InferCrane feature qualification matrix

This is the authoritative inventory for whole-product qualification. AUTOMATED_QUALIFIED means the strongest safe local boundary is green in repository CI. AWS_REAL_PASSED records the narrower real L40S boundary in AWS real-infrastructure evidence; SCALEWAY_BREZEL_REAL_PASSED records the bounded private-computer lifecycle in Scaleway Brezel evidence. Neither qualifies another provider, topology, model, or failure mode. REAL_INFRA_REQUIRED and MANUAL_REQUIRED remain explicit and are never satisfied by fixtures. Fixture evidence never qualifies real provider or GPU behavior.

Feature	User-visible / operational behavior	Implementation	Existing evidence/tests	Qualification level and automated strategy	Manual requirement	Current status
CLI bootstrap/config contexts	`init`, context selection and missing-config remediation	`cmd/infercrane`, `internal/config`	command/config tests	Unit + local API CLI E2E	None expected	AUTOMATED_QUALIFIED
Doctor	Dependency, auth and configuration diagnostics	`cmd/infercrane`, `internal/doctor`, control API	doctor/API/CLI tests	Deterministic fixture + malformed config	Provider credential validity is external	AUTOMATED_QUALIFIED
Fleet inbox	Rank persisted endpoint/deployment attention without mutation	CLI over tenant-scoped fleet APIs	command/model tests + live local API	Severity/order/limits/unknown state/partial-read fail-closed/content exclusion	Fresh diagnosis and provider truth remain explicit follow-up	AUTOMATED_QUALIFIED
Semantic plan	Side-effect-free deployment diff	`internal/planning`, CLI/API	planning and command tests	Unit + stack assertion of zero mutation	Provider price display needs sourced data	AUTOMATED_QUALIFIED
Intent plan	Resolve bounded user text to one exact reviewed model and an editable architecture/configuration, or return explicit missing choices	`intentplan`, control API	intent planner and API tests	Strict 16 KiB JSON, deterministic recipe matching, no silent model substitution, exact deployment-comparable provider price ranking separated from connection readiness, sourced per-replica price disclosure, and explicit unmeasured performance/unknown capacity	Natural-language breadth is intentionally limited; arbitrary Hugging Face discovery, real capacity, and performance qualification remain separate	AUTOMATED_QUALIFIED (DETERMINISTIC CONTRACT)
Hermetic product demo	One command proves adoption through rejected-candidate cleanup	`scripts/demo-product.sh`, Docker stack, public CLI	acceptance safety + black-box execution	Isolated Compose project, fixture label, exact active-revision invariant, zero provisioning commands	Cannot qualify GPU, provider, model quality, or performance	AUTOMATED_QUALIFIED
Deployment apply/deploy	Durable provider-neutral lifecycle	`internal/workflows`, `store`, control API	workflow/store/stack/fault suites	PostgreSQL + fixture provider E2E/restart	Real provider lifecycle	AUTOMATED_QUALIFIED
Durable operations	Watch/resume/cancel after disconnect with a persisted current phase	`internal/operations`, `store`, CLI	operations, failure recovery, acceptance	State-machine + process restart + CLI disconnect, semantic checkpoint deduplication, exact current-step API/CLI rendering	Real in-flight provider call crossed create/restart/cancel/cleanup on the bounded RunPod GLM transfer; successful runtime completion remains tuple-specific	AUTOMATED_QUALIFIED / REAL_INFRA_REQUIRED
Provider ensure adoption	Lost-create-response and exactly-one intent	`internal/provision`, `workflows`, conformance	provider contracts/fault fixtures	Fault injection + repeated idempotency	Real provider adoption semantics	AUTOMATED_QUALIFIED
Readiness/model identity	Health plus intended served model	`runtime`, `reconcile`, `workflows`	runtime/reconcile/stack tests	Fake vLLM malformed/delayed/wrong identity	AWS vLLM/SGLang/custom OCI passed; other providers remain separate	AUTOMATED_QUALIFIED / AWS_REAL_PASSED
Safe delete/orphans	Route withdrawal, drain, idempotent cleanup, inventory	workflows/store/routes/acceptance	fault and acceptance fixtures	Restart/delete boundary + inventory equality	AWS worker and root-volume cleanup passed; other providers remain separate	AUTOMATED_QUALIFIED / AWS_REAL_PASSED
Immutable DeploymentSpec/revisions	Versioned strict spec, immutable candidate history, and exact accelerator type/count	`spec`, `store`, `domain`	spec/migration/store tests	Schema/property + topology persistence tests	None expected	AUTOMATED_QUALIFIED
Rollout/provision/promote/reject/rollback	Safe revision lifecycle	workflows/releaseguard/routes	rollout/guard/fault tests	E2E active/candidate and stale action rejection	Real generation-safe GPU rollout	AUTOMATED_QUALIFIED
