Specira sample artefact. Rendered from the governed default template on a fictional company. Names, figures and dates are illustrative.All artefacts →
SAMPLE
seeded demo data · specira.ai
Specira AI Orchestration Blueprint: Governed Template Rendering
Governed template rendering reference: AI Orchestration Blueprint default v2 (draft) definition 2d63546b…e56b
AI Orchestration Blueprint · Project artifact SPECIRA

AI Orchestration Blueprint: Dispatch Modernization

Meridian Field Services: instructional example, not project evidence

Draft · watermark policy: draft_only template ai_orchestration_blueprint v2 · pack: specira_default_delivery workflows beat agents until proven; the justification cuts both ways
§1

Agent Architecture

mandatory 1 decision1 evidence rule validators: every_feature_has_a_named_orchestration_pattern · every_agent_or_multiagent_choice_carries_a_complexity_justification · every_guardrail_layer_is_mapped_to_input_or_output_stage

Cheapest pattern first: single call, chain, router, parallelization, orchestrator-workers, evaluator-optimizer, agent, multi-agent. Anything agentic carries a written justification; so does the rejection.

FeaturePatternModelRationaleFallback
Suggestion ranking single constrained call ranking role[1] One bounded task (rank the eligible set), no tool use, no iteration Full-list degradation (suggestions_unavailable[2])
Exception-review search single retrieval call none (retrieval-only at pilot) The US-3 surface over the override-note index (retrieval[3]) Keyword filter fallback
Complexity justification, cutting both ways

An agent framework was REJECTED for suggestion ranking: the task has fixed inputs, a closed output space, and a two-second budget; agent tiers trade latency and cost for autonomy the task cannot use; multi-agent burns 4 to 220x tokens on work this shape (orchestration pattern[4]).

Guardrails, two layers. Input: the intake validation rules (validation[5]). Output: schema-constrained ranked ids from the eligible set only; constrained decoding per ADR-5[4], the strongest output guardrail. Observability: every call a span with model, input and output tokens, operation, and provider, landing in the LLM trace stream (instrumentation[6]).

§2

LLM Provider Strategy

mandatory 1 decision1 evidence rule validators: every_provider_dependency_has_a_fallback_or_documented_spof · provider_egress_cites_security_and_classification
Decision: single provider behind the platform gateway

Why considered: one AI feature at pilot; a second provider's resilience buys less than its gateway-maintenance and eval-duplication costs; direct integration hard-wires the vendor shape into application code. Findings: single-provider-with-gateway wins; the gateway is the egress enforcement point; anonymized codes only cross (SR-014[7], INT-4[8]); switching exposure stated: prompts and evals are provider-portable by design, the gateway isolates the API shape (provider strategy[4]).

SPOF, documented and owned: provider outage degrades suggestions to the full list (never blocks dispatch); accepted by A. Reyes; revisit trigger is a second AI feature entering the roadmap.

§3

Prompt Design Patterns

mandatory 1 decision1 evidence rule validators: no_prompt_content_in_blueprint · constrained_outputs_cite_their_decision

Orchestration-level shape only; the prompt CONTENT lives in the prompt spec, cited never restated.

The ranking call carries the system framing at the gateway layer; job attributes and candidate codes enter as typed variables; the instruction hierarchy is architectural: the system framing is not overridable by job-note content (injection defenses[9]). Structured output: the ranking feature feeds a machine (the board), so schema-constrained decoding is enforced at the gateway, citing ADR-5[4]; no free-text output path exists at pilot. Prompt content: one entry, PR-01[9].

§4

State Management

mandatory 1 decision1 evidence rule validators: every_checkpointed_workflow_declares_its_persistence_backend · stateless_features_carry_rationale

Statelessness is a claim with rationale; the absence of a checkpointer is a decision, not an omission.

