterionId": "retrieval_operations", "operator": "equals", "value": "yes" } ], "any": [ { "criterionId": "fresh_facts", "operator": "equals", "value": "yes" }, { "criterionId": "source_citations", "operator": "equals", "value": "yes" } ] } }, { "claimIds": [ "CLAIM-FINE-TUNING-BEHAVIOR", "CLAIM-EXAMPLES-REQUIRED", "CLAIM-LORA-METHOD" ], "id": "RULE-FINE-TUNING-CANDIDATE", "kind": "decision", "nextSteps": [ "Compare prompt-only and fine-tuned candidates on target behavior, held-out examples, regressions, and rollback readiness." ], "outcomeId": "FINE_TUNING", "priority": 90, "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION", "SOURCE-MICROSOFT-RAG-VS-FT", "SOURCE-LORA-PAPER" ], "when": { "all": [ { "criterionId": "behavior_change", "operator": "equals", "value": "yes" }, { "criterionId": "representative_examples", "operator": "equals", "value": "yes" }, { "criterionId": "training_operations", "operator": "equals", "value": "yes" }, { "criterionId": "stable_narrow_task", "operator": "equals", "value": "yes" } ], "none": [ { "criterionId": "fresh_facts", "operator": "equals", "value": "yes" }, { "criterionId": "source_citations", "operator": "equals", "value": "yes" } ] } }, { "claimIds": [ "CLAIM-EVAL-FIRST" ], "id": "RULE-BASELINE-NO-PRIMARY-GAP", "kind": "decision", "nextSteps": [ "Measure the current prompt or workflow baseline and document a specific failure class before adding architecture." ], "outcomeId": "BASELINE_FIRST", "priority": 100, "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION" ], "when": { "all": [ { "criterionId": "fresh_facts", "operator": "equals", "value": "no" }, { "criterionId": "source_citations", "operator": "equals", "value": "no" }, { "criterionId": "behavior_change", "operator": "equals", "value": "no" } ] } }, { "claimIds": [ "CLAIM-EVAL-FIRST" ], "id": "RULE-INSUFFICIENT-FALLBACK", "kind": "decision", "nextSteps": [ "Document which requirement, data asset, or operating capability prevents the available evidence from matching a candidate rule." ], "outcomeId": "INSUFFICIENT_EVIDENCE", "priority": 110, "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION", "SOURCE-PHAROS-GUIDE" ], "when": { "all": [ { "criterionId": "fresh_facts", "operator": "not_equals", "value": "unknown" } ] } }, { "claimIds": [ "CLAIM-LEXICAL-SEMANTIC", "CLAIM-RERANKING" ], "id": "ADVISORY-LEXICAL-RETRIEVAL", "kind": "advisory", "nextSteps": [ "Benchmark BM25 or another lexical path alongside embeddings, then evaluate fusion and reranking on exact-term failures." ], "priority": 200, "sourceIds": [ "SOURCE-ANTHROPIC-CONTEXTUAL-RETRIEVAL" ], "when": { "all": [ { "criterionId": "exact_term_retrieval", "operator": "equals", "value": "yes" } ] } }, { "claimIds": [ "CLAIM-RAG-PROVENANCE" ], "id": "ADVISORY-CITATION-FAITHFULNESS", "kind": "advisory", "nextSteps": [ "Evaluate whether each material generated claim is supported by its cited passage; do not treat citation presence as faithfulness." ], "priority": 210, "sourceIds": [ "SOURCE-RAG-PAPER", "SOURCE-ANTHROPIC-CONTEXTUAL-RETRIEVAL" ], "when": { "all": [ { "criterionId": "source_citations", "operator": "equals", "value": "yes" } ] } }, { "claimIds": [ "CLAIM-RUNTIME-MEASUREMENT" ], "id": "ADVISORY-RUNTIME-BENCHMARK", "kind": "advisory", "nextSteps": [ "Benchmark end-to-end latency and cost on the intended provider, corpus, traffic shape, and quality target." ], "priority": 220, "sourceIds": [ "SOURCE-MICROSOFT-RAG-VS-FT", "SOURCE-OPENAI-MODEL-OPTIMIZATION" ], "when": { "all": [ { "criterionId": "strict_runtime_budget", "operator": "equals", "value": "yes" } ] } }, { "claimIds": [ "CLAIM-PRIVACY-TOPOLOGY" ], "id": "ADVISORY-PRIVACY-TOPOLOGY", "kind": "advisory", "nextSteps": [ "Retain the verified data-flow map and recheck provider retention, training, access, and deletion controls before release." ], "priority": 230, "sourceIds": [ "SOURCE-OPENAI-DATA-CONTROLS" ], "when": { "all": [ { "criterionId": "privacy_topology", "operator": "equals", "value": "yes" } ] } }, { "claimIds": [ "CLAIM-PRIVACY-TOPOLOGY" ], "id": "ADVISORY-PRIVACY-UNRESOLVED", "kind": "advisory", "nextSteps": [ "Map storage, retrieval, training, retention, access, and deletion paths; architecture labels alone do not establish privacy." ], "priority": 240, "sourceIds": [ "SOURCE-OPENAI-DATA-CONTROLS", "SOURCE-MICROSOFT-RAG-VS-FT" ], "when": { "any": [ { "criterionId": "privacy_topology", "operator": "equals", "value": "no" }, { "criterionId": "privacy_topology", "operator": "equals", "value": "unknown" } ] } }, { "claimIds": [ "CLAIM-HYBRID-PATTERN", "CLAIM-DUAL-OPERATIONS" ], "id": "ADVISORY-HYBRID-OPERATIONS", "kind": "advisory", "nextSteps": [ "Version the corpus, index, training dataset, base model, adapter or tuned model, and evaluation suite as separate release inputs." ], "priority": 250, "sourceIds": [ "SOURCE-RAFT-PAPER" ], "when": { "all": [ { "criterionId": "retrieval_operations", "operator": "equals", "value": "yes" }, { "criterionId": "training_operations", "operator": "equals", "value": "yes" }, { "criterionId": "behavior_change", "operator": "equals", "value": "yes" } ], "any": [ { "criterionId": "fresh_facts", "operator": "equals", "value": "yes" }, { "criterionId": "source_citations", "operator": "equals", "value": "yes" } ] } } ], "version": "2026-08-18.1" }, "scenarios": { "scenarios": [ { "candidateOutcome": "RAG", "claimIds": [ "CLAIM-RAG-FRESHNESS-FIT", "CLAIM-RAG-PROVENANCE" ], "evidencePattern": [ "fresh_facts=yes", "source_citations=yes", "authoritative_corpus=yes", "retrieval_operations=yes", "evaluation_set=yes" ], "firstExperiment": "Compare a prompt-only baseline with retrieval and score passage relevance, claim support, answer quality, and stale-index failures.", "id": "SCENARIO-CHANGING-KNOWLEDGE-BASE", "situation": "Support answers must reflect maintained policies and link the user to the passages used.", "sourceIds": [ "SOURCE-RAG-PAPER", "SOURCE-MICROSOFT-RAG-VS-FT" ], "title": "Changing support knowledge base", "why": "The primary gap is a maintained knowledge path with traceability, not a new behavioral contract." }, { "candidateOutcome": "FINE_TUNING", "claimIds": [ "CLAIM-FINE-TUNING-BEHAVIOR", "CLAIM-EXAMPLES-REQUIRED", "CLAIM-LORA-METHOD" ], "evidencePattern": [ "behavior_change=yes", "representative_examples=yes", "training_operations=yes", "stable_narrow_task=yes", "evaluation_set=yes" ], "firstExperiment": "Compare prompt-only and tuned candidates on held-out schema validity, task errors, regressions, and operational rollback.", "id": "SCENARIO-STABLE-STRUCTURED-TASK", "situation": "A stable input class must produce a consistent schema, and representative permitted examples already exist.", "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION", "SOURCE-LORA-PAPER" ], "title": "Stable structured extraction task", "why": "The target is repeatable task behavior rather than changing external facts." }, { "candidateOutcome": "HYBRID", "claimIds": [ "CLAIM-HYBRID-PATTERN", "CLAIM-DUAL-OPERATIONS" ], "evidencePattern": [ "fresh_facts=yes", "behavior_change=yes", "representative_examples=yes", "authoritative_corpus=yes", "retrieval_operations=yes", "training_operations=yes", "stable_narrow_task=yes", "evaluation_set=yes" ], "firstExperiment": "Test retrieval and behavior adaptation separately; combine them only if each produces a measured improvement without unacceptable regressions.", "id": "SCENARIO-DOMAIN-ASSISTANT", "situation": "The assistant must use a changing domain corpus and also follow a stable response or reasoning contract represented in examples.", "sourceIds": [ "SOURCE-RAFT-PAPER", "SOURCE-MICROSOFT-RAG-VS-FT" ], "title": "Domain assistant with specialized behavior", "why": "The knowledge and behavior gaps are independent and both lifecycle capabilities are present." }, { "candidateOutcome": "INSUFFICIENT_EVIDENCE", "claimIds": [ "CLAIM-EVAL-FIRST" ], "evidencePattern": [ "evaluation_set=no" ], "firstExperiment": "Create the evaluation set and baseline before implementing either