AI Strategic Resilience: Avoiding Model Lock-In and Staying Adaptive as Models Change
On 27 February 2026, according to the Congressional Research Service, the US President directed federal agencies to stop using Anthropic's technology, and the Defense Secretary designated the company a supply-chain risk. Firms that had built workflows on those models had done nothing to cause it: this is AI model lock-in exposed by events nobody in IT controls. AI model lock-in is the state where replacing the model behind a workload takes so long or costs so much that you cannot do it inside the notice you will be given. This lesson shows why that notice is shorter and less predictable than your architecture assumes, then runs a half-day workshop and a 24-hour switch drill that leave you with a Model Resilience Plan: a dependency map, fallback and switch runbook for every workload that matters.
The AI Model Lock-In Problem and What It Costs
The day a vendor became an AI supply chain risk
The Congressional Research Service report, updated 14 October 2026, sets out the sequence. The directive followed what CRS calls a reported months-long dispute between the Defense Department and the company over certain uses of its technology; the company says it had asked for two exceptions, fully autonomous weapons and mass domestic surveillance of Americans. Agencies such as the Defense Department were given a six-month phase-out period.
On Monday 2 March 2026, the first working day after the directive, the private sector had to decide. CRS notes that some defense contractors stopped using the models while others waited.
Then came the courts. CRS lists a preliminary injunction on 26 March 2026. On 27 August 2026 the same judge vacated and permanently enjoined enforcement of one designation (under 10 U.S.C. section 3252). On 25 September 2026 the D.C. Circuit upheld the other (under 41 U.S.C. section 4713), so the Defense Department may keep removing the company's products from its systems, barring further court action. Check for later rulings before you repeat this.
This lesson takes no side in that dispute; treat it purely as a dependency problem. The customers' prompts were fine and their evaluations passed, yet for some of them using a model they had built workflows around became a legal question overnight.
Access to a model can change for reasons that have nothing to do with you: a government dispute, a competitor conflict, an acquisition, an export rule. Your plan has to work whichever one arrives.
Two pieces of advice from that episode shape the workshop. VentureBeat, on 27 February 2026, in "Anthropic vs. the Pentagon: what enterprises should do", was blunt: keep a "warm standby" provider, and if you cannot change providers within a day, your supply chain is fragile. Do not hard-code agents to one provider's API, and benchmark on your own workloads. The Cloud Security Alliance's research note of 11 March 2026 added the governance half: inventory model provenance, including models inside SaaS products; assess concentration; estimate migration time for critical workloads only; and add contract clauses on continuity, portability and notification of government action.
The routine version arrives far more often
Every firm will face the routine case. Picture an illustrative mid-size company: a polite email says the model behind its claims agents, service assistant and document pipeline retires in 60 days. It lands in a shared mailbox nobody owns, and by the time an architect forwards it, eleven days have gone. Which applications use the model? "We think fourteen", each calling the API directly with its own copy of the prompts, tuned for months to that version's habits. Is any substitute good enough for claims triage? Nobody has tested one.
The CIO assumes procurement negotiated longer notice. Procurement assumes the architects built for portability. Finance assumes someone budgeted for running two models in parallel. Business continuity assumes the disaster recovery plan covers AI. So nobody owns the one question that matters: if this model goes away, how long until the work is running again at acceptable quality?
Two more ways a model disappears
- A competitive conflict. Anthropic cut the coding tool Windsurf's direct access to Claude models in early June 2025, amid reports that OpenAI was acquiring Windsurf (TechCrunch, 3 June 2025); by Windsurf's account, with less than five days of notice.
- An export rule. The same CRS report records that on 12 June 2026 the US Commerce Department required Anthropic to restrict foreign nationals' access to two newly released models. The company disabled both for all customers until the controls were lifted on 30 June 2026: an 18-day gap. Only the CRS report records this episode, so treat the commercial details as unconfirmed. For readers outside the US, this is the sovereignty story: a rule about nationality switched off a model for everyone.
How much notice do AI providers give before retiring a model?
For routine retirements, expect about two months' notice, often. Anthropic's published model deprecation policy promises at least 60 days' notice before retiring public models. Its log shows ten deprecation announcements since September 2024, eight between June 2025 and September 2026: about one every two months. Recent notice periods ran 60 to 62 days: Claude Sonnet 4 and Opus 4 were announced for retirement on 14 April 2026 and retired on 15 June 2026.