Release Guard	Deterministic persisted accept/reject/monitor/rollback, with threshold or seeded paired-bootstrap quality comparison	`releaseguard`, store, API/CLI	unit/store/API/acceptance	Strong local candidate E2E; deterministic 10,000-resample bootstrap records seed/sample counts/interval; insufficient or overlapping evidence waits	Real performance/cost/quality evidence; InferCrane enforces signed external semantic evidence but does not itself judge answer quality	AUTOMATED_QUALIFIED
External task-quality evidence	Ingest aggregate evaluator results or bounded content-free paired score distributions, bind/sign/attach exact revision evidence	`qualityevidence`, evaluation CLI, control API	strict schema/content/tamper/key/API tests	Versioned distribution schema, sample-count equality, pairing digest, finite bounds, unknown fields, oversize/trailing JSON and attachment failures fail closed	Evaluator calibration, pairing validity, representative private datasets, key custody and real task quality remain customer evidence	AUTOMATED_QUALIFIED
Router supervision/generations	Instance-owned router lifecycle and safe generations	`router`, `routes`, workflows	router/routes/HA tests	Process crash/restart and active stream route swap	Real long-running vLLM stream	AUTOMATED_QUALIFIED
OpenAI Chat Completions	Buffered and streaming proxy	`gateway`, `openaicompat`	gateway/stack/SDK tests	Local fake runtime E2E success/failure/cancel	Real vLLM compatibility	AUTOMATED_QUALIFIED
Responses API	Faithful capability-gated proxy	gateway/runtime contract	gateway/protocol tests	Qualified fixture, malformed/upstream errors	Qualified real runtime	AUTOMATED_QUALIFIED
Embeddings	Faithful capability-gated proxy	gateway/runtime contract	gateway/protocol tests	Qualified fixture and dimensional payload preservation	Qualified real model/runtime	AUTOMATED_QUALIFIED
Completions	Legacy compatible proxy	gateway/runtime contract	gateway/protocol tests	Fixture contract + streaming	Qualified real runtime	AUTOMATED_QUALIFIED
Chat batch	Online batch proxy/capability	gateway/runtime contract	gateway/protocol tests	Fixture body/response/error preservation	Qualified real runtime	AUTOMATED_QUALIFIED
Tool calling/structured output	Protocol pass-through and capability claim	gateway/integration profiles	acceptance/contract tests	Local schema/tool payload preservation	AWS vLLM/Qwen path passed; every additional runtime/model pair remains separate	AWS_REAL_PASSED / REAL_INFRA_REQUIRED
Request IDs/tracing/retries	Stable request/trace identity and bounded retries	gateway/telemetry/admission	gateway tests	Failure sequence, cancel and no streaming retry	None expected	AUTOMATED_QUALIFIED
Telemetry/metrics	OTel GenAI identity, latency/token/error/cold metrics plus fresh revision-bound hardware evidence	gateway/metrics/store/control API/CLI	metrics/gateway/store/API/CLI tests	Local request E2E, bounded DCGM parsing, tenant/revision/idempotency/freshness, and content exclusion	Real DCGM/GPU field availability and complete fleet labeling	AUTOMATED_QUALIFIED
Stable logical models/environments/endpoints	Stable app identity independent of backend	store/routes/API/CLI	endpoint/store/API tests	Tenant/reference/digest/route E2E	None expected	AUTOMATED_QUALIFIED
Backend bindings/serving plans	Active/candidate/fallback plus exactly-two-provider active-active weighted plan compilation	store/routes/control API	endpoint/routes/PostgreSQL tests	Same explicit upstream model, distinct provider identities, deterministic weights, health decay/recovery hysteresis and generation pinning	Real cross-provider network, health propagation, streaming continuity and failover timing are unqualified	AUTOMATED_QUALIFIED (LOCAL CONTRACT) / REAL_INFRA_REQUIRED
Managed external endpoint bindings	Stage authenticated OpenRouter or generic OpenAI-compatible capacity behind a stable endpoint without browser-visible credentials	external/store/reconcile/control API/CLI	strict policy, tenant, budget, credential-reference, authenticated model-probe and route-compilation tests	PostgreSQL-backed hard reservation, fail-closed consent, candidate-only console flow, successful-probe caching	Real provider protocol, streaming and billing semantics remain target-specific	AUTOMATED_QUALIFIED
