The LLM Gateway Is a Policy Engine, Not a Proxy

Normalize capability without erasing provider semantics

A gateway should give applications a stable contract while preserving the provider-specific controls required for quality, safety, and performance. The production boundary is larger than the feature that makes the technique attractive: callers arrive with different authorization, freshness, latency, and quality requirements, while the underlying model and data keep changing. A design is therefore incomplete until it states which inputs are part of identity, what may be reused, what must be recomputed, and which failure should be visible to the caller. Treating those choices as an explicit contract prevents an optimization from quietly becoming a correctness or governance bug.

A canonical request captures messages, tools, schema, latency class, data policy, and quality target, while adapters map only supported semantics and expose degradation. This mechanism couples layers that teams often measure separately. A change in routing alters traffic shape; a change in traffic shape alters cache behavior and queue time; a change in retained context alters cost, privacy exposure, and model behavior. Preserve the intermediate decisions needed for replay: normalized inputs, policy version, selected route, relevant hashes, model revision, timing, and outcome. The final text alone cannot explain why the system behaved as it did, and it cannot support a safe rollback when the frontier model changes.

Reject silent coercion, version adapter behavior, and return capability metadata whenever a requested feature is approximated or unavailable. Use rate of semantically degraded requests by provider and feature as the primary operating signal, but review it beside tail latency, unit cost, policy violations, and an outcome metric owned by the product. Roll out behind a versioned decision with a comparison cohort and an explicit reversal trigger. Representative evaluation must include adversarial boundaries, stale state, partial dependency failure, and the busiest traffic window. That evidence converts a plausible architecture into an operable system and gives the team a durable baseline for the next model, workload, or regulation.

Routing is a constrained decision

Choosing the cheapest model that usually works is insufficient when data residency, tool support, context length, latency, or failure consequence are hard requirements. The production boundary is larger than the feature that makes the technique attractive: callers arrive with different authorization, freshness, latency, and quality requirements, while the underlying model and data keep changing. A design is therefore incomplete until it states which inputs are part of identity, what may be reused, what must be recomputed, and which failure should be visible to the caller. Treating those choices as an explicit contract prevents an optimization from quietly becoming a correctness or governance bug.

Policy filters ineligible routes first, then a scorer ranks the remaining models using predicted quality, queue time, unit cost, and recent health. This mechanism couples layers that teams often measure separately. A change in routing alters traffic shape; a change in traffic shape alters cache behavior and queue time; a change in retained context alters cost, privacy exposure, and model behavior. Preserve the intermediate decisions needed for replay: normalized inputs, policy version, selected route, relevant hashes, model revision, timing, and outcome. The final text alone cannot explain why the system behaved as it did, and it cannot support a safe rollback when the frontier model changes.

Log eligible routes and scoring evidence, run challengers in shadow, and prevent product teams from bypassing non-negotiable policy through raw provider credentials. Use quality-adjusted cost within each latency and policy class as the primary operating signal, but review it beside tail latency, unit cost, policy violations, and an outcome metric owned by the product. Roll out behind a versioned decision with a comparison cohort and an explicit reversal trigger. Representative evaluation must include adversarial boundaries, stale state, partial dependency failure, and the busiest traffic window. That evidence converts a plausible architecture into an operable system and gives the team a durable baseline for the next model, workload, or regulation.

Retries must understand side effects

A generic network retry can duplicate tool calls, charges, messages, or external mutations even when the model request itself appears idempotent. The production boundary is larger than the feature that makes the technique attractive: callers arrive with different authorization, freshness, latency, and quality requirements, while the underlying model and data keep changing. A design is therefore incomplete until it states which inputs are part of identity, what may be reused, what must be recomputed, and which failure should be visible to the caller. Treating those choices as an explicit contract prevents an optimization from quietly becoming a correctness or governance bug.

The gateway classifies stages as replayable or consequential and coordinates idempotency keys, stream state, tool receipts, and model sampling state. This mechanism couples layers that teams often measure separately. A change in routing alters traffic shape; a change in traffic shape alters cache behavior and queue time; a change in retained context alters cost, privacy exposure, and model behavior. Preserve the intermediate decisions needed for replay: normalized inputs, policy version, selected route, relevant hashes, model revision, timing, and outcome. The final text alone cannot explain why the system behaved as it did, and it cannot support a safe rollback when the frontier model changes.

