Executive briefThe endpoint can move quickly while the capability remains trapped

A provider announces a model retirement. Engineers redirect an OpenAI-compatible endpoint within a day and the replacement accepts every request. Structured outputs still parse. Yet omission behaviour changes, safety refusals move, retrieval filters no longer match the embedding space and several agent tasks lose in-flight state. The interface migrated; the business capability did not.

Architectural portability is the tested ability to preserve an essential outcome, its decision contract, control evidence and recoverable state when a model or provider becomes unavailable, uneconomic, non-compliant or strategically unacceptable. It does not promise zero lock-in or perfect substitutability. A replacement can be syntactically compatible and still fail task quality, latency, authority, audit or customer continuity.

The Exit Surface Map in this article treats the governed business capability as the unit. It records twelve dependency layers, from outcome and decision semantics through data, state, tools, evidence, commercial terms and transition machinery. Each layer names an owner, export format, semantic conversion, recovery target, rehearsal evidence, expiry and residual loss.

The Portability Claim Ladder then prevents a lower-level success from masquerading as full exit. Syntactic portability, execution portability, behavioural portability, control portability and outcome continuity are five separate claims. The lowest unproven rung is the active migration risk. Open standards and adapters can reduce selected switching costs; they cannot prove equivalent behaviour or results.

The strongest rival: deliberate concentration

Deep use of one provider can improve quality, speed and operating simplicity. Premature abstraction can flatten useful features, add latency and create a permanently under-tested duplicate stack. The defensible alternative is bounded concentration: price switching loss, preserve critical assets, fund a manual or technical contingency and name triggers that reopen the exit decision. Concentration becomes unmanaged only when its dependencies, residual loss or recovery plan remain unknown.

Exit Surface MapTwelve layers expose what must move, rebuild or recover

Classify each dependency as portable, rebuildable, substitutable, recoverable only by fallback or accepted concentration. The label is supported by a restore, conversion or rehearsal result. “Export available” is not a recovery claim until another environment consumes the asset and reconciles it with the business state.

Portability Claim LadderProve five rungs without allowing implication to climb ahead of evidence

Syntactic portability means another endpoint accepts and returns compatible shapes. Execution portability adds required tools, data, state and operating limits. Behavioural portability requires the replacement to meet task cases and critical-failure bounds. Control portability preserves authority, privacy, provenance, monitoring, appeal and audit. Outcome continuity demonstrates the capability within its disruption tolerance after cutover.

Each rung needs a separate artefact and rehearsal. OpenAPI 3.2.0 can describe an HTTP interface. OCI image specifications can package content-addressed software artefacts. ONNX 1.24.0 represents model graphs and operators. SLSA 1.2 can support build provenance. These standards reduce particular movement or integrity costs, yet runtime operators, semantics, hardware, evidence and outcomes can still diverge.

Record the lowest unproven rung. If a candidate executes but fails refusal or omission cases, behavioural portability is the blocker. If behaviour passes but decision receipts or appeal records disappear, control portability is unproven. If all controls transfer but customers experience unacceptable disruption, the final outcome claim fails. Technical completion cannot close a business or control exception.

Hidden couplingThe endpoint is the visible tip of a capability dependency

A gateway can redirect calls and a stable schema can reduce integration change. Neither moves the prompt policy, evaluation set, context rights, workflow checkpoints, tool side effects, operational knowledge or appeal evidence. Hidden technical debt research describes entanglement, undeclared consumers, data dependencies and configuration debt. Model exit concentrates those liabilities into one deadline.

Open weights are not automatically portable. Operation may depend on a particular accelerator, quantisation method, serving engine, licence interpretation, safety configuration and scarce team. A managed API can be more recoverable when behaviour cases, data rights, state and controls remain outside the supplier. Ownership of weights is one asset classification, not the capability verdict.

Inventory consumers that do not appear in the interface catalogue. A generated field may feed a manual queue, a data product or a downstream control. A change in refusal or omission behaviour can alter those consumers even while schemas pass. Dependency discovery combines configuration, telemetry, lineage, contracts, owner interviews and live rehearsal rather than trusting architecture diagrams alone.

