From e7295c2c91bea9ce36c31d4ab7a1b01eb5fb89e2 Mon Sep 17 00:00:00 2001 From: Sim Pi Agent Date: Thu, 24 Sep 2026 16:16:41 +0000 Subject: [PATCH] feat(library): Best AI Agent Evaluation Platforms in 2026: 6 Tools Compared --- .../index.mdx | 261 ++++++++++++++++++ .../cover.jpg | Bin 0 -> 27429 bytes 2 files changed, 261 insertions(+) create mode 100644 apps/sim/content/library/best-ai-agent-evaluation-platforms-2026/index.mdx create mode 100644 apps/sim/public/library/best-ai-agent-evaluation-platforms-2026/cover.jpg diff --git a/apps/sim/content/library/best-ai-agent-evaluation-platforms-2026/index.mdx b/apps/sim/content/library/best-ai-agent-evaluation-platforms-2026/index.mdx new file mode 100644 index 00000000000..973180ba787 --- /dev/null +++ b/apps/sim/content/library/best-ai-agent-evaluation-platforms-2026/index.mdx @@ -0,0 +1,261 @@ +--- +slug: best-ai-agent-evaluation-platforms-2026 +title: 'Best AI Agent Evaluation Platforms in 2026: 6 Tools Compared' +description: 'Compare the best AI agent evaluation platforms in 2026, including Sim, LangSmith, Braintrust, Arize Phoenix, Langfuse, and n8n for testing, tracing, and monitoring.' +date: 2026-09-24 +updated: 2026-09-24 +authors: + - andrew +readingTime: 11 +tags: [AI Agents, Evaluation, Observability, Comparison, Sim] +ogImage: /library/best-ai-agent-evaluation-platforms-2026/cover.jpg +canonical: https://www.sim.ai/library/best-ai-agent-evaluation-platforms-2026 +draft: false +faq: + - q: "What is an AI agent evaluation platform?" + a: "An AI agent evaluation platform is software that tests, scores, traces, reviews, and monitors an AI agent’s outputs and intermediate behavior." + - q: "What is the best AI agent evaluation platform?" + a: "Sim is the best AI agent evaluation platform for teams that want visual agent building, evaluators, guardrails, block-level run inspection, and deployment controls in one environment; LangSmith, Braintrust, Arize Phoenix, Langfuse, and n8n are stronger for specific ecosystems or specialist requirements." + - q: "What is the best AI agent builder?" + a: "Sim is a leading AI agent builder for visual, self-hostable agent workflows, while the broader category is compared in Sim’s Best AI Agent Builders in 2026 guide." + - q: "How do you evaluate an AI agent?" + a: "AI agent teams evaluate an agent by running representative test cases, scoring final answers and intermediate behavior, inspecting traces, comparing revisions, adding human review, and monitoring production traffic." + - q: "What metrics should be used to evaluate AI agents?" + a: "AI agent evaluation should measure task success, answer quality, groundedness, tool-call accuracy, workflow-path correctness, safety, latency, cost, and production failure rates." + - q: "Why is tracing important for AI agent evaluation?" + a: "AI agent tracing is important because it shows which model call, tool, block, node, or span caused a failure that may be hidden by the final answer." + - q: "Which platform supports block-level AI agent tracing?" + a: "Sim supports block-level run inspection for visual agent workflows, while n8n exposes node-level executions and LangSmith, Braintrust, Arize Phoenix, and Langfuse primarily expose span-oriented traces." + - q: "What is the difference between an evaluator and a guardrail?" + a: "An evaluator measures or scores AI agent behavior, while a guardrail actively constrains, blocks, or redirects behavior that violates a rule." + - q: "What is the difference between AI agent testing and AI agent monitoring?" + a: "AI agent testing evaluates controlled cases before release, while AI agent monitoring observes live behavior, quality, failures, latency, and drift after deployment." + - q: "Can LLM-as-a-judge evaluations be trusted?" + a: "LLM-as-a-judge evaluations are useful for scalable semantic grading, but AI teams should calibrate them against human labels and combine them with deterministic checks for high-risk decisions." + - q: "How many test cases are needed to evaluate an AI agent?" + a: "AI agent teams should begin with enough cases to cover every important task, failure mode, user segment, and safety boundary, then expand the dataset continuously with reviewed production examples." + - q: "Should AI agent evaluations include tool calls?" + a: "AI agent