Retry only before irreversible output or tool execution, expose partial-stream failure, and require the application to resume consequential workflows deliberately. Use duplicate-side-effect incidents and safe retry success rate as the primary operating signal, but review it beside tail latency, unit cost, policy violations, and an outcome metric owned by the product. Roll out behind a versioned decision with a comparison cohort and an explicit reversal trigger. Representative evaluation must include adversarial boundaries, stale state, partial dependency failure, and the busiest traffic window. That evidence converts a plausible architecture into an operable system and gives the team a durable baseline for the next model, workload, or regulation.

Evidence is a first-class gateway product

Centralization is valuable only if it produces decision evidence without becoming an uncontrolled warehouse of sensitive prompts. The production boundary is larger than the feature that makes the technique attractive: callers arrive with different authorization, freshness, latency, and quality requirements, while the underlying model and data keep changing. A design is therefore incomplete until it states which inputs are part of identity, what may be reused, what must be recomputed, and which failure should be visible to the caller. Treating those choices as an explicit contract prevents an optimization from quietly becoming a correctness or governance bug.

Structured traces retain route, policy, model, token, latency, error, and outcome identifiers while payload capture follows tenant-specific redaction and retention. This mechanism couples layers that teams often measure separately. A change in routing alters traffic shape; a change in traffic shape alters cache behavior and queue time; a change in retained context alters cost, privacy exposure, and model behavior. Preserve the intermediate decisions needed for replay: normalized inputs, policy version, selected route, relevant hashes, model revision, timing, and outcome. The final text alone cannot explain why the system behaved as it did, and it cannot support a safe rollback when the frontier model changes.

Separate operational metadata from content, make raw capture opt-in and encrypted, and connect gateway traces to product outcomes through stable run identifiers. Use percentage of production failures diagnosable without raw sensitive payloads as the primary operating signal, but review it beside tail latency, unit cost, policy violations, and an outcome metric owned by the product. Roll out behind a versioned decision with a comparison cohort and an explicit reversal trigger. Representative evaluation must include adversarial boundaries, stale state, partial dependency failure, and the busiest traffic window. That evidence converts a plausible architecture into an operable system and gives the team a durable baseline for the next model, workload, or regulation.

Research foundation and scope

OpenTelemetry's generative-AI conventions support consistent model-operation telemetry, while NIST and OWASP define broader risk and application-security responsibilities [OTEL-GENAI] [NIST-GAI] [OWASP-LLM]. A gateway is valuable when it enforces these shared boundaries. These sources establish the mechanism, but they do not remove the need to measure the local implementation. A paper's result belongs to its checkpoint, dataset, hardware, traffic, and experimental protocol; a standard describes a contract or risk practice, not a guarantee that a particular product follows it. The useful engineering move is to convert each cited idea into a hypothesis that can be falsified with representative inputs and observable intermediate state.

A common request format must not silently erase provider-specific semantics. Model capability, data policy, region, tools, schemas, streaming, retries, and side effects determine whether adaptation is faithful. Write these conditions into the design review before selecting a library or provider. Specify the authoritative source of truth, the authenticated principal, the versions that influence behavior, the maximum consequence of an error, and what the system will do when required evidence is missing. This boundary statement makes later optimization honest: teams can compare speed, cost, or convenience only after candidates satisfy the same correctness, authorization, privacy, and recovery contract.

Read [OTEL-GENAI], [NIST-GAI], [OWASP-LLM] together rather than treating one source as a recipe. OpenTelemetry Generative AI semantic conventions supplies one part of the foundation; Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile supplies a second perspective; OWASP Top 10 for LLM Applications 2025 helps connect the mechanism to evaluation, governance, or operations. The citations are deliberately primary papers, standards, or official documentation. Product-specific claims still require local evidence, and any dated behavior must be rechecked against the cited version before an article or architecture decision is refreshed.

A production implementation blueprint

Define a versioned canonical contract, expose semantic degradation, authenticate applications and users, enforce budgets and data policy, filter eligible routes, coordinate idempotency, and emit structured decision evidence without collecting unnecessary content. Begin with one thin vertical path that preserves all decision evidence. Give each run a stable identifier; capture normalized input, authenticated scope, relevant dependency versions, selected policy, timing, result status, and any external receipt. Keep large or sensitive payloads in a separately governed store and pass bounded references through the orchestration layer. This produces a debuggable system without making raw customer content the default telemetry substrate.

