Eight-Layer Lock-In MapFind where the organisation would lose meaning, not merely files

Model behaviour includes provider-specific system instructions, tool schemas, safety rules, version aliases, tuning artefacts, latency ceilings, context handling and rate limits. A compatible request shape does not guarantee equivalent refusal, extraction, structured output or transaction behaviour. The version name also matters: an alias that silently advances can invalidate the evidence attached to yesterday's configuration.

Context and data covers source rights, transformations, chunking, embeddings, indexes, metadata, feedback links, labels, lineage and retention. The buyer may own the original documents but not possess the relationships needed to rebuild retrieval. A file archive without stable source identifiers, transformation recipes and deletion status preserves content while losing operational meaning.

Evaluation includes cases, rubrics, human labels, graders, baselines, thresholds, telemetry and release decisions. When the provider owns both the system and the only measurement console, it can mediate the buyer's view of substitute quality. Retaining prompts is insufficient if grader versions, adjudication, historical score distributions and production slices disappear.

Workflow and state includes orchestration, prompts, tools, identities, policies, approvals, queues, retries, memory, active cases and compensation logic. Code portability is not state portability. The risky moment is not starting one new case on a replacement. It is deciding what happened to every in-flight action, unknown commitment, pending approval and recovery obligation.

Human practice holds tacit procedures, training, review habits, exception handling, support and professional judgement. Supplier-operated reviewers may quietly repair model failure or classify edge cases. If those interventions are not observed and transferred, the replacement inherits an incomplete process description and apparently regresses for reasons the buyer cannot diagnose.

Control and evidence contains audit events, provenance, monitoring, incident records, assurance mappings, retention and regulator access. Downloaded logs do not reproduce an assurance record when event semantics, identities, configuration references and derivation links are absent. W3C PROV-O offers a portable vocabulary for entities, activities, agents and derivations, but the organisation must still capture those relationships at the operational source.

Infrastructure and operations spans cloud services, accelerators, deployment formats, network controls, observability, capacity, recovery and service management. Open weights can remain captive to specialist kernels, hardware, a particular serving stack or knowledge concentrated in one team. A second model on the same region and vector store is diversity at one layer, not independent continuity.

Commercial and legal includes licences, minimum commitments, egress charges, intellectual-property rights, subcontractors, pricing changes, assistance, deletion and remedies. A promise of termination assistance is not executable unless it names formats, time, quality, responsibilities, cost and evidence. The map distinguishes expensive visible dependencies from hidden edges. The hidden edge is more dangerous because no budget, owner or closing test exists for it.

Exit Complexity GraphTrace the critical path from dependency to damaged capability

The Lock-In Map inventories layers. The Exit Complexity Graph makes their relationships actionable. Each directed edge connects a source component to the capability, state or evidence that depends on it. The record names an owner, substitutability, export format, semantic loss, validation work, migration time, interruption risk, cost, rights and supporting evidence. It also states the condition under which the edge is considered closed.

Consider a proprietary embedding service. Its obvious target is the vector index. Less visible targets may include chunk identifiers in evaluation cases, feedback records tied to neighbours, monitoring thresholds based on retrieval scores and incident evidence that refers to provider event IDs. Re-embedding the documents creates a new representation, not a continuation of the old one. The graph forces the team to traverse every dependent edge, decide which histories can be translated, and mark where comparisons must restart.

Architecture shows what calls what. Exit complexity shows what institutional result is lost when an edge breaks and how the loss will be repaired. Supplier specialists, incident judgement and licensed test corpora belong in the graph because they govern practical operability.

Paths should be tested for conjunction. The replacement can parse the export, but cannot satisfy the identity policy. It can reproduce representative answers, but the active queue has no stable transaction identifiers. It can run in a second cloud, but both paths depend on the same control plane. The critical path is the longest credible sequence through rights, export, transformation, validation, rehearsal, migration, reconciliation and termination. Adding nominal providers does not shorten it unless a shared edge is removed.

Capability Continuity EnvelopeDefine what must survive before selecting the substitute

The envelope names the minimum state that a replacement must preserve. It begins with the bounded business outcome, authorised population and accepted ranges for quality, harmful effect, abstention, latency and cost. It preserves identity, tool and policy semantics so that an apparently better model cannot acquire broader authority or bypass a prohibited action. It accounts for active work, including safe completion, cancellation, rollback and reconciliation.

