Trigger-to-Concept Cheat Sheet

Use this to recognise what an interviewer is testing and decide your next move. A trigger suggests a question or requirement; it does not automatically justify a component.

HEAR THE TRIGGER → RECALL THE CONCEPT
                           ↓
        CLARIFY → REQUIREMENT → CHOICE → TRADE-OFF

Answer pattern: “That suggests ___. I’d clarify ___ before choosing ___. The main trade-off is ___.”

Recall discovery: WHY → TODAY → WHO/HOW MUCH → WHAT → HOW FAR → BOUNDARIES.

API Gateway   = User → Application control
Model Gateway = Application → Model control

Router: Which capability?
Orchestrator: Which step / transition next?
State store: Where did the task stop?

Debug RAG:
Source → Parse/index → Access/version filters
       → Retrieve/rank → Final context → Supported answer

RAG = retrieval-augmented generation. Recall concerns finding relevant evidence; precision concerns how much retrieved evidence is relevant. Groundedness concerns support, not whether the source is current and correct for the case.

Agent recommends → Policy decides → Executor acts
                          ↓
                 Approval if required

HITL = human-in-the-loop. High model confidence is not permission to act. Approval does not grant unrelated access or reverse completed transactions.

Rate limit: May this tenant send this much work?
Admission:  Can the system accept it now?
Queue:      Where can accepted work wait within limits?

Latency: Trace → Critical path → Optimize → Recheck quality
Cost: Total cost across attempts ÷ successful tasks

Streaming does not necessarily reduce computation. Async does not automatically shorten completion. A queue does not create capacity. TTFT marks only when output starts; useful user output and confirmed action completion are separate measurements.

Rule: timeout ≠ confirmed failure; checkpoint ≠ exactly-once external effect; DLQ ≠ completion. Never generate a fresh logical action key simply because the previous attempt timed out.

Good wording ≠ completed task
Offline evaluation ≠ real production outcome
Shadow ≠ canary
Rollback ≠ undo completed payment

Logs: Specific events
Metrics: Aggregate patterns
Traces: Correlated path
Audit: Business decision/action trail

Observe tool operations, state transitions, decision summaries, and evidence, not private model chain-of-thought, which is often hidden and not always faithful. Shadow candidates must not duplicate real actions. Keep telemetry access and retention appropriate to sensitive data.

Distinguish: authentication = who; authorisation = allowed resource/operation; business policy = whether this action meets the rule; tenant fairness = shared-capacity allocation.

“I’m choosing X because of requirement Y. The trade-off is Z. I’d verify it using W.”

Need AI?       → WHY + TODAY
What must do?  → FRs
How well?     → NFRs
Documents?    → Authorized RAG
Live state?   → API/tool
Dynamic steps?→ Bounded agent; otherwise workflow
Business write?→ Policy + approval if required + executor
More users?   → Workload + bottleneck + capacity/fairness
Slow?         → Critical-path trace
Expensive?    → Cost per successful task
Timeout write?→ Reconcile; preserve action identifier
Ready to ship?→ Evaluate + version + gate + scoped rollout
Why happened? → Observability or audit?
Restricted data?→ Enforce resource/tenant boundaries
Why this choice?→ Requirement + trade-off + verification

Closing check: every important box should have a customer requirement, a failure behaviour, and a way to verify that it worked.

• OWASP GenAI LLM07:2025 System Prompt Leakage — access control before model context; permission-aware retrieval
• OWASP GenAI LLM06:2025 Excessive Agency — access control before model context; untrusted retrieved and tool input
• Claude docs - Rate limits — provider rate limits cap throughput
• Temporal blog - Idempotency and durable execution — checkpoints don't stop duplicate side effects; at-least-once redelivery
• LangGraph docs - Interrupts — checkpoints don't stop duplicate side effects
• Anthropic Engineering, How we built our multi-agent research system — multi-agent overhead
• Cemri et al. 2025, Why Do Multi-Agent LLM Systems Fail? (MAST, arXiv 2503.13657) — multi-agent overhead
• Chen et al. 2025, Reasoning Models Don't Always Say What They Think (arXiv 2505.05410) — chain-of-thought is not always faithful
• Stripe API reference - Idempotent requests — idempotency key reuse and scope
• IETF draft-ietf-httpapi-idempotency-key-header-07 (expired) s2.7 — idempotency key reuse and scope
• OpenAI API docs - Batch API — batch discounts
• Claude docs - Models overview — batch discounts
• Zheng et al. 2023, Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena (arXiv 2306.05685) — LLM-judge bias, calibration
• OWASP GenAI LLM08:2025 Vector and Embedding Weaknesses — permission-aware retrieval
• OWASP GenAI LLM01:2025 Prompt Injection — untrusted retrieved and tool input
• Amazon Bedrock User Guide: Cross-Region inference — cross-region inference
• Amazon Bedrock User Guide: Geographic cross-Region inference — cross-region inference