Stateless, as a claim: suggestion ranking is stateless per-request; every call carries its full context (job attributes plus the eligible set), nothing resumes, nothing checkpoints. Rationale: the task completes inside one budget window and the board re-requests on staleness (board read model[10]). The exception-review search is likewise stateless. No checkpointer, no persistence backend, no conversation state, and the first conversational feature on the roadmap re-opens this section.

§5

Cost Architecture

mandatory 1 decision1 evidence rule validators: every_feature_has_a_cost_per_task_budget · budget_breach_rule_wired_to_observability

Cost per task is designed, not discovered on the invoice.

FeatureBudget / taskCachingMeasured actual
Suggestion ranking 1,800 tokens (1,500 in / 300 out) Prompt-prefix caching on the stable system framing (the eligible-set payload varies, the framing does not) 1,410 tokens mean over pilot-shape trials (eval runs[1])

Semantic caching: rejected at pilot: dispatch contexts rarely repeat within the cache window; recorded so the rejection is a decision. Cost per thousand suggestions is budgeted and tracked on the value-case dashboard (dashboards[6]). Breach rule: a day above 120% of budget raises a ticket to A. Reyes; the alert is wired (alert catalog[6]).

§6

Error Recovery & Fallbacks

mandatory 1 decision1 evidence rule validators: every_external_call_has_a_timeout_and_retry_budget · degrade_paths_cite_api_degradation_contracts
RungSuggestion ranking
Timeout 2 s, the boundary the process flow owns (FLW-01[11])
Retry budget One retry inside the window, none after; a late suggestion is discarded, never late-applied (AI cases[12])
Fallback chain DEGRADED, declared: the full technician list with the degradation notice (suggestions_unavailable[2])
Circuit breaker 5 consecutive provider failures → open 5 min; suggestions degrade without per-request timeouts stacking
Posture Degrade before error: dispatch never blocks on the model

The failure paths themselves are owned by the process flows and edge-case analysis; cited, not remodeled.

§7

Open Questions

mandatory1 decision
QuestionOwnerAnswer byBlocks
Does the exception-review search add a generation step (summarize similar overrides) or stay retrieval-only through pilot? A. ReyesSep 26, 2026 The pattern table's second row only
Refs

References & Package Contents

In the Specira workspace

specira [1] Model selection: the ranking role, eval runs app.specira.ai/projects/dispatch-modernization/artifacts/model-selection
specira [2] API contracts: suggestions_unavailable degradation contract app.specira.ai/projects/dispatch-modernization/artifacts/api-contracts#errors
specira [3] Embedding & vector strategy: the override-note retrieval app.specira.ai/projects/dispatch-modernization/artifacts/embedding-vector-strategy
specira [4] Decision log: orchestration pattern, provider strategy, ADR-5 app.specira.ai/projects/dispatch-modernization/artifacts/decision-log
specira [5] Business rules: validation family app.specira.ai/projects/dispatch-modernization/artifacts/business-rules#validation
specira [6] Observability: LLM instrumentation, dashboards, alert catalog app.specira.ai/projects/dispatch-modernization/artifacts/observability-monitoring
specira [7] Security requirements: SR-014 gateway-only egress app.specira.ai/projects/dispatch-modernization/artifacts/security-requirements#sr-014
specira [8] Integration inventory: INT-4 model provider app.specira.ai/projects/dispatch-modernization/artifacts/integration-inventory#int-4
specira [9] Prompt engineering spec: PR-01, injection defenses app.specira.ai/projects/dispatch-modernization/artifacts/prompt-engineering-spec
specira [10] Architecture: the board read model app.specira.ai/projects/dispatch-modernization/artifacts/architecture
specira [11] Process flows: FLW-01 boundary timer app.specira.ai/projects/dispatch-modernization/artifacts/process-flows#flw-01
specira [12] Edge case analysis: the AI families app.specira.ai/projects/dispatch-modernization/artifacts/edge-case-analysis#ai
Generated by Specira · template ai_orchestration_blueprint v2 (draft) · pack specira_default_delivery lineage 2d63546b…e56b · page 1 of 5