Evidence is part of the capability. Source lineage, decision receipts, configuration, retention and historical baselines must remain available to the people who release, investigate, audit and challenge the system. Monitoring, incident handling, escalation and recovery need working owners and channels after migration. Human roles, tacit knowledge and support capacity move or are deliberately replaced. Legal rights and regulator access continue for the relevant retention period even after the old provider is gone.

The envelope defines equivalence where it matters and allowed difference where it does not. A replacement may use a different embedding model or orchestration engine, but must meet the intended outcome, respect authority, preserve evidence meaning and handle in-flight work within approved limits. Functional equivalence is a starting idea; AI continuity also carries behaviour, human practice and evaluation.

Write the envelope before a renewal, outage or commercial dispute. During calm operation, leaders can decide which histories may be retired, which evidence must remain comparable and which population can tolerate an interruption. During an incident, every unresolved question lengthens the critical path. The envelope turns “maintain service” into a testable obligation and exposes when safe stop is the only honest form of continuity.

Replaceability BudgetPrice the option before paying abstraction tax

The budget sets four independent limits: maximum switching time, maximum service interruption, maximum one-off cost and maximum tolerated outcome or evidence loss. These are gates, not ingredients of a single portability score. A replacement that is cheap but interrupts a critical public service for too long fails. A fast migration that destroys the assurance history required for challenge also fails. Consequence determines the budget, not a universal preference for open or closed technology.

Switching time includes discovery, rights clearance, export, transformation, validation, human preparation, shadow operation, active-work migration, reconciliation and termination. Interruption covers degraded capacity and safe-stop periods, not merely network downtime. One-off cost includes engineering, duplicate operation, data movement, re-evaluation, control repair, retraining, assurance and supplier assistance. Tolerated loss names which behaviour, comparability, history or evidence may expire and who accepts that residual risk.

The Lock-In Decision compares dependency rent with switching option value. Rent includes price premium, constrained negotiation, change exposure and exit work. Option value is avoided damage if the provider changes terms, quality, service or direction. A proprietary capability can win when differentiated value is material and rehearsed exit stays within budget. Remove an abstraction that forces inferior behaviour while protecting a low-value switch.

Do not pay abstraction tax everywhere. Protect the seams that carry institutional truth: identifiers, source lineage, case sets, thresholds, authority decisions, active state, receipts and recovery. Provider-specific prompts, optimised runtimes or specialised tools can live behind those seams. Selective replaceability preserves the option that matters without pretending every component is interchangeable.

Portability Proof LadderClimb from words to a completed exit

Level one is a contractual statement. It creates a right and a basis for remedy, but no evidence that the provider can perform it. Level two produces the export. The team checks completeness, format, timing and cost. Level three parses it independently in an environment controlled by the buyer or replacement. This is where missing identifiers, undocumented schemas and inaccessible metadata become visible.

Level four runs representative and challenge cases on the replacement. Evaluation remains buyer-controlled, with versioned graders, human adjudication and historical baselines. Level five completes a shadow operating window under realistic load, tools, policy, reviewers, monitoring and recovery. A good fresh-work result may still hide queue, exception or tail-latency failure, so the window uses the full continuity envelope.

Level six migrates or safely closes active work. Pending approvals, possible commitments, retries, memory, escalations and compensation obligations receive explicit terminal states. Receiving systems, not the departing agent or provider, confirm external effects. Level seven verifies that old provider access is revoked, retained data is deleted or lawfully retained under a named schedule, subcontractor duties are closed, and the evidence needed for audit or dispute remains accessible.

The ladder prevents a backup from being described as migration proof. It also prevents a contract clause from becoming permanent comfort. Each high-consequence dependency has a target rung and rehearsal frequency. A low-impact drafting assistant may need only independent parsing and representative cases. A capability that can commit payments, determine access or carry regulated records may require active-work migration and verified termination before expansion.

Selective replaceabilityProtect the seams, then permit differentiated machinery

Selective abstraction begins with observed differences. A common interface preserves material variation in tool schema, safety policy and context economics. Build the seam after the exit graph identifies option value, not before.

Dual qualification is appropriate when consequence is high, demand can be split safely and the alternate path is genuinely independent. The second provider needs current cases, capacity, rights, operations and active-state handling. It should not share the critical cloud region, vector service, identity broker, grader or supplier-operated review team without that concentration being acknowledged. Warm readiness is a continuing operating duty, not a logo on an architecture diagram.