Managed Model API prepaid billing	Reserve customer wallet funds before supplier transmission, settle from returned token usage, and preserve unknown usage for operator reconciliation	`managedbilling`, gateway, store, control API, console	deterministic calculator, insufficient-funds/no-supplier-call, buffered/SSE settlement, PostgreSQL idempotency, append-only ledger, pending-usage listing, explicit settlement, verified-no-charge release, fixed-price Stripe Checkout, signed-webhook mode/amount/tenant verification, duplicate-event/session tests, immutable cost-basis/margin validation, and expired-rate no-transmission	Tenant routes cannot set retail prices; only signed paid events grant balance; browser redirects never grant credit; request content excluded; missing usage never becomes an invented charge or an automatic release; supplier cost basis stays private	Stripe sandbox journey, live payment collection, tax, refunds, automated supplier invoice reconciliation, overhead allocation, and real supplier usage semantics remain unqualified	AUTOMATED_QUALIFIED (LOCAL BILLING CONTRACT) / REAL_INFRA_REQUIRED
Managed Model API discovery catalog	Present a curated model shelf and one InferCrane service contract without exposing procurement routing	`modelapicatalog`, control API, console	strict versioned catalog validation, bounded filters/pagination, curated default shelf, customer-contract redaction tests, expired-rate downgrade, responsive browser journey	No supplier identity, internal offer ID, region, adapter, URL, credential, or upstream rate provenance crosses the customer API; missing or expired rate/context remains unknown	Managed availability, real pricing, performance, and production qualification require a separately operated supply catalog and real evidence	AUTOMATED_QUALIFIED (LOCAL CATALOG CONTRACT) / REAL_INFRA_REQUIRED
Managed Model API supply planning	Compile private offers into an immutable primary/fallback plan for one exact logical model	`modelapisupply`, `modelapicatalog`, control API	deterministic order-independent plans, content digest, exact-model/protocol/capability/region gates, freshness/health/rate/margin rejection, exact-tuple minimum-sample SLO evidence, supplier-diverse fallback, no-workload no-cost-ranking tests	Missing or stale evidence fails closed; no silent model substitution; a price rank is emitted only for a declared token shape; suppliers remain private	Real supplier availability, billing semantics, correlated failure domains, streaming failover, reconciliation and sustained gross margin remain unqualified	AUTOMATED_QUALIFIED (LOCAL PLANNING CONTRACT) / REAL_INFRA_REQUIRED
Managed Model API RunPod routing and COGS	Move one stable logical model gradually onto a private RunPod Serverless vLLM target while retaining upstream fallback and exact economics	`modelapitarget`, `modelapirouting`, `modelapiqualification`, `modelapireconciliation`, `supplieradapter`, store	immutable offer-bound target bindings, endpoint-config digest checks, strict buffered/SSE usage contracts, deterministic basis-point selection, future-request circuit fallback, exact-tuple measurement digest, integer micro-USD COGS/margin reconciliation, PostgreSQL migration and store tests	RunPod candidates cannot publish without one current exact binding; no retry occurs after supplier transmission; supplier rate identity is pinned before execution; ambiguous usage stays unsettled	A real RunPod endpoint, supplier credential, representative load, cold-start behavior, invoice reconciliation, and progressive production canary remain required	AUTOMATED_QUALIFIED (LOCAL CONTRACT) / REAL_INFRA_REQUIRED
External agent sandbox composition	Reference an externally owned sandbox and issue one expiring endpoint-scoped inference credential	integration catalog, store, control API, gateway	PostgreSQL identity/TTL/rotation/revocation, endpoint-scope enforcement, metadata content exclusion and no-external-mutation tests	InferCrane does not run commands, store files/content, claim isolation, or treat plain containers as a hostile-code boundary	Managed sandbox, gVisor, Kata and Kubernetes Agent Sandbox adapters each require real isolation, egress, cleanup, snapshot and credential-injection qualification	AUTOMATED_QUALIFIED (COMPOSITION CONTRACT) / REAL_INFRA_REQUIRED
InferCrane private computers on Brezel	Give one InferCrane tenant persistent isolated CPU workspaces through approved immutable environments	`brezelsandbox`, `sandboxprovider`, control API, metadata store, usage ledger, console proxy	Customer-ID ownership checks, cross-tenant rejection, HTTPS and owner-only token enforcement, durable workspace compensation, incremental command NDJSON, bounded file streaming, re-brokered preview leases, content-free receipts/events/usage, lifecycle idempotency, provider-identity redaction, no-internet default, plus bounded production lifecycle evidence	Private-tenant preview only; no shared-hosted, GPU, PTY/SSH, arbitrary OCI, controlled arbitrary egress, host-loss durability, or hardware-attestation claim; usage is not an invoice	Production creation, commands, files, preview, receipts, pause/resume persistence, model connector, usage and cleanup passed; host-loss recovery, capacity exhaustion, host restart, preview-expiry timing and adversarial isolation remain separate drills	SCALEWAY_BREZEL_REAL_PASSED (BOUNDED CPU LIFECYCLE) / REAL_INFRA_REQUIRED