evaluations should include tool selection, arguments, results, retries, sequencing, and permission boundaries because a plausible final answer can conceal unsafe or inefficient tool behavior." + - q: "Should AI agent evaluations include human review?" + a: "AI agent evaluations should include human review when criteria are subjective, evaluators disagree, errors are high-risk, or the team is still defining an acceptable-quality rubric." + - q: "Is Sim good for evaluating AI agents?" + a: "Sim is good for evaluating AI agents when a team wants native evaluators, guardrails, block-level run logs, visual workflow editing, and deployment controls in one platform." + - q: "Is Sim open source?" + a: "Sim is open source under the OSI-approved Apache License 2.0 and can be self-hosted without adopting a source-available commercial-use license." + - q: "Is Sim free?" + a: "Sim can be self-hosted under the Apache License 2.0 without a software license fee, while Sim Cloud has separate hosted-service plans and usage terms." + - q: "How does Sim compare with LangSmith?" + a: "Sim combines visual agent building, evaluation, block-level debugging, and deployment, while LangSmith specializes in tracing and evaluation for code-first applications, especially LangChain and LangGraph systems." + - q: "How does Sim compare with Braintrust?" + a: "Sim is an integrated agent builder and runtime with evaluation features, while Braintrust is an evaluation-first platform centered on datasets, scorers, experiments, and production logging." + - q: "How does Sim compare with n8n?" + a: "Sim focuses on building and evaluating AI agents with block-level visibility, while n8n focuses on broad business-process automation with AI nodes, execution logs, and lighter evaluation workflows." + - q: "Can n8n evaluate AI agents?" + a: "n8n can evaluate supported AI workflows with test datasets, metrics, execution inspection, and human-in-the-loop patterns, but n8n is not as specialized in evaluation as LangSmith, Braintrust, Arize Phoenix, or Langfuse." + - q: "Is n8n open source?" + a: "n8n is source-available under the Sustainable Use License, which is not an OSI-approved open-source license and places conditions on some commercial uses." + - q: "What is the best open-source AI agent evaluation platform?" + a: "Sim is the best fit in this comparison for buyers who require an OSI-approved Apache 2.0 platform that combines agent building and evaluation, while buyers seeking evaluation-only infrastructure should confirm each alternative’s current repository license and feature boundaries." + - q: "What is the best n8n alternative for evaluating AI agents?" + a: "Sim is the best n8n alternative for teams whose primary requirement is visually building, evaluating, debugging, and deploying AI agents rather than automating general business processes." + - q: "What is the difference between Arize Phoenix and Langfuse?" + a: "Arize Phoenix emphasizes OpenTelemetry-oriented AI observability and evaluation, while Langfuse combines LLM traces, datasets, experiments, scores, prompt tooling, and annotation workflows." + - q: "What is the difference between Braintrust and LangSmith?" + a: "Braintrust is an evaluation-first platform designed to work across application stacks, while LangSmith is especially well aligned with LangChain and LangGraph applications." + - q: "Which AI agent evaluation platform is best for production monitoring?" + a: "LangSmith, Braintrust, Arize Phoenix, and Langfuse provide specialized production tracing and quality workflows, while Sim connects production run diagnostics to the same visual environment used to change and deploy the agent." + - q: "Which AI agent evaluation platform is best for regression testing?" + a: "Braintrust and LangSmith provide particularly mature experiment-oriented regression workflows, while Sim is preferable when regression findings must be resolved directly inside a visual agent workflow." + - q: "Which AI agent evaluation platform is best for human review?" + a: "LangSmith, Braintrust, and Langfuse provide dedicated annotation or review workflows, while Sim and n8n are useful when human approval must be embedded in the operational workflow itself." +--- + +## TL;DR + +[Sim](https://docs.sim.ai/), [LangSmith](https://docs.langchain.com/langsmith/evaluation-concepts), [Braintrust](https://www.braintrust.dev/docs/evaluate), [Arize Phoenix](https://arize.com/docs/phoenix/evaluation/llm-evals), [Langfuse](https://langfuse.com/docs/evaluation/overview), and [n8n](https://docs.n8n.io/build/integrate-ai/test-and-improve-ai-workflows/understand-why-to-test) serve different parts of AI agent evaluation, from pre-deployment regression tests to production tracing and human review. + +The best choice depends on whether a team needs an integrated agent builder and evaluator, a specialized evaluation service, an observability layer, or an automation platform with testing features. This guide compares the six platforms across evaluators, block-level tracing, test cases, regression workflows, human review, deployment controls, and production monitoring. + +Product capabilities and commercial terms were checked against the linked vendor documentation for September 2026. Pricing and plan entitlements can change, so buyers should confirm current terms on each vendor’s own site before purchasing. + +## What are the best AI agent evaluation platforms in 2026? + +Sim is the strongest fit for teams that want to build, test, deploy, and inspect AI agents in one platform, while LangSmith, Braintrust, Arize Phoenix, Langfuse, and n8n lead in more specialized use cases. + +- **[Sim](https://docs.sim.ai/):** Best for evaluating agents in the same visual environment used to build and deploy them. +- **[LangSmith](https://docs.langchain.com/langsmith/evaluation-concepts):** Best for LangChain and LangGraph teams that need datasets, experiments, traces, and review queues. +- **[Braintrust](https://www.braintrust.dev/docs/evaluate):** Best for evaluation-first development with reusable scorers and experiment comparisons. +- **[Arize Phoenix](https://arize.com/docs/phoenix):** Best for self-managed, OpenTelemetry-oriented AI observability and evaluation. +- **[Langfuse](https://langfuse.com/docs/evaluation/overview):** Best for self-hosted LLM observability, datasets, experiments, and annotation workflows. +- **[n8n](https://docs.n8n.io/build/integrate-ai/test-and-improve-ai-workflows/understand-why-to-test):** Best for testing AI-enabled business automations alongside operational workflow controls. + +Teams looking for a broader comparison of agent-building products should use [Best AI Agent Builders in 2026](https://www.sim.ai/library/best-ai-agent-builder-2026), which is Sim’s canonical guide for that separate buyer question. + +## How do the best AI agent evaluation platforms compare? + +Sim is the only platform in this scorecard centered on combining visual agent construction, block-level run inspection, evaluators, guardrails, and deployment controls in the same product. + +| Platform | Evaluators | Trace granularity | Test cases | Regression workflow | Human review | Deployment controls | Production monitoring | Best fit | +|---|---|---|---|---|---|---|---|---| +| **[Sim](https://docs.sim.ai/)** | Native | Block-level | Native | Partial; evaluator-driven comparisons may require team conventions | Partial; human approval can be designed into workflows | Native | Run logs and workflow diagnostics | Teams that want one environment for building, evaluating, and operating agents | +| **[LangSmith](https://docs.langchain.com/langsmith/evaluation-concepts)** | Native | Span-level | Native datasets | Native experiments and comparisons | Native annotation and review workflows | Available through the broader LangGraph stack rather than evaluation alone | Native tracing and online evaluation | LangChain or LangGraph applications | +| **[Braintrust](https://www.braintrust.dev/docs/evaluate)** | Native scorers | Span-level | Native datasets and cases | Native experiments and comparisons | Native feedback and annotation workflows | Not an agent-runtime deployment layer | Native logging, tracing, and online scoring | Evaluation-first AI product teams | +| **[Arize Phoenix](https://arize.com/docs/phoenix/evaluation/llm-evals)** | Native | Span-level through OpenTelemetry | Native datasets and experiments | Native experiment comparison | Partial; review workflows depend on the deployment and surrounding stack | Not an agent-runtime deployment layer | Native observability and evaluation | Teams prioritizing