architecture.", "id": "SCENARIO-NO-EVALUATION", "situation": "Stakeholders request RAG or tuning, but no versioned tasks or acceptance criteria distinguish a real improvement from a demo.", "sourceIds": [ "SOURCE-OPENAI-MODEL-OPTIMIZATION" ], "title": "Architecture request without an evaluation set", "why": "The tool declines to choose when alternatives cannot be measured against the same workload contract." } ], "version": "2026-08-18.1" }, "sources": { "sources": [ { "id": "SOURCE-PHAROS-GUIDE", "published": "2026-03-30", "publisher": "Pharos Production", "title": "RAG vs Fine-Tuning: When to Use Each for AI Projects", "type": "practitioner-guide", "updated": "2026-06-29", "url": "https://pharosproduction.com/insights/engineering/rag-vs-fine-tuning/", "useBoundary": "Practitioner framing only; quantitative claims are not reused without primary support.", "verificationStatus": "VERIFIED", "verifiedOn": "2026-08-18" }, { "id": "SOURCE-RAG-PAPER", "published": "2020-05-22", "publisher": "arXiv", "title": "Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks", "type": "research-paper", "url": "https://arxiv.org/abs/2005.11401", "useBoundary": "Supports the RAG architecture definition and provenance-oriented claims, not universal production outcomes.", "verificationStatus": "VERIFIED", "verifiedOn": "2026-08-18" }, { "id": "SOURCE-MICROSOFT-RAG-VS-FT", "publisher": "Microsoft Learn", "title": "Augment language models with RAG or fine-tuning", "type": "official-documentation", "updated": "2026-01-30", "url": "https://learn.microsoft.com/en-us/azure/developer/ai/augment-llm-rag-fine-tuning", "useBoundary": "Supports qualitative selection factors; vendor examples are not universal thresholds.", "verificationStatus": "VERIFIED", "verifiedOn": "2026-08-18" }, { "id": "SOURCE-ANTHROPIC-CONTEXTUAL-RETRIEVAL", "published": "2024-09-19", "publisher": "Anthropic", "title": "Introducing Contextual Retrieval", "type": "official-engineering-article", "url": "https://www.anthropic.com/engineering/contextual-retrieval", "useBoundary": "Supports hybrid lexical-semantic retrieval and reranking patterns; reported experiments remain provider-specific.", "verificationStatus": "VERIFIED", "verifiedOn": "2026-08-18" }, { "id": "SOURCE-OPENAI-MODEL-OPTIMIZATION", "publisher": "OpenAI", "title": "Model optimization", "type": "official-documentation", "url": "https://developers.openai.com/api/docs/guides/model-optimization", "useBoundary": "Supports an eval-first optimization loop and fine-tuning with representative examples; product availability can change.", "verificationStatus": "VERIFIED", "verifiedOn": "2026-08-18" }, { "id": "SOURCE-LORA-PAPER", "published": "2021-06-17", "publisher": "arXiv", "title": "LoRA: Low-Rank Adaptation of Large Language Models", "type": "research-paper", "url": "https://arxiv.org/abs/2106.09685", "useBoundary": "Supports the LoRA method definition; benchmark results are not generalized to every model or workload.", "verificationStatus": "VERIFIED", "verifiedOn": "2026-08-18" }, { "id": "SOURCE-RAFT-PAPER", "published": "2024-03-15", "publisher": "arXiv", "title": "RAFT: Adapting Language Model to Domain Specific RAG", "type": "research-paper", "url": "https://arxiv.org/abs/2403.10131", "useBoundary": "Supports a specific retrieval-aware fine-tuning pattern, not a universal claim that hybrid systems outperform.", "verificationStatus": "VERIFIED", "verifiedOn": "2026-08-18" }, { "id": "SOURCE-OPENAI-DATA-CONTROLS", "publisher": "OpenAI", "title": "Data controls in the OpenAI platform", "type": "official-documentation", "url": "https://developers.openai.com/api/docs/guides/your-data#default-usage-policies-by-endpoint", "useBoundary": "Used only to show that retention and training treatment depend on service and endpoint configuration.", "verificationStatus": "VERIFIED", "verifiedOn": "2026-08-18" } ], "verifiedOn": "2026-08-18", "version": "2026-08-18.1" } } }, "release_version": "0.1.0-candidate.1", "review_status": "NOT_REVIEWED", "schema_version": "2.0.0-candidate.1" }