Migration controlPrepare, freeze, reconcile and cut over with a usable rollback

A planned transition begins by freezing the decision contract and representative cases, exporting recoverable assets and proving the candidate in shadow. A bounded dual run compares behaviour, controls, cost, latency and capacity under a decision rule and end date. Before cutover, stop new admissions where necessary, drain or checkpoint work and reconcile in-flight commitments by business identifier.

An emergency exit compresses the same states rather than skipping them. Freeze unsafe admissions, activate the manual or technical fallback, preserve evidence and identify work whose commit state is uncertain. Restore from tested assets, reconcile side effects, then admit a limited population to the candidate route. A backup that has never been restored and joined to live business state is an aspiration, not continuity evidence.

Verification continues after traffic moves. Confirm critical cases, authority, provenance, monitoring, appeal access and the capability recovery objective. Roll back when a named critical failure, tail condition, control gap or outcome threshold occurs. Rollback must handle state created during the candidate interval; it is not simply a DNS reversal. Decommission only after evidence archives, data return or deletion, contract closure and residual dependencies have owners.

Recoverable contextMove meaning, state and commitments before changing the model

A RAG capability needs more than source documents. Preserve document versions, licences, chunk rules, transformations, embedding model, vector dimensions, index configuration, entitlement metadata, freshness logic and evaluation queries. Exported vectors without the source, chunk identity and access rules cannot be safely interpreted or rebuilt. Re-embedding changes neighbourhoods and may reorder retrieved evidence, so behavioural regression is required after a technically successful import.

An agent capability carries conversations, task state, memory, checkpoints, tool receipts, idempotency keys and external commitments. Classify each item as authoritative business state, reconstructable context, ephemeral scratch state or prohibited retention. Rebuild authoritative state from systems of record where possible. Do not copy hidden reasoning. For every in-flight task, determine whether the prior provider completed, failed or left an uncertain side effect before resuming elsewhere.

Tool semantics can differ even when function names and JSON arguments match. The replacement must preserve downstream authorisation, preconditions, confirmation, transaction receipts, cancellation and compensation. Event schemas and orchestration checkpoints need version-aware conversion. If an agent wrote a record just before the freeze, reconciliation prevents the new route from repeating the effect under a fresh invocation identifier.

Control evidence moves with the case. Route receipts show which candidate handled an attempt; decision receipts link evidence, authority and commitment. Retain their identifiers and access rules across telemetry backends. OpenTelemetry 1.60.0 supports vendor-neutral signal concepts, while PROV-O can preserve derivation and agency relationships. Neither supplies the domain meaning or retention duty automatically.

Strategic economicsPrice concentration instead of pretending it can be eliminated

Portability has a carrying cost. Abstraction requires design, testing and sometimes performance compromise. Duplicate operation requires capacity, licences, skills, security, monitoring and continual regression. Switching loss includes outage, conversion, revalidation, customer impact and stranded assets. Compare those costs over an explicit horizon rather than treating optionality as free.

Bounded concentration is rational when a capability is non-critical, the manual fallback is credible, switching loss is tolerable and duplicate operation costs more than plausible disruption. Record the accepted dependency, maximum loss, protected assets, contingency funding and trigger. A trigger can be deprecation, sustained effective-cost change, control failure, geographic constraint, capacity deterioration or contract milestone.

Concentration is not only provider count. Two suppliers can share a cloud, base model, accelerator, dataset, identity plane, integrator or scarce skill. Nominal dual sourcing can therefore preserve the same failure mode. Conversely, one managed provider plus portable decision cases, data rights, state and a tested manual process may provide meaningful recovery. Measure dependency topology rather than logos.

Dual running needs a hypothesis, population, comparison rule, budget and end date. Permanent shadow capacity can become an expensive comfort signal if it never receives realistic state or traffic. Close the rehearsal when the claim is proven, rejected or superseded. Keep only the readiness needed for the recovery target.

Rehearsal windowUse notice time to convert unknown dependencies into tested evidence

Start before a supplier notice. Maintain an owner and expiry for each surface, exercise representative restores and keep contract notice periods visible. When an exit trigger fires, establish one migration clock tied to maximum tolerable disruption and the contractual deadline. Separate asset export, semantic conversion, behaviour calibration, operational proving and controlled cutover.