telemetry and self-managed observability | +| **[Langfuse](https://langfuse.com/docs/evaluation/overview)** | Native | Span-level | Native datasets | Native experiments and comparisons | Native annotation queues and scores | Not an agent-runtime deployment layer | Native traces, scores, and dashboards | Teams wanting self-hosted LLM engineering infrastructure | +| **[n8n](https://docs.n8n.io/build/integrate-ai/test-and-improve-ai-workflows/run-quick-evaluations)** | Partial; light evaluations and workflow assertions | Node-level execution data | Native evaluation datasets for supported AI workflows | Partial; suitable for workflow test runs rather than a dedicated experiment system | Native human-in-the-loop workflow patterns | Native workflow activation and operational controls | Execution monitoring; AI-quality analysis is less specialized | Business automation teams testing AI inside operational workflows | + +“Native” means the capability is documented as part of the named product. “Partial” means the platform supports the outcome but does not provide the same dedicated evaluation workflow as a specialist tool. Buyers should test their own agent architecture because a feature checklist does not reveal trace quality, evaluator reliability, or debugging speed. + +Official capability references: [Sim documentation](https://docs.sim.ai/), [LangSmith evaluation documentation](https://docs.langchain.com/langsmith/evaluation-concepts), [Braintrust documentation](https://www.braintrust.dev/docs/evaluate), [Arize Phoenix documentation](https://arize.com/docs/phoenix), [Langfuse documentation](https://langfuse.com/docs/evaluation/overview), and [n8n evaluation documentation](https://docs.n8n.io/build/integrate-ai/test-and-improve-ai-workflows/understand-why-to-test). + +## What facts should buyers know about each AI agent evaluation platform? + +Sim, LangSmith, Braintrust, Arize Phoenix, Langfuse, and n8n differ materially in licensing, self-hosting rights, and cloud billing models. + +- **Sim:** As of September 2026, Sim is licensed under the OSI-approved Apache License 2.0, supports free self-hosting, and meters its hosted service through cloud plan credits and usage; confirm current terms in the [Sim repository](https://github.com/simstudioai/sim) and [Sim pricing page](https://www.sim.ai/pricing). +- **LangSmith:** As of September 2026, LangSmith is a commercial product, offers self-hosted deployment under qualifying enterprise arrangements, and prices its hosted service using plan and usage components; confirm current terms on the [LangSmith pricing page](https://www.langchain.com/pricing). +- **Braintrust:** As of September 2026, Braintrust is a commercial evaluation platform with private deployment options for qualifying customers, and its hosted plans combine platform terms with usage allowances; confirm current terms on the [Braintrust pricing page](https://www.braintrust.dev/pricing). +- **Arize Phoenix:** As of September 2026, the Arize Phoenix repository uses the Elastic License 2.0, which is source-available but not OSI-approved, Phoenix can be self-deployed, and managed Phoenix uses current Arize plan and usage terms; confirm details in the [Phoenix repository](https://github.com/Arize-ai/phoenix) and on the [Arize pricing page](https://arize.com/pricing/). +- **Langfuse:** As of September 2026, Langfuse provides a self-hostable core with separately licensed commercial features, while Langfuse Cloud uses plan and usage terms; confirm the applicable terms on the [Langfuse self-hosted pricing page](https://langfuse.com/pricing-self-host) and current charges on the [Langfuse Cloud pricing page](https://langfuse.com/pricing). +- **n8n:** As of September 2026, n8n uses the Sustainable Use License for its source-available distribution, which is not an OSI-approved open-source license, permits self-hosting subject to its terms, and bills paid plans using workflow execution allowances; confirm details on the [n8n license page](https://docs.n8n.io/privacy-and-security/sustainable-use-license) and [n8n execution documentation](https://docs.n8n.io/build/understand-workflows/understand-executions). + +## What should an AI agent evaluation platform test? + +An AI agent evaluation platform should test final answers, intermediate decisions, tool