Terms differ by provider and contract. AI model deprecation planning needs a working assumption, and a reasonable one (an inference from logs like this, not a published fact) is that a model version lives roughly 12 to 18 months. The previous lesson had you time a requalification; if it took nine weeks against a 60-day window, you are locked in, whatever your architecture diagram says.
What model lock-in costs, beyond the scramble
No published source gives a measured cost of a forced model migration, so do not let anyone quote you one. The evidence does show large stakes.
- The spend is material. AI spending runs at about 3.3% of revenue (BCG press release, 30 September 2026, survey self-report). That belongs in the risk register.
- Supply may stay tight. The IEA's 2026 update of "Key Questions on Energy and AI" reports data-centre electricity demand of 485 TWh in 2025 and forecasts about 950 TWh by 2030, with a high-bandwidth-memory shortage to at least the end of 2027. Plan reserved throughput and a second source.
Where this lesson sits on the ladder
This is Level 5, AI Native, lesson 5 of 5, and the last lesson of the AI-Ready Foundation program. Lessons 21 to 23 gave you a Process Redesign Charter, an Agent Register and a Feedback Loop Canvas with your own eval set. The previous lesson, Continuous AI Assurance, wrapped them in an Assurance Evidence Matrix and measured how long requalification takes. Now you ask what all four depend on: what happens when the model underneath changes, with 60 days' notice or none? There is no next lesson; the closing section sends you back to Lesson 1 to re-walk the whole ladder.
How Strategic Resilience Works: The Hospital Generator
A hospital does not wire a backup generator to every vending machine and corridor light. Someone decides which circuits are life-critical: the operating theatres, intensive care, the blood fridge. Those, and only those, switch over automatically.
Then it does the part most buildings skip. On a fixed schedule, on purpose, it cuts mains power and times how long the generator takes to pick up the load. Too slow is a finding, fixed before the real outage. And in the facilities office sits a fuel contract saying how fast a delivery arrives.
Now map it. The critical circuits are your tier 1 workloads. The generator is a qualified fallback model. The scheduled test is the switch drill. The fuel contract is your notice and continuity clause. The vending machines are low-stakes features that can go dark for a week.
What is AI model lock-in, precisely?
AI model lock-in is when replacing the model behind a workload would take longer, or cost more, than the notice you will be given. It is about time against notice, not vendor count. A firm with one provider and a rehearsed two-day switch is less locked in than one with three providers and no tested way to move anything. Avoiding AI vendor lock-in starts with a stopwatch, not a tender.
Lock-in builds up from three habits, none of which looks dangerous on the day:
- Direct calls from many applications. Each team calls the provider's software development kit directly, so a change of model becomes fourteen projects with fourteen backlogs.
- Prompts fitted to one model's habits. On the next model they fail quietly: longer answers, a different refusal style, a date format that breaks a downstream parser.
- No evaluation set of your own. Without one, a substitute cannot be qualified, only hoped for. You built it in Lesson 8 and made it living in Lesson 23. Here it tells you whether the generator can carry the load.
What is strategic resilience for AI?
Strategic resilience is the ability to keep critical AI workloads running, at acceptable quality and cost, when a model is retired, restricted, repriced or withdrawn.
Two more terms. A model gateway (or abstraction layer) is a single internal service that all your applications call, which routes each request to the model approved for that workload. It is the heart of a model-agnostic AI architecture: applications name a workload, never a model. The gateway in Lesson 16's internal AI platform is the start of one. An open-weight model is one whose trained parameters you can download and run on infrastructure you control: nobody else's dispute can switch it off, but you own the hosting and the security review.
The three layers of portability
Portability is the property that lets you move a workload to another model inside your notice window. It has three layers, and most firms build only the first.
| Layer | What it contains | The hospital equivalent |
|---|---|---|
| Technical | A gateway in every request path; prompts, tool definitions and evals in provider-neutral form in your own repository; an eval set ready to requalify a substitute | The changeover switch and the wiring to critical circuits |
| Contractual | Notice period, continuity commitments, a transition period, change-of-control and assignment clauses, export of your data and fine-tunes, notification of government action | The fuel delivery contract |
| Organizational | A rehearsed switch, one named owner for model changes, a budget for running two models side by side during a move | The scheduled test and the facilities manager who signs it off |
The layers fail independently. A perfect gateway does not help if the contract allows short notice, a great contract does not help if your tool definitions take a week to rewrite, and neither helps if nobody has ever flipped the switch.