Make state transitions explicit and make every external boundary return a typed result. Distinguish validation failure, permission denial, missing evidence, rate limit, timeout, dependency outage, user cancellation, and internal defect. The distinction matters because each permits a different response: clarify, retry, fall back, defer, compensate, or stop. A generic exception handler that asks the model to try again turns uncertainty into repeated cost and can duplicate irreversible effects. Retries require an idempotency strategy and a stored receipt, not optimism.

Roll out the implementation as a versioned policy. Record which cohort receives it, what baseline it replaces, which metrics can stop the rollout, and how to reverse stateful artifacts such as indexes, caches, adapters, or workflow histories. Review privacy and retention alongside performance. If raw content is needed temporarily for diagnosis, state the purpose, access, sampling, encryption, and deletion window. Production readiness means the team can explain, contain, and reverse behavior under load—not merely demonstrate the happy path once.

Failure lab and evaluation plan

Run adapter conformance across providers, replay routing cohorts, test policy denial and audit, compare provider semantics for tools and schemas, and chaos-test failures before and after stream or side-effect boundaries. Keep the arrival pattern, data distribution, authorization, and dependency versions close to production. Use a paired baseline where possible and retain per-case results so averages cannot hide a severe slice. Offline evaluation should localize the failing stage; shadow evaluation should expose integration and traffic effects; a canary should confirm user-facing outcomes with a small blast radius. Each stage needs an exit criterion rather than a vague request to 'look good.'

Include token leakage, bypass through direct credentials, invalid model capability claim, partial stream retry, provider outage, duplicate tool execution, policy-version skew, redaction miss, and gateway overload becoming a fleet-wide dependency. Add adversarial boundary cases and partial failures, not only malformed inputs. Terminate a worker after an external system may have committed, revoke access after a cache was populated, change one dependency version, and delay one branch until its data is stale. Confirm that cancellation releases resources and that a fallback preserves semantics. The test passes only when the external outcome, audit evidence, and user-visible state agree; a fluent final message is not proof of completion.

Evaluation data needs lineage equal to production data. Record why each case exists, who reviewed the expected behavior, which source or policy supports it, and when it must be refreshed. Separate generated cases from production incidents and prevent test cases from leaking into training. When a metric changes, inspect changed examples before accepting an aggregate. A statistically small regression can still block release if it crosses an authorization, privacy, safety, or irreversible-action boundary.

Operating metrics and decision record

Use semantically degraded requests, policy denials and escapes, route quality-adjusted cost, adapter failures, p95 latency, safe retry, duplicate effects, trace completeness, sensitive-content capture, and gateway availability. Define every metric with a numerator, denominator, unit, time window, exclusions, and responsible owner. Segment by workload and consequence so a dominant easy class cannot hide a rare costly failure. Pair service signals with product outcomes: latency without completion encourages fast useless answers, while answer quality without queue and cost data can produce an uneconomic service. Review percentiles and distributions rather than relying on a single average.

Centralize only contracts and controls the organization can operate consistently, while preserving provider differences and application-owned authorization rather than masking them. Put that choice in a short architecture decision record containing context, alternatives, evidence, assumptions, selected policy, rollout, reversal trigger, and unresolved risk. Link the evaluation snapshot and source versions. This document prevents a benchmark from becoming folklore and tells the next engineer which evidence must be repeated when the model, provider, traffic, data, regulation, or product promise changes.

Monitor leading and lagging signals after release. Leading signals reveal pressure—queue age, cache allocation, disagreement, validation failure, policy denial, or retry growth—before users report harm. Lagging signals show whether the job was actually completed and trusted. Alert on a service objective tied to user consequence, and make the run identifier available to support without exposing private content. Every serious miss should become a reproducible regression case and, when relevant, a new threat or data-quality control.

Primary sources and further reading

1. OpenTelemetry Generative AI semantic conventionsOpenTelemetry · standard
2. Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence ProfileNational Institute of Standards and Technology · standard
3. OWASP Top 10 for LLM Applications 2025OWASP Foundation · standard