Hosted control-plane deployment profile	Run the API, gateway, reconciler, and durable-operation worker on one always-on Fly.io CPU machine with external PostgreSQL behind a provider-neutral host contract	Docker image, `deploy/fly/control-plane.toml`, production config, portable hosting contract	TOML and portable-host invariant tests, production Compose/entrypoint checks, readiness and shutdown contract tests	No embedded secret, app name, shared replica identity, host volume, or scale-to-zero; Fly is not GPU supply; PostgreSQL remains the durable boundary; Render/ECS/Kubernetes are portability targets, not qualified deployments	Real Fly deployment, restart, region loss, database interruption, cost, external-client traffic shift, and two-machine rolling upgrade	CONFIGURATION_QUALIFIED / REAL_INFRA_REQUIRED
Reusable provider connections	Configure Baseten, Fireworks, OpenRouter, or an operator-supplied OpenAI-compatible Modal, RunPod Serverless API, Fly.io service, or generic endpoint once and bind it without exposing credentials or silently authorizing spend	store/control API/CLI	tenant/idempotency/adapter/credential-reference/binding compilation and authenticated discovery tests	Reference-only persistence, bounded bearer-authenticated `/v1/models` discovery, explicit per-binding consent and hard budgets, non-destructive delete; provider labels do not imply provisioning	Real provider credential, generation/streaming, availability, performance, and billing semantics	AUTOMATED_QUALIFIED (PROTOCOL CONTRACT) / REAL_INFRA_REQUIRED
Endpoint Release Guard	Compare and govern serving plans	store/releaseguard/API	endpoint guard tests	Deterministic persisted candidate comparison	Real cross-backend evidence	AUTOMATED_QUALIFIED
Import/adopt ownership	Observe-only then traffic-managed adoption	store/control API/CLI	adoption/API/route tests	Ownership enforcement and idempotent atomicity	Real existing endpoint	AUTOMATED_QUALIFIED
Bare-metal GPU discovery	Read-only concrete local NVIDIA inventory before endpoint adoption	`nodediscovery`, CLI	parser/command-boundary/CLI tests	No shell, five-second deadline, bounded output/count/fields, explicit unavailable state, no lifecycle transfer or compatibility claim	Real NVIDIA driver matrix; AMD/ROCm and remote hosts unavailable	AUTOMATED_QUALIFIED (LOCAL CONTRACT) / REAL_INFRA_REQUIRED
Request Inspector	Reconstruct request metadata without content	store/API/CLI	diagnostic/store/API tests	Tenant/content exclusion E2E	Provider-hidden timing	AUTOMATED_QUALIFIED
Deterministic Doctor/explain	Evidence-backed operational findings	`doctor`, explain CLI/store	doctor/CLI tests	Rule coverage + stale/missing evidence	None expected	AUTOMATED_QUALIFIED
Signed alerts	Policy evaluation and signed webhook retry	`alert`, store/API/CLI	alert/API tests	Local TLS webhook failure/retry/signature	Slack/PagerDuty adapters not implemented	AUTOMATED_QUALIFIED
Admission control and request deadlines	Concurrency/queue/size/token/priority/load shedding plus one endpoint-wide latency budget	`admission`, gateway, store, CLI/API/console	admission/gateway/store/API/CLI/browser tests	Concurrent local E2E, cancellation, queue expiry, bounded buffered retries, streaming expiry, PostgreSQL persistence, and cancelled-context evidence recording	Real-runtime cancellation and overload calibration remain exact-tuple evidence	AUTOMATED_QUALIFIED / REAL_INFRA_REQUIRED
Tenant request quota	Distributed bounded lease without DB data path	`requestquota`, store/gateway	quota/store tests	Concurrency/property/failure tests	Multi-host timing at scale	AUTOMATED_QUALIFIED
Async inference	Idempotent encrypted queued jobs, cancel, signed webhook	`asyncinference`, API/CLI	async/API tests	Worker/restart/deadline/retry/cancel E2E	Real long-running inference	AUTOMATED_QUALIFIED
Autoscaling	Min/max, SLO-aware replica target from persisted request attainment and exact-tuple one-replica benchmark goodput, safe scale down	`autoscale`, workflows/store	autoscale/workflow/store/acceptance	Minimum 20 samples, exact immutable model/runtime/version/args/provider/region/GPU tuple, queue fallback with `capacity evidence absent`, hysteresis/cooldown/router fence and persisted snapshots	Real GPU provider scaling and benchmark calibration	AUTOMATED_QUALIFIED (DECISION CONTRACT) / REAL_INFRA_REQUIRED