Do we need a backup model for every workload?
No. The Cloud Security Alliance's note of 11 March 2026 explicitly does not recommend engineering redundancy for every workload. Tier your workloads by business impact and keep a tested warm standby only for the top tier. A warm standby is a fallback model whose credentials are live, routing configured and capacity available, so switching is a configuration change rather than a project. That also answers the CFO: a standby for every feature doubles testing work, and no standby saves money until the claims operation stops.
When the exit plan is a legal duty: DORA Article 28
For an EU bank, insurer, investment firm or other financial entity, the Model Resilience Plan is not optional good practice. Since 17 January 2025, Article 28(8) of the Digital Operational Resilience Act (DORA), Regulation (EU) 2022/2554, has required financial entities to "put in place exit strategies" for ICT services supporting critical or important functions. A DORA exit strategy for AI is that duty applied to a model provider.
The strategy must cover provider failure, "a deterioration of the quality of the ICT services provided", disruption, material risk to the service, and contract termination. Exit must not disrupt the business, limit regulatory compliance or harm service to clients. Plans "shall be comprehensive, documented" and "sufficiently tested and reviewed periodically", with identified alternative solutions and transition plans. Article 28(3) adds a register of information on all contractual arrangements for ICT services, and Article 30(3)(f) requires contracts for critical or important functions to include exit strategies with a mandatory adequate transition period.
Read against this lesson, deterioration of quality is a silent model change, the alternative is your qualified fallback, and "sufficiently tested" is the switch drill. DORA defines ICT services broadly, as digital and data services provided through ICT systems on an ongoing basis, so assume (an interpretation, not a supervisory ruling) that a model API behind a critical function belongs in your register and exit plans, and confirm scope with compliance against the Official Journal text.
What is a multi-model AI strategy?
A multi-model AI strategy means choosing the right model for each workload and keeping the ability to change that choice, rather than standardizing on one provider. It is additive, not a divorce from the frontier labs. A September 2026 survey by ETR of 200 enterprises, reported by Techstrong.ai on 22 September 2026, found open-weight models carried 34% of enterprise AI tokens, up from 23% a year earlier, with respondents expecting 41% within 12 months (a forecast). Among production users, 60% had moved some workloads off proprietary models, while 93% said their proprietary usage was still growing.
Read those two numbers together: firms are spreading workloads, not changing camps. On open-weight models for enterprise use, the survey found cost drove 69% of adoption, and export-control and sanctions exposure had become three times more common as a consideration. Among firms piloting open-weight models, 62% named incomplete security and compliance reviews as the obstacle to production. The sample is small, so read direction, not census.
Adaptivity is a cadence, not a one-off decision. JPMorgan Chase, whose in-house abstraction layer Lesson 16 describes, runs a model-agnostic platform that uses at least OpenAI and Anthropic models and is refreshed every eight weeks (AI News, 16 December 2025). Copy the useful part: a fixed rhythm for re-evaluating your models, notice or no notice. A hospital that tested its generator once, at installation, would not call itself resilient.
The Workshop: Build Your Model Resilience Plan
The workshop has two parts: a half-day session (about 4 hours) to map and decide, and a 24-hour switch drill held within 30 days of it. The half day gives you a draft Model Resilience Plan. The drill gives you the one number that says whether the plan is real.
Who is in the room, and what to bring
- Chair: the CIO or CTO, because the outcome is an architecture and investment decision.
- Enterprise architect and AI platform lead: the request paths and the gateway.
- Procurement, legal and FinOps: the contracts, token spend by workload, and the dual-running budget.
- Business continuity lead: owns the drill, in the same calendar as every other continuity test.
- Operational-resilience or compliance lead, if DORA applies: owns the register of information and the exit plans.
- Owners of the two most critical workloads: they define "acceptable quality".
Bring the Assurance Evidence Matrix and requalification time (Lesson 24), the Agent Register (Lesson 22), your eval sets, model contracts, token spend by workload (Lesson 14) and each provider's deprecation calendar.