Bounded lock-in is appropriate when a provider-specific feature creates material value and exit remains exercisable. The contract pins required rights and change notice, the team owns evaluation and institutional state, the budget is authorised, and a scheduled drill shows how the system degrades or migrates. The business owner accepts the residual dependency because the benefit exceeds the priced option cost.

Decomposition is necessary when multiple layers are trapped inside one managed object. Extract identity, evidence, active state or evaluation from the vendor boundary before renewal or scale. Stop expansion when rights are absent, exports cannot be interpreted, active effects cannot be reconciled, the substitute cannot be validated, or the critical path exceeds the approved budget. A graceful restriction now is cheaper than discovering an impossible exit during an incident.

Open and closed labels do not decide practical freedom. Open weights may improve inspection, deployment choice or continuity, yet training data, licence conditions, specialist runtime, hardware kernels and scarce operating skill can still form hard edges. A closed service may offer strong, tested export and assistance rights. The same map, graph, envelope, budget and ladder should judge both.

Ten worked casesTraverse all eight layers before deciding what to keep

1. The API-compatible model that changes the work

The model layer changes tool arguments, refusal patterns, response length and token economics while the interface remains stable. The graph reaches workflow retries, reviewer queues, cost thresholds and customer completion. The continuity envelope requires the same authorised outcome, not identical wording. The time and interruption budget is tight. The team preserves its typed tool contracts and buyer-owned regression cases, then climbs the ladder through representative work and a shadow window. Decision: preserve provider flexibility through a tested abstraction at the tool and evidence seams, while allowing model-specific prompting behind them.

2. Documents move; retrieval meaning does not

Source files export, yet embeddings, chunk identifiers, metadata, feedback links and transformation lineage remain trapped. The map marks context, evaluation and evidence red. Graph traversal shows that prior relevance judgements and incident records cannot attach to reconstructed passages. The envelope requires source lineage and comparable decision evidence; the tolerated-loss budget rejects silent history loss. An independent parser proves only the raw export, so the ladder stops at level three. Decision: decompose the retrieval layer before renewal, establish stable buyer identifiers and recipes, rebuild representative cases, and postpone scale until continuity is demonstrated.

3. The incumbent grader prefers the incumbent voice

A candidate model performs well under human task review but loses against a proprietary automated grader tuned to the old provider's style. Evaluation, telemetry and release evidence are captive even though data and workflow are portable. The graph reveals a circular edge from provider output to provider judge to historical baseline. The envelope protects task outcome and harmful-error thresholds, not stylistic similarity. The budget permits a short re-baselining period. The buyer exports cases and labels, versions independent graders, adjudicates disagreement and reruns the shadow window. Decision: decompose evaluation ownership; no model comparison is valid until the proof ladder uses a buyer-controlled measurement path.

4. Portable workflow code, trapped active state

The orchestration definition can be copied, but approvals, retries, queues, conversation memory, possible tool commitments and compensation state live inside a managed service. The graph's critical path runs through every active case and receiving system. The envelope refuses a fresh-start migration because unknown effects could be duplicated or abandoned. Switching time and interruption exceed budget. Export and parser tests reach level three, but active-work proof fails. Decision: decompose the workflow before renewal or scale by externalising stable case and transaction state, then rehearse safe completion, cancellation, rollback and reconciliation.

5. Source code without operating knowledge

The buyer owns repositories and infrastructure definitions, yet the supplier's two specialists resolve exceptions through undocumented judgement. Human practice, support, incident evidence and recovery remain captive. The graph ties specialist decisions to service restoration and high-risk case escalation. The continuity envelope requires runbooks, trained roles and demonstrated capacity. The budget includes rehearsal and workforce time rather than engineering alone. A clean deployment is only level four evidence if realistic exceptions still require the incumbent. Decision: accept no claim of operability until paired staff complete a shadow window and recovery drill; price transitional support and knowledge-transfer remedies explicitly.

6. Downloadable logs that cannot prove a decision

Audit files exist, but event names change, identities are ambiguous, model configuration is missing and provenance cannot connect a source to a decision and effect. The map marks control and evidence as trapped. The graph reaches regulatory response, incident reconstruction and customer challenge. The envelope requires stable semantic events and lineage for the retention period. Zero service interruption does not compensate for intolerable evidence loss. The ladder stays below independent parsing because the content cannot be interpreted. Decision: stop expansion of consequential use, define portable event contracts and provenance links, and require assurance reconstruction before the provider can claim an exit-ready log export.