External governed fallback	Explicit privacy plus hard request/cost budget	`external`, store/gateway/CLI	external/store/gateway tests	Concurrent budget reservation and fail closed	Real external provider	AUTOMATED_QUALIFIED
Burst Guard	Fresh sustained health/queue/cost bounded overflow	`burstguard`, store/API/CLI	burst unit/API contracts	State transitions, stale evidence and budget boundary	Real overflow/provider cost	AUTOMATED_QUALIFIED
Provider-native serverless	Durable endpoint, zero/warm/cold/cancel/delete	provision/workflows/gateway	serverless fixture/fault acceptance	Lost response, cancellation, orphan fixture	RunPod Serverless real cycle	REAL_INFRA_REQUIRED
Cold-start intelligence	Grounded timing boundaries, exact-tuple per-stage history, and explanations	gateway/store/metrics/planning	cold-start/store/planning/explain tests	Provider markers only; per-stage p50 after 3 successes and p95 after 20 matching immutable model/runtime/version/args/provider/region/GPU observations; absent stages remain unknown	Real provider marker completeness and hidden stages	AUTOMATED_QUALIFIED (EVIDENCE CONTRACT) / REAL_INFRA_REQUIRED
ModelArtifact/HF identity	Resolve public, private, or gated mutable HF ref to an immutable artifact	`artifact`, store	artifact unit/HTTP fixture/store/workflow	Tenant-scoped secret reference, token passed only through child environment, immutable revision, cache metadata validation	Real gated HF/Xet transfer and provider-native runtime secret delivery	AUTOMATED_QUALIFIED
Artifact cache observations/prefetch	Expiring evidence, delegated idempotent intent, and explicit bounded provider preparation workflows	`artifactcache`, store/API, guarded scripts	artifactcache/provider/API and hermetic script safety tests	PostgreSQL tenant/idempotency/expiry, stable lost-response adoption, exact-model AWS snapshot, GCP zonal disk, Kubernetes PVC, RunPod identity-named network-volume attachment, resumable RunPod population plan with hard cost/watchdog/cleanup gates	Real provider attachment, byte integrity, filesystem behavior, cache-hit timing, and node-local/provider-native caches remain external	AUTOMATED_QUALIFIED (CONTRACT) / REAL_INFRA_REQUIRED
AIPerf benchmark/history	Explicit reproducible measurement	`benchmark`, store/API/CLI	benchmark/API/store tests	Fake AIPerf process + exact args/results/failure	AWS L40S qualification samples passed; representative performance campaigns remain separate	AWS_REAL_PASSED / REAL_INFRA_REQUIRED
Recipes	Immutable content-addressed verified serving recipe	`recipe`, store/API/CLI	recipe/store/API tests	Evidence matching/digest/idempotency/tenant	Community/public registry not implemented	AUTOMATED_QUALIFIED
Inference Lab	Compare measured candidate evidence	`lab`, store/API/CLI	lab/store/API tests	Measured-only/no hidden provisioning/no invented cost	Real candidate catalog evidence	AUTOMATED_QUALIFIED
Optimization proposals	Generate bounded immutable catalog or modeled candidates without mutation	`optimizer`, pinned AIConfigurator adapter, integration/recipe catalogs, CLI	optimizer/CLI tests plus real upstream package contract	Multi-model catalog, deterministic inputs, typed evidence state, immutable custom-runtime profiles, exact GPU type/count, credential scrubbing, bounded process output, explicit fallback, topology/mutation-owner safety, strict loadable specs, no modeled-as-measured claim	Estimator calibration and real-hardware candidate qualification	AUTOMATED_QUALIFIED / REAL_INFRA_REQUIRED
Optimization campaign coordination	Persist, inspect, approve with an expiring hard cost cap, advance proof boundaries durably, rank measured exact-workload candidates, cancel and clean candidates, and require explicit durable activation or guarded promotion	`optimizationcampaign`, Lab, store, control API, CLI	state-machine/coordinator/composite-driver/ranker/operation/store/API/CLI tests plus black-box modules journey	Exact sourced price authority, PostgreSQL durability, explicit new-endpoint versus evolve-endpoint intent, no fabricated guard baseline, mandatory guard for evolution, direct-revision AIPerf, idempotency conflict, stale transition fencing, all-candidate measurement barrier, proposal-order independence, workload/cost/SLO fail-closed ranking, distinct benchmark/quality/Lab/guard evidence, lost-response adoption, approval expiry-to-cleanup, cancellation fencing, tenant isolation, transactional stable-alias publication, route-generation wait, explicit activation, and zero implicit promotion	Real provider/GPU execution, runtime performance, and cleanup remain separate qualification	AUTOMATED_QUALIFIED (LOCAL CONTROL PLANE) / REAL_INFRA_REQUIRED