The nine steps
- Map dependencies (45 minutes). For each AI workload, record the model, the provider and the route: a direct API call, a cloud marketplace, or a model embedded in a SaaS product. The SaaS rows will embarrass you. Write "unknown" honestly; it becomes an action, not a guess.
- Tier the workloads (30 minutes). Judge by the impact of a week's outage or a forced downgrade. Tier 1: a regulated, customer-facing or revenue-critical process stops or breaks. Tier 2: people slow down and complain. Tier 3: a nice-to-have goes dark. Be strict: more than a handful in tier 1 means the room is avoiding a decision. If DORA applies, every workload supporting a critical or important function is tier 1.
- Name and qualify a fallback for each tier 1 workload (45 minutes, then a week of testing). Choose a substitute, ideally from a different provider and route. Run your own eval set against it and record the score next to the primary's; the gap is your quality delta. Then make it warm: credentials live, routing configured, capacity available.
- Put a gateway between applications and models (decide now, deliver over a quarter). Agree the single entry point every application will call. Move prompts, tool definitions and evals into provider-neutral, versioned form in your own repository (the practice from Lesson 9). A model change should be a routing change in one place, not fourteen releases.
- Rehearse the switch (a 24-hour drill within 30 days). Route real or replayed traffic for one tier 1 workload to the fallback through the rollback route from LLMOps in Practice, Lesson 11: a switch is a rollback to a different model. Raise the change ticket in advance. Time it from the decision until the fallback serves traffic at agreed quality, and record the quality drop and every manual step. VentureBeat's one-day test is a proposal, not a benchmark, so measure your own.
- Track deprecation calendars (15 minutes to assign). Give each provider's notices a named watcher and a monitored queue, not a personal inbox. Requalification starts the day a notice arrives, using the Lesson 24 pipeline.
- Renegotiate the contract terms that matter (30 minutes to list, a budget cycle to land). For each tier 1 vendor: notice before retirement or material change; continuity commitments; a transition period in which the old service keeps running while you migrate; change-of-control and assignment clauses; export of your data and fine-tuned models; notification of government action. Score each vendor on days of notice, and use only API keys your organization controls under a business contract.
- Decide where open-weight models fit (30 minutes). Use them where cost, control or sovereignty justify the hosting and security-review effort; the review is the usual bottleneck. Check each model's export-control and sanctions exposure as routine supplier due diligence.
- Plan capacity and cost, and set the cadence (30 minutes). Set a token budget per workflow and a spending cap per agent that stops a runaway loop rather than just reporting it, agree reserved throughput for tier 1, and track cost per successful outcome. Fix a re-evaluation cadence (eight weeks is a good default to argue from) and name a standing team that owns model changes.
Worked example: a European insurer runs the workshop
Illustrative only. Picture a 5,000-person European insurer, and so a DORA financial entity, with dozens of AI features. After step 2, only one workload is tier 1; the table adds one tier 2 and one tier 3 to show the pattern.
| Workload | Tier | Primary route | Fallback | Eval score, primary vs fallback | Switch time in drill | Decision |
|---|---|---|---|---|---|---|
| Claims triage agents | 1 | Proprietary model, direct API | Second proprietary model via cloud marketplace | 91% vs 86% | 31 hours (failed) | Move tool definitions behind the gateway; re-drill within 30 days |
| Policy-wording search | 2 | Proprietary model via gateway | Open-weight model, self-hosted | 88% vs 84% | Not drilled | Qualified on paper; drill yearly |
| Email summarizer in office suite | 3 | Embedded in SaaS, model unknown | None | No fallback | Not drilled | Ask vendor for provenance; accept the risk |
The drill result is the finding. The gateway handled the routing change in about an hour. The other 30 hours went on two things nobody had listed as risks: tool definitions written in one provider's format, and three prompts that depended on that model's output style. One downstream step read the summaries by position, and the fallback put the reasons first. Nothing errored. Cases simply went to the wrong queue until a claims handler noticed.
Three more findings. The largest provider takes 78% of model spend. The five-point quality gap is fine for a week but not a quarter, so the claims owner caps fallback running at 30 days before a human review. And because claims handling supports what the insurer classes as an important function, the compliance lead files the claims row as its DORA exit strategy, the drill log as test evidence, and the missing transition period as procurement's first ask.