The rehearsal receipt lists candidate and versions, surface classifications, lowest unproven rung, cases executed, recovery time, recovery point, control results, residual loss and owner. Failed tests are valuable when they expose a dependency early. A rehearsal that uses synthetic requests but omits live-scale queues, human roles, appeal access or in-flight state proves a narrower claim and must say so.

Board oversight focuses on critical capabilities, concentrated dependencies, expired proof, unfunded recovery and triggers near breach. Executives fund the transition mechanism and accept residual loss. Product and domain owners judge outcome continuity. Platform teams execute migration. Procurement and legal secure assistance, data return and termination rights. Risk and assurance challenge evidence without owning the daily service.

Capability testsTen cases separate movement from continuity

The second and seventh cases are deliberate negative controls. In one, a replacement succeeds technically but fails the business behaviour boundary. In the other, outputs meet task tests while evidence and appeal continuity fail. Neither cutover is complete. The tenth case rejects architectural theatre by allowing a priced and governed concentration outcome.

Every case reports current rung, lowest unproven rung, assets at risk, capability recovery target, rehearsal result and residual dependency. Public benchmarks can help construct tests but cannot replace them. Acceptance uses the actual task, cohort, tools, data, controls, load and human operating model.

A worked exit drill for a consequential service

Consider a claims-support capability whose model supplier announces retirement in six weeks. The service retrieves policy and case evidence, prepares a recommendation for a human reviewer and can create a payment instruction after approval. Its maximum tolerable disruption is one business day, while a constrained manual workflow can serve urgent claims at lower throughput. The capability owner opens the transition record, fixes the decision contract and stops feature changes that would invalidate comparison.

The first rehearsal inventories the surface rather than selecting a model immediately. It finds an OpenAPI contract, external prompt repository, labelled decision cases and source documents under enterprise control. It also finds provider-managed embeddings, a hidden chunk identifier in citations, reviewer sessions stored in the agent runtime and a payment tool whose idempotency key is derived from the old task identifier. The apparent endpoint swap is therefore only syntactically portable.

Data and state owners export documents with versions, licensing and entitlement metadata, reconstruct chunk lineage and rebuild the index with a new embedding model. They measure retrieval changes on evidence-bearing cases rather than comparing vector files. In-flight sessions are classified into unstarted, working, input-required, approved and commitment-uncertain states. The last group is reconciled against the payment system before any task is resumed, preventing a second instruction after a hidden commit.

The candidate then runs in shadow against the frozen evaluation population. Interface conformance passes, but behavioural tests reveal that the candidate omits one exclusion more often in complex cases and refuses a permitted class differently. The team adjusts the behaviour specification and surrounding deterministic checks, then repeats representative, adversarial and critical-failure cases. This is a rebuild behind the stable decision contract, not evidence that the models became interchangeable.

A bounded dual run follows for cases that create no automatic effect. The old and candidate paths receive the same authorised snapshot, while reviewers see only the production result to avoid contaminating judgement. The rehearsal compares decision-quality rubrics, refusal and omission classes, retrieval citations, first-useful and completed latency, effective cost, reviewer changes and evidence continuity. It has a fixed traffic population, budget, end date and rule for selecting the next state.

Before cutover, the platform freezes new admissions, checkpoints remaining work and reconciles every external commitment. The candidate opens to a limited cohort after control tests prove principal, policy, provenance, telemetry and appeal access. A critical evidence omission, missing payment receipt, tail-latency breach or unrecoverable state mismatch triggers rollback. The rollback procedure drains candidate work, reconciles its effects and restores the prior or manual route using the same business case identifiers.

Post-cutover verification observes completed claims through the declared disruption window. Only then can the owner claim outcome continuity. Decommissioning archives the model, route and decision receipts under their retention rules, returns or deletes provider data, revokes credentials, closes monitoring and preserves the regression history. The exit receipt records that proprietary historical embeddings were not portable but were safely rebuildable, and that manual throughput remains the funded fallback until the next rehearsal.