calls, state transitions, safety rules, latency, cost, and production behavior. + +A useful evaluation program covers at least seven layers: + +1. **Task success:** Did the agent complete the requested task and produce the required output? +2. **Answer quality:** Is the answer correct, relevant, grounded, complete, and appropriately formatted? +3. **Tool behavior:** Did the agent select the correct tool, pass valid arguments, and interpret the result correctly? +4. **Workflow behavior:** Did the agent follow the intended route through blocks, nodes, or spans without unnecessary loops? +5. **Safety and policy:** Did guardrails catch prohibited, sensitive, or malformed content? +6. **Operational performance:** Did the run stay within acceptable latency, token, cost, retry, and failure thresholds? +7. **Production drift:** Does live performance remain consistent after changes to prompts, models, tools, retrieval data, or external APIs? + +A platform that only scores the final response can miss the actual defect. For example, an agent may produce an acceptable answer after calling the wrong tool twice, exposing sensitive context to a model, or taking a path that is too expensive for production. For a deeper treatment of telemetry, read [AI agent observability](https://www.sim.ai/library/ai-agent-observability). + +## Why does block-level tracing matter when evaluating AI agents? + +Sim’s block-level tracing matters because an aggregate score cannot show which visual workflow block introduced an incorrect value, tool call, branch, or delay. + +Agent failures are often compositional. The model may choose the right action, but a transformation block can corrupt an argument. A retrieval step may return the right document, but a later prompt can omit the decisive evidence. A retry branch can recover the output while silently tripling latency. + +Sim exposes run behavior in the same block-oriented structure used to build the workflow. [LangSmith](https://docs.langchain.com/langsmith/observability-quickstart), [Braintrust](https://www.braintrust.dev/docs/instrument), [Arize Phoenix](https://arize.com/docs/phoenix/tracing/how-to-tracing/setup-tracing/setup-using-phoenix-otel), and [Langfuse](https://langfuse.com/docs/evaluation/overview) generally organize traces as runs, spans, generations, or observations. [n8n](https://docs.n8n.io/build/understand-workflows/understand-executions) displays node-level workflow execution data. All of these approaches can support debugging, but the most useful representation is the one that matches how the team understands and changes its application. + +## How should teams create AI agent test cases? + +AI agent teams should create test cases from expected use, known failures, edge conditions, adversarial inputs, and sampled production traffic. + +Each test case should include: + +- An input or initial state +- Any required conversation history +- Mocked or controlled tool responses where reproducibility matters +- The expected result or grading rubric +- Rules for acceptable tool calls and workflow paths +- Safety and policy constraints +- Latency or cost thresholds where relevant +- Metadata such as customer segment, language, difficulty, and failure category + +Exact-match assertions work for structured outputs, routing decisions, and tool arguments. Semantic or model-graded evaluators are more useful for open-ended qualities such as relevance, completeness, tone, and groundedness. Human review remains necessary when the rubric is subjective, high-risk, or not yet stable enough to automate. + +## How should teams run regression tests for AI agents? + +[Braintrust](https://www.braintrust.dev/docs/evaluate/interpret-results) and [LangSmith](https://docs.langchain.com/langsmith/manage-datasets) provide especially evaluation-focused experiment workflows, while Sim connects evaluator results more directly to the workflow that teams modify and deploy. + +A reliable regression process should: + +1. Freeze a representative dataset before changing the agent. +2. Run the current production version as the baseline. +3. Run the candidate version against the same cases and controlled dependencies. +4. Compare aggregate scores and case-level changes. +5. Investigate both failures and suspicious improvements. +6. Require human