The steering committee sets two targets: a tier 1 switch under 24 hours by the end of the following quarter, and no workload above tier 3 with unknown provenance. It buys no standby for the email summarizer: like the vending machine, it can go dark.
The Model Resilience Plan template
This is the artifact you leave with: one row per workload, tier 1 rows complete, tier 3 rows short. Store it next to the Agent Register and the Assurance Evidence Matrix, not in a slide deck.
| Field | What to record | Filled-in example (claims triage) |
|---|---|---|
| Workload, owner and tier | Name, accountable owner, tier with one-line rationale | Claims triage agents; Head of Claims Operations; tier 1, regulated customer outcome |
| Primary model, provider and route | Model version, provider, direct API, marketplace or embedded; for SaaS, the model inside per the vendor | Provider A, direct API, built in-house |
| Qualified fallback | Substitute model, route, date it passed your eval set, scores and quality delta | Provider B via cloud marketplace; passed nine days before the drill; 91% vs 86% |
| Portability status | Gateway in path? Prompts, tools and evals provider-neutral and versioned? | Gateway yes; tools provider-specific; three prompts model-dependent |
| Last drill | Date, switch time, quality drop, manual steps | Drill 1; 31 hours; 5 points; tool rewrite, prompt fixes |
| Contract terms | Notice days, continuity, transition period, change of control, export rights, government-action notification | 60 days; no continuity clause; no transition period; change of control silent; data export only |
| Regulatory exit duties | DORA in scope? Register of information entry, exit plan reference, date last tested | In scope; Provider A in register; drill 1 logged as exit-plan test |
| Open-weight decision | Used or not, why, hosting, sanctions and export-control check | Not for tier 1 in this plan cycle; security review pending |
| Capacity and cost | Token budget, agent spending cap, reserved throughput, cost per successful outcome | Budget and cap set; throughput requested; cost per triaged claim tracked |
| Cadence, watch and owner | Re-evaluation interval, deprecation watcher and alert route, standing team | Every 8 weeks; platform lead watches the AI platform queue; AI platform change board |
Fill it in during the half day, gaps and all. Every "unknown" cell is backlog.
Failure Modes, Metrics, and Closing the Loop
How do you spot model lock-in before the notice arrives?
Look for evidence, not intentions: each failure below leaves a trace in code, contracts or calendars long before the notice arrives.
| Failure mode | How to spot it | Fix |
|---|---|---|
| Hard-coded to one API | Application code imports a provider's SDK directly; no gateway in the request path | Gateway, tier 1 first |
| Redundancy for everything or nothing | Every workload has a paid standby, or none has | Re-tier; standby for tier 1 only |
| Untested fallback | A fallback is named, with no drill date or switch time recorded | Schedule the drill |
| Exit plan on paper only | A DORA exit strategy exists as a document, with no drill result and no qualified alternative | Log the switch drill as the exit-plan test |
| Hidden models inside SaaS | The vendor inventory leaves the model column blank | Ask for provenance at renewal |
| Sovereignty as a slogan | A board paper mentions sovereignty and no workload has a hosting or origin decision | Decide hosting and origin per tier 1 workload |
| Notice-day surprise | Notices reach an unmonitored mailbox; requalification starts weeks late | Named watcher (step 6) |
On sovereignty: BCG (29 May 2026) describes AI sovereignty in Europe as becoming a CEO-level question, with a "structurally fragmented picture" of what firms expect from it. A sovereign AI strategy becomes real only as rows: this workload, this model origin, this hosting choice, this reason.
What metrics show you are staying adaptive?
Six numbers, reviewed quarterly. None has a published industry target, so baseline them from the first drill.
| Metric | How to measure it |
|---|---|
| Time to switch, tier 1 | Hours from decision to fallback serving traffic at agreed quality, from the drill |
| Share of spend with one provider | Largest provider's spend divided by total model spend (the concentration the CSA tells you to assess) |
| Critical workloads with a tested fallback | Tier 1 workloads drilled inside 12 months, divided by all tier 1 |
| Days of notice in contracts | Shortest contractual notice across tier 1 vendors |
| Quality delta on fallback | Eval-score gap, primary against fallback |
| Cost per successful outcome | Model cost divided by completed cases, per workload |
One comparison ties them together. If time to switch plus requalification time (Lesson 24) is longer than your shortest notice, that workload is locked in, and the plan says so in plain words.