The executive review sees no portability percentage. It sees the proven rung, the rebuilt surfaces, actual recovery time and point, cases that failed and later passed, residual supplier and skill dependencies, fallback capacity and the next proof expiry. Procurement confirms termination assistance and data return, legal confirms continuing evidence duties, finance confirms transition cost, and the capability owner accepts the remaining outcome risk. This evidence turns a notice period into an accountable capability transition.

Rehearsal evidence expires when dependencies materially change. A new model version, embedding pipeline, tool contract, policy rule, supplier term, accelerator or operating team can invalidate a previously proven rung. The owner records the trigger, narrows any continued claim and repeats the affected tests before presenting the capability as recoverable again.

Interactive instrumentModel Exit Rehearsal Board

Select a named case or change the dependency record. The board returns one of exactly five actions. It reports a rung rather than a portability score because strength at one layer cannot compensate for a missing state, behaviour or control proof.

Evidence and governance boundaryExit is an operating capability, not a procurement clause

The Exit Surface Map, Portability Claim Ladder and Model Exit Rehearsal Board are this article's synthesis. Hidden technical debt research exposes system coupling. OpenAPI, OCI, ONNX, OpenTelemetry, SLSA and PROV-O reduce specific description, packaging, representation, telemetry, provenance and integrity costs. None proves behavioural equivalence, control continuity or business recovery.

Current versions checked on 2 September 2026 include OpenAPI 3.2.0, OpenTelemetry 1.60.0, SLSA 1.2 and ONNX 1.24.0. Pin the versions actually used by an exit mechanism. A “latest” reference can change after the rehearsal, and implementation-defined or unsupported features can break apparent compatibility.

NIST AI, supply-chain and security guidance supports lifecycle, supplier, configuration, contingency and evidence controls. ISO/IEC 42001 supports organisational AI management and continual improvement. DORA contains ICT concentration, contract, transition and exit expectations for applicable financial entities. United States banking guidance addresses planning, due diligence, contracting, monitoring and termination. The EU AI Act may affect records and responsibilities for applicable systems and actors. Scope is contextual; this design is not a legal conclusion.

Contracts should secure notice, assistance, audit, data return, deletion, geography, intellectual property and termination rights, but operational proof remains necessary. Procurement cannot export an index, reconcile an agent effect or validate a new refusal boundary. Likewise, engineering cannot accept business disruption or residual regulatory exposure without the authorised owner.

Review the surface map when models, routes, data, retrieval, tools, state, controls, supplier terms, skills or criticality change. Rehearse before proof expires. Exercise an urgent retirement as well as a planned commercial move. Measure recovery from the business capability, not from the first successful replacement response. Include human fallback throughput and evidence access.

A board should see critical capability, tolerated disruption, current proven rung, lowest unproven rung, concentrated dependencies, rehearsal age, residual loss, funding and triggers. It should not receive a universal portability score. One unproven control or state layer can govern the exit even if eleven other surfaces are strong.

The closing standard is practical: preserve the capability's meaning and evidence outside any single model where economically justified, accept remaining concentration explicitly, and prove transition by live rehearsal. Portability is the capacity to continue accountable business service after cutover, not the ability to make another endpoint return familiar JSON.

Source recordWorks cited

1. Sculley, D. et al. Hidden Technical Debt in Machine Learning Systems.
2. NIST. Artificial Intelligence Risk Management Framework.
3. NIST. Generative Artificial Intelligence Profile.
4. NIST. SP 800-161 Rev. 1.
5. NIST. SP 800-53 Rev. 5.
6. ISO. ISO/IEC 42001.
7. European Union. Regulation (EU) 2022/2554.
8. European Union. Regulation (EU) 2024/1689.
9. US banking agencies. Interagency Guidance on Third-Party Relationships.
10. CISA. Secure by Design.
11. OpenAPI Initiative. OpenAPI Specification.
12. Open Container Initiative. Image Specification.
13. SLSA. Specification 1.2.
14. OpenTelemetry. Specification.
15. W3C. PROV-O.
16. ONNX. Introduction to ONNX.
17. Mitchell, M. et al. Model Cards for Model Reporting.