Optimization capability registry	Compile only exact qualified model/runtime/version/precision/GPU mechanisms and fail closed on unknown or conflicting combinations	`optimizationcapability`, `servingcontract`, runtime profiles, optimizer	registry/compiler/deferred-translation and multi-model tests	Exact tuple matching and conflict rejection; LMCache and Dynamo/NIXL prefill-decode are registered as Deferred for argument translation but compilation always fails closed	LMCache and NIXL execution, transport, routing, failure and GPU performance require mechanism-specific qualification; no support claim	AUTOMATED_QUALIFIED (CURRENT TUPLES; DEFERRED REGISTRATION) / REAL_INFRA_REQUIRED
Accelerator Lab	Profile exact target, plan bounded kernel opportunities, optionally generate profile-bound source, build in Brezel, qualify on target hardware, and stop before promotion	`acceleratorlab`, `kernelplanner`, `brezelexecutor`, operations/control API	engine/client/catalog/workflow/Brezel/config/API tests	Worker-declared capability catalog; immutable model/runtime/workload/hardware/topology; 24-hour authority and spend cap; same-origin content-addressed artifacts; existing-first source search; generated-source profile binding; hidden/adversarial shapes; sanitizer, quality, serving replay, end-to-end, memory and cost gates; no implicit promotion	Deploy the baked optimization runner and exact NVIDIA/AMD/TPU/Neuron worker pools; collect per-tuple profiler, multimodal quality, topology and serving evidence	AUTOMATED_QUALIFIED (ORCHESTRATION AND FAIL-CLOSED GATES) / REAL_INFRA_REQUIRED
Optimized artifact provenance	Plan and attest immutable quantized checkpoints, speculators, and TensorRT engines through external digest-pinned builders	`optimizedartifact`, store, control API, CLI	artifact state/store/API/CLI tests	Immutable base/tool/config/calibration/hardware/license identity, duplicate/tamper/failure/cross-tenant checks, exact deployment/revision/base-model binding, mismatched-model rejection, and mandatory signed quality evidence	Real LLM Compressor, ModelOpt, Speculators, and TensorRT builds plus GPU qualification	AUTOMATED_QUALIFIED (PROVENANCE) / REAL_INFRA_REQUIRED
Production telemetry import and Inference Replay	Import bounded content-free OTel/Datadog/Baseten/Fireworks request shapes, then capture immutable replay history	`replay`, gateway/store/API/CLI	import/replay/gateway/store/API tests	Strict unknown-field rejection, deterministic retry identity, content exclusion, digest/window, tenant-scoped list + opt-in CLI/console	Source-specific managed pull adapters, real workload/GPU approximation, and content-bearing shadow traffic	AUTOMATED_QUALIFIED (CONTENT-FREE CONTRACT) / REAL_INFRA_REQUIRED
Capacity Intelligence	Tenant-scoped observed success/failure/latency plus deterministic cache-aware placement	store/workflows/API/CLI, `capacity`	workflow/store/capacity contracts	PostgreSQL aggregation partitioned by exact GPU type/count, sparse samples, percentiles, expired-cache rejection, readiness/reliability/cost objectives, readiness SLO, and inventory-order independence	Real provider distributions and automatic multi-provider placement remain external	AUTOMATED_QUALIFIED (DECISION CONTRACT) / REAL_INFRA_REQUIRED
SLO/recommendations	Qualified deterministic serving recommendation	`decision`, store/API/CLI	decision/store/API tests	Missing/stale/cost/qualification fail closed	Real evidence quality	AUTOMATED_QUALIFIED
FinOps	Sourced dated hourly cost, no invented savings	`finops`, OpenCost parser, store/API/CLI	finops/store/API/CLI tests	Exact allocation selection, currency/source/window/expiry, active-revision binding, retry conflicts, persistence	Real OpenCost deployment and invoice reconciliation	AUTOMATED_QUALIFIED
Advisory Autopilot	Immutable recommendation plan and human approval, no mutation	store/API/CLI	API/contract tests	Idempotency/tenant/state/audit/no side effects	Execute candidate manually	AUTOMATED_QUALIFIED
Inference Passport	Signed canonical offline release evidence	`passport`, store/API/CLI	passport/store/API tests	Sign/verify/tamper/byte preservation/tenant	Organizational key custody	AUTOMATED_QUALIFIED
Context Passport/session affinity	Durable logical identity, reliability-first hints, no KV promise	`contextpassport`, routes/store/API/CLI	directory/routes tests	Persistence/expiry/tenant/route fallback E2E	Multi-instance snapshot propagation and real worker loss	AUTOMATED_QUALIFIED
Request survival capability	Delegated qualified backend claim	`integration` contract	integration test	Invalid/unknown/unqualified matrix	Real Dynamo/qualified backend migration	REAL_INFRA_REQUIRED