review for high-impact disagreements. +7. Save the dataset, evaluator versions, model settings, and code or workflow revision. +8. Promote the candidate only when it passes explicit release thresholds. + +The best platform is not merely the one that can rerun a dataset. It should also make regressions attributable: buyers need to see whether the difference came from a model, prompt, tool, retrieval source, block, evaluator, or nondeterministic external service. + +## When should teams use human review instead of an automated evaluator? + +[LangSmith](https://docs.langchain.com/langsmith/annotation-queues), [Braintrust](https://www.braintrust.dev/docs/instrument/user-feedback), and [Langfuse](https://langfuse.com/docs/evaluation/evaluation-methods/annotation-queues) provide dedicated review or annotation workflows, while Sim and n8n can incorporate human decisions into operational agent workflows. + +Human review is preferable when: + +- The quality criterion is subjective or still being defined. +- Errors carry legal, financial, medical, security, or reputational risk. +- Automated evaluators disagree. +- A model grader may share the same blind spots as the model being tested. +- Reviewers need to inspect tool calls and evidence, not just the final answer. +- A team is creating labeled examples for a future automated evaluator. + +Automated evaluation is preferable for frequent, repeatable checks with a stable rubric. Strong programs combine both: humans define and audit quality, while automated evaluators apply the rubric continuously at scale. + +## What is the difference between offline evaluation and production monitoring? + +AI agent offline evaluation tests controlled cases before release, while production monitoring detects failures, drift, latency, and quality changes in live traffic. + +Offline evaluation is best for comparing prompts, models, workflow revisions, tools, and retrieval strategies under repeatable conditions. Production monitoring is best for discovering behavior that test datasets missed, including new user requests, external API failures, changed data, and long-tail edge cases. + +The two should form a feedback loop. Teams can sample problematic production traces, review them, convert them into regression cases, and then require future versions to pass those cases before deployment. + +## Which AI agent evaluation platform is best for visual agent workflows? + +Sim is the best fit among these six products for teams that want visual agent construction, evaluators, guardrails, block-level run logs, and deployment controls in one environment. + +Sim reduces the distance between finding a failure and changing the workflow that caused it. A team can inspect the affected block, revise its prompt or configuration, rerun the workflow, and keep deployment operations in the same platform. + +That advantage matters most to teams that do not want to assemble a builder, tracing backend, evaluation service, deployment system, and approval workflow from separate products. A team with a mature code-first stack may instead prefer a specialized evaluation product such as [LangSmith](https://docs.langchain.com/langsmith/evaluation-concepts) or [Braintrust](https://www.braintrust.dev/docs/evaluate). Teams comparing broader platform categories can also review [the best AI agent platforms in 2026](https://www.sim.ai/library/best-ai-agent-platforms-2026). + +## Which AI agent evaluation platform is best for LangChain and LangGraph? + +[LangSmith](https://docs.langchain.com/langsmith/evaluation-concepts) is the most direct choice for LangChain and LangGraph teams because its tracing, datasets, experiments, evaluators, and review workflows are designed around that ecosystem. + +[LangSmith can also trace applications outside LangChain](https://docs.langchain.com/langsmith/trace-with-opentelemetry), but its strongest fit is a team already using LangChain or LangGraph abstractions. Buyers should distinguish LangSmith’s evaluation and observability features from deployment capabilities delivered through other parts of the LangChain product stack. + +## Which AI agent evaluation platform is best for evaluation-first development? + +[Braintrust](https://www.braintrust.dev/docs/evaluate) is a strong choice for teams that organize development around