Closing the loop: re-walk the whole ladder
This is the last lesson, so the bridge points backwards. Resilience at Level 5 rests on every earlier level; re-run the earlier artifacts with Level 5 eyes to find the weak rung.
- Level 1, AI Aware. Your dependency map only covers models you know about. Go back to Lesson 1 and re-run the shadow AI discovery, then the data inventory (Lesson 2). "Model unknown" SaaS rows are Level 1 gaps showing up at Level 5.
- Level 2, AI Capable. Lesson 10 taught you to set a baseline before any pilot; without one, neither a redesign nor a model switch can be judged. Check every workload you might switch still has one.
- Level 3, AI Operational. Your model inventory (Lesson 15) must now hold agents (Lesson 22) and assurance tiers (Lesson 24), and your rollback runbook (Lesson 11) should be the route your switch drill uses.
- Level 4, AI Scaled. Lesson 19's ROI measures should include cost per successful outcome, and Lesson 20's operating model should name who owns model changes.
- Level 5, AI Native. Review the five artifacts together each quarter: Process Redesign Charter, Agent Register, Feedback Loop Canvas and eval set, Assurance Evidence Matrix, Model Resilience Plan.
Two honest framings for that reassessment. First, maturity is not readiness: maturity is the capability you can evidence today; readiness asks whether you can execute the next move now, such as a forced model change. Second, score each dimension separately and let the weakest one cap your level, instead of averaging to a flattering number. A firm with superb evals and no tested fallback is not Level 5 on resilience.
Picture where this leaves you. When a better or cheaper model appears, you qualify it and route a workload to it inside one eight-week cycle, while a rival spends a quarter finding out what would break. When a provider retires, restricts or reprices a model, you lose hours of routing work, not a quarter, and you negotiate with a tested alternative on the table. That is AI Native in practice: the model is a component you choose each cycle, not a foundation you are stuck on.
Then book a rerun of the readiness assessment from Lesson 5. The ladder is climbed once and re-walked every quarter, because the models under it will have changed.
Frequently Asked Questions
Is using several AI vendors enough to avoid lock-in?
No. Lock-in is time against notice, not vendor count: three providers with no tested way to move a workload leave you more exposed than one with a rehearsed two-day switch. What protects you is a gateway, your own eval set and a drill that proves the switch time.
How do we switch LLM providers without breaking production?
Route every application through a gateway, requalify the new model on your own eval set, and rehearse the cutover on one tier 1 workload before you need it. Budget for running both models during the move, and check downstream steps that read the old model's output format; they fail quietly.
What should an AI vendor contract say about model retirement?
It should fix a minimum notice period before retirement or material change, plus a transition period in which the old service keeps running while you migrate. Add continuity, change-of-control, data and fine-tune export, and government-action notification terms.
Should we move to open-weight models to avoid lock-in?
Use them where cost, control or sovereignty justify the hosting and security-review effort, not as a wholesale switch. A September 2026 ETR survey of 200 firms found open-weight models carry 34% of enterprise AI tokens while proprietary usage still grows, so most firms mix both.
Does DORA require an exit strategy for our AI model provider?
If you are an EU financial entity and the model supports a critical or important function, assume yes. Since 17 January 2025, DORA Article 28(8) has required documented, periodically tested exit strategies with identified alternatives for such ICT services; confirm scope with your compliance team.
Key Takeaways
- AI model lock-in is time against notice: if switching and requalifying a workload takes longer than the notice you will get, you are locked in, however many vendors you use.
- Model access changes for reasons unrelated to you: routine retirements (about every two months in Anthropic's log), plus cutoffs driven by competitors, governments and export rules.
- Build portability in three layers: technical (gateway, provider-neutral prompts, tools and evals), contractual (notice, transition period, continuity, change of control) and organizational (drill, owner, dual-running budget).
- Be proportionate, like a hospital generator: tier strictly and keep a tested warm standby only for tier 1.
- A fallback is real only once drilled; a failed drill is the quarter's most useful finding, and for EU financial entities it is also DORA exit-plan evidence.
- Multi-model is additive: the right model per workload, open-weight where it justifies the review effort, re-evaluated on a fixed cadence.
- Review the Model Resilience Plan with the other four Level 5 artifacts every quarter, and re-walk the ladder from Level 1, letting your weakest dimension set your level.
Skill.re