Existing-target provider	Register and route customer-managed endpoint	store/routes/API/CLI	target/store/stack tests	Local fake OpenAI endpoint E2E	Customer network endpoint	AUTOMATED_QUALIFIED
RunPod elastic adapter	Elastic GPU lifecycle/inventory	provision/workflows profiles	fixtures/contracts	Deterministic provider HTTP fixture	Real RunPod GPU	REAL_INFRA_REQUIRED
RunPod native OCI Pods	Immutable image/argv/GPU lifecycle without SSH for built-in vLLM/SGLang and custom OCI workloads	provision/workflows profiles	native REST fixture/contracts + bounded GLM evidence + Qwen evidence	Lost-response adoption, conflict, redaction, exact-region and identity-bound network-volume attachment, persistent Hub/materialized-model paths, delete fixture; exact Qwen3-8B/vLLM 0.22.1/L40S durable deploy, buffered/streaming requests, c1/c8/c32 benchmark, guard rejection, restart-delete, and zero inventory passed	Every other image-model-runtime-GPU-region tuple; Qwen3.8 TP2 B200 remained nondeterministic and is unqualified	RUNPOD_REAL_PASSED (EXACT BASELINE TUPLE) / REAL_INFRA_REQUIRED
RunPod serverless adapter	Provider endpoint lifecycle	provision/workflows profiles	fixtures/fault proxy	Deterministic API fault fixture	Real RunPod Serverless	REAL_INFRA_REQUIRED
AWS EC2/ASG/EKS/SageMaker/Bedrock profiles	BYOC and imported/managed product boundaries	provision/integration profiles	contract/manifest tests	CLI/API fixture and translation conformance	AWS account/GPU/private network	REAL_INFRA_REQUIRED
GCP Compute/MIG/GKE/Vertex profiles	BYOC and managed product boundaries	provision/integration profiles	contract/manifest tests	CLI/API fixture, immutable dependency preflight, Private Google Access enforcement, observed regional/global GPU quota fail-fast with unknown preservation, exact-model zonal Persistent Disk verification and read-only retained attachment, translation conformance, and leak-inventory safety	GCP account/GPU/PSC, stock and real cache mount	REAL_INFRA_REQUIRED
CoreWeave CKS profile	Kubernetes-first provider profile	integration/Kubernetes	contract/manifest tests	Manifest/provider contract	CoreWeave account/GPU	REAL_INFRA_REQUIRED
SkyPilot boundary	Infrastructure provisioning reuse contract	integration/provision docs	profile/contract tests	Ensure core remains adapter neutral	Real SkyPilot/provider execution	REAL_INFRA_REQUIRED
vLLM runtime	Default inference runtime and router	runtime/integration/router	runtime fixture/stack + RunPod Qwen evidence	Local fake runtime protocol/health/drain; exact RunPod Qwen3-8B/vLLM 0.22.1/L40S buffered, streaming, benchmark and cleanup passed	Other models, providers, GPUs and representative repeated performance campaigns require fresh qualification	RUNPOD_REAL_PASSED (EXACT BASELINE TUPLE) / REAL_INFRA_REQUIRED
SGLang runtime	Second qualified runtime profile	runtime/integration	profile/conformance tests	Hermetic runtime contract	AWS L40S SGLang 0.5.12 passed; smaller runtime image and other providers remain separate	AWS_REAL_PASSED / REAL_INFRA_REQUIRED
Custom OCI runtime	Digest-pinned portable workload contract	runtime/spec/provision	conformance/spec/provider tests	Complete executable argv, entrypoint override, validation/translation/probes/cancel/drain, and exact GPU count on AWS/GCP/Kubernetes/SkyPilot compilers	AWS L40S portable vLLM workload passed historically; every other image/model/provider/topology tuple needs real qualification	AWS_REAL_PASSED / REAL_INFRA_REQUIRED
Provider/runtime conformance kit	Idempotency, adoption, inventory, capability evidence	`conformance`, integration	contract qualifier	Run all deterministic adapters repeatedly	Real provider qualification tier	AUTOMATED_QUALIFIED
Kubernetes provider	Namespace/SSA/RBAC/restart/delete lifecycle	provision/manifests/tools	manifest + Kind suites	Real Kind API, restart/lost response/zero resources, operator pre-pull contract with fail-closed `imagePullPolicy: Never`, exact-model PVC	GPU scheduling and node-level pre-pull timing	AUTOMATED_QUALIFIED
KServe integration	Standard CRD-gated single-owner inference service	provision/manifests	manifest/Kind tests	Missing/present CRD and field ownership	Real KServe GPU runtime	AUTOMATED_QUALIFIED
NVIDIA Dynamo DGD adapter	One immutable parent graph for the aggregated baseline; disaggregated NIXL intent is Deferred and fails before mutation	servingcontract/provision/workflows/spec	contract, fixture, API, persistence, strict DGD-subset Kind and fail-closed translation tests	Lost apply response, changed intent, stale/failed status, exact ownership, server-side dry-run/apply/delete-to-zero for aggregated mode	Real operator/GPU; NIXL transport, disaggregated routing, cache allocation, streaming rollout and performance are unqualified	AUTOMATED_QUALIFIED (AGGREGATED ONLY) / REAL_INFRA_REQUIRED