7. Two models, one hidden failure boundary

A gateway routes across two model providers, but both use the same cloud region, vector database, identity broker, monitoring service and human review supplier. Model diversity hides infrastructure, data, control and people convergence. The graph shows one critical path and one recovery team. The envelope requires service and evidence continuity through a relevant shared failure. The budget does not tolerate the common-region interruption. A representative model test reaches level four but proves no independent operating path. Decision: maintain two qualified providers only after removing or governing the shared dependencies that matter; otherwise describe the design as model routing, not continuity.

8. Differentiated capability earns bounded lock-in

A proprietary multimodal feature materially improves an industrial inspection outcome that substitutes cannot yet match. The eight-layer map shows provider-specific model and infrastructure edges, while data, evaluation, workflow state, authority and evidence stay under buyer control. The graph's exit path is visible and timed. The envelope permits temporary feature loss under a safe manual mode. The budget funds migration and a short interruption. The ladder has completed a shadow exit on the nearest substitute. A generic adapter imposes latency and removes the valuable feature, so it is deleted. Decision: accept bounded lock-in with priced exit, change rights and scheduled proof; abstraction cost exceeds the option it protects.

9. Open weights, captive runtime

The model weights are available, but efficient operation depends on hardware-specific kernels, a restrictive ancillary licence, a scarce serving team and provider-tuned safety components. The map correctly marks model access as open and infrastructure, people, control and commercial edges as constrained. The graph shows that a nominal deployment elsewhere misses capacity and policy behaviour. The envelope requires operating ranges, authority and recovery, not possession of files. The budget exposes a long training and optimisation path. The ladder has produced and parsed artefacts but not survived a shadow window. Decision: accept only bounded use while building a qualified runtime or choose a simpler substitute.

10. The termination drill discovers the real service

A scheduled exercise exports representative and active cases, parses them independently, runs the replacement under live-like load, reconciles pending effects, validates evidence continuity and verifies deletion. The eight-layer map appears mostly green until supplier-operated reviewers fail to transfer rare-case judgement. The graph adds this hidden people edge and extends the critical path. The envelope requires the affected cases to enter a restricted queue while staff train. The interruption remains inside budget. After a second shadow window, the ladder reaches verified termination. Decision: preserve tested flexibility and amend the contract, runbook and owner map so the newly visible human dependency cannot hide again.

Interactive decision artefactLock-In Escape Planner

The Planner returns one dependency posture from observed conditions. It does not calculate a portability score or recommend multi-cloud by default. Each named case changes the same layer, graph, envelope, budget and proof fields used for a real renewal decision.

Exit operating cadenceMake replaceability evidence expire unless it is renewed

At capability design, create the eight-layer map and establish buyer-owned identifiers for sources, cases, decisions, active work and receipts. Define the continuity envelope and four-part budget before selecting an architecture. Classify seams by consequence and option value. Permit specialist features where they earn their dependency, and resist adapters whose permanent cost is greater than the realistic option protected.

At contracting, translate graph edges into precise rights. Name export formats, semantic documentation, testing access, knowledge transfer, assistance time, price, subcontractor duties, deletion evidence and remedies. OMB M-25-22 and the current GAO acquisition evidence support attention to knowledge transfer, portability, licensing, transparent pricing, lifecycle testing and post-contract rights. Terms create the ability to demand evidence; the proof ladder establishes whether the ability works.

At release and change, keep evaluation cases, grader versions, human labels, thresholds, telemetry and decision histories under accountable control. A material model, data, tool, policy, infrastructure, supplier or pricing change updates the graph. Re-run the affected continuity evidence instead of assuming that an unchanged interface preserves the result. NIST lifecycle and supply-chain guidance is valuable here because substitution is a controlled change, not a one-time purchasing event.

At each rehearsal, produce and parse exports independently, run representative and challenge cases, operate a shadow window, migrate or close active work, reconcile receiving-system effects, test human capacity and verify access and deletion. Record actual time, interruption, cost, loss and assistance. Update the budget with observed data. A failed drill can lead to decomposition, restricted scope, stronger rights, alternate qualification or a stop.

At renewal, compare differentiated value with dependency rent and current option value. Do not renew solely because switching is already expensive; that converts sunk cost into permanent surrender. Do not switch solely to demonstrate independence; that can destroy value. The accountable owner presents the current heat map, critical path, envelope, budget, ladder level and unresolved hidden edges. The decision is explicit: tested flexibility, dual qualification, bounded lock-in, decomposition or stopped expansion.