PostgreSQL persistence/migrations	Source of truth, upgrade safety, concurrency/checksum	`store`, 51 migrations	store/migration/restore tests	Every-prefix, concurrent startup, downgrade/checksum	Customer managed HA database	AUTOMATED_QUALIFIED
HA/fencing/mixed-version	Stateless API, leased workers, stale writer rejection	store/operations/serve	HA/failure/restore suites	Multi-process Docker crash/reclaim/protocol overlap	Multi-host customer topology	AUTOMATED_QUALIFIED
Backup/restore	Guarded backup, ledger preservation, reconcile startup	scripts/store/compose	safety + Docker drill	Corrupt/mismatch/restore/reconcile	Customer backup system/RTO	AUTOMATED_QUALIFIED
TLS/mTLS/private identity	Production secure transport and rotation checks	config/safehttp/serve	config/production tests	Local TLS/mTLS invalid/rotation	Customer CA/network/IAM	AUTOMATED_QUALIFIED
Authentication/RBAC/scopes	Principal auth, hosted Clerk organization identity, roles, scopes, rotation/revocation	authn/authz/store/API/console	auth/API/store/browser tests	Boundary matrix, tenant mapping, deny-by-default entitlement, and concurrent revoke	Production SAML/OIDC domain verification, SCIM, and IdP lifecycle require hosted qualification	AUTOMATED_QUALIFIED / MANUAL_HOSTED
Tenant isolation/audit	Every resource scoped and mutation audited	store/API/gateway	store/API security tests	Cross-tenant route/API/property tests	None expected	AUTOMATED_QUALIFIED
Secret references/redaction	Reference-only secrets and no leakage	secrets/store/API/logging	secret/security tests	Canary secret through errors/logs/reports	External secret manager access	AUTOMATED_QUALIFIED
Embedded terminal UI	Read/operate workspace, responsive navigation	CLI UI	UI snapshot/interaction tests	PTY sizes, empty/error/action handoff	Human visual/keyboard review	MANUAL_REQUIRED
Browser console	Endpoint-first stable identity, workload, operation, evidence, API-key and team UI	separate web app/control API	unit + contract fixture tests	Production build + protected auth + responsive Playwright + axe + visual baselines + disposable real Go/PostgreSQL control-plane browser qualification	Hosted Clerk/private-preview and final hosted visual review	AUTOMATED_QUALIFIED / MANUAL_HOSTED
OpenAPI route contract	Server/API schema parity	apicontract/codegen	route drift tests	Generate/check every route/schema	None expected	AUTOMATED_QUALIFIED
Python SDK	Generated API plus streaming client/package	`sdk/python`	unittest/wheel/stack	Package and local stack E2E	Public-beta PyPI package; stable publication remains release-gated	AUTOMATED_QUALIFIED
TypeScript SDK	Generated API plus streaming client/package	`sdk/typescript`	node tests/npm pack/stack	Package and local stack E2E	Public-beta npm package; stable publication remains release-gated	AUTOMATED_QUALIFIED
Terraform provider	Deployment/SLO CRUD/import/adoption	`integrations/terraform`	provider tests/automation	Protocol tests against deterministic API	Public-beta GitHub binaries; Terraform Registry publication deferred	AUTOMATED_QUALIFIED
GitHub delivery integration	CI/release plan and evidence artifact handling	`.github/workflows`, scripts	repository/release checks	Static workflow validation + local scripts	Hosted Actions run	MANUAL_REQUIRED
Documentation/Mintlify	Accurate navigation, links, API and commands	`docs`	docs check/context drift	Mintlify build/link/claim audit	Hosted Mintlify rendering	MANUAL_REQUIRED
Release packaging	Multi-arch archives/checksums/SBOM/native/Homebrew	release scripts/GoReleaser	qualify/release verification	Exact version build and native smoke	Public-beta GitHub/Homebrew distribution; stable publication remains release-gated	AUTOMATED_QUALIFIED
Docker images/Compose	Provider-neutral control plane and opt-in overlays	Dockerfiles/compose	container/stack/production tests	Build/start/health/stop and config matrix	GHCR candidate publication deferred until the release vulnerability gate passes; exact runtime/provider qualification remains separate	AUTOMATED_QUALIFIED
Acceptance orchestration	Approval lock, run identity, resume, cleanup, report	acceptance scripts	safety tests/local qualification	Shell fault/safety and dry local workflow	Paid suites with user	MANUAL_REQUIRED

Start here

Deploy and connect

Operate

Evidence and decisions

Feature qualification matrix

InferCrane feature qualification matrix

​InferCrane feature qualification matrix

InferCrane feature qualification matrix