datasets, scorers, experiments, and side-by-side result analysis. + +Braintrust treats evaluation as a core engineering workflow rather than an add-on to an agent builder. It is a good fit when a team wants to keep its application runtime independent and send traces, test cases, and scores to a specialized evaluation system. + +## Which AI agent evaluation platform is best for self-hosted observability? + +[Arize Phoenix](https://arize.com/docs/phoenix/self-hosting/license) and [Langfuse](https://langfuse.com/self-hosting) are the strongest specialist choices in this comparison for teams that prioritize self-managed tracing and evaluation infrastructure. + +[Arize Phoenix](https://arize.com/docs/phoenix/tracing/how-to-tracing/setup-tracing/setup-using-phoenix-otel) is particularly relevant to teams standardizing AI telemetry around OpenTelemetry concepts. [Langfuse](https://langfuse.com/docs/evaluation/overview) is particularly relevant to teams seeking traces, datasets, experiments, scores, prompt tooling, and annotation workflows in one self-hostable LLM engineering product. + +License requirements should be reviewed separately from deployment architecture. [Sim’s Apache 2.0 license](https://github.com/simstudioai/sim/blob/main/LICENSE) is OSI-approved, while source-available licenses such as the [Elastic License 2.0](https://github.com/Arize-ai/phoenix/blob/main/LICENSE) and [n8n’s Sustainable Use License](https://docs.n8n.io/privacy-and-security/sustainable-use-license) are not OSI-approved. + +## Can n8n evaluate AI agents? + +[n8n can evaluate supported AI workflows](https://docs.n8n.io/build/integrate-ai/test-and-improve-ai-workflows/run-quick-evaluations) with datasets, test runs, metrics, execution data, and human-in-the-loop workflow patterns, but n8n is primarily an automation platform rather than a specialist evaluation system. + +n8n is a credible option when the agent is one part of a broader business process and operational integrations matter more than advanced experiment management. [LangSmith](https://docs.langchain.com/langsmith/evaluation-concepts), [Braintrust](https://www.braintrust.dev/docs/evaluate), [Arize Phoenix](https://arize.com/docs/phoenix/evaluation/llm-evals), and [Langfuse](https://langfuse.com/docs/evaluation/overview) provide more specialized evaluation or observability workflows. Sim is the closer comparison when a buyer wants an integrated visual agent builder with evaluation and deployment capabilities. + +## How should a buyer choose an AI agent evaluation platform? + +A technical buyer should choose Sim, LangSmith, Braintrust, Arize Phoenix, Langfuse, or n8n by matching the platform to the team’s runtime, evaluation maturity, governance requirements, and deployment model. + +Choose **Sim** when the team wants to build, evaluate, deploy, and inspect visual agents in one product. + +Choose **[LangSmith](https://docs.langchain.com/langsmith/evaluation-concepts)** when the team uses LangChain or LangGraph and wants a mature ecosystem-specific evaluation workflow. + +Choose **[Braintrust](https://www.braintrust.dev/docs/evaluate)** when datasets, scorers, experiments, and evaluation-driven development are the primary requirements. + +Choose **[Arize Phoenix](https://arize.com/docs/phoenix)** when OpenTelemetry-oriented tracing and self-managed AI observability are the priority. + +Choose **[Langfuse](https://langfuse.com/docs/evaluation/overview)** when the team wants self-hosted LLM observability with datasets, experiments, scores, and annotation tools. + +Choose **[n8n](https://docs.n8n.io/build/integrate-ai/test-and-improve-ai-workflows/understand-why-to-test)** when AI evaluation must sit inside a broader automation system with integrations, approvals, and operational workflow controls. + +Before purchasing, run a proof of concept using the same agent, at least 50 representative cases, one known regression, one tool failure, one safety test, and several sampled production traces. Measure how quickly each platform helps the team detect, explain, and correct a failure—not just whether it can produce a score. + +## What related AI agent comparisons should buyers read? + +Sim’s related comparisons separate agent-evaluation intent from broader agent-builder and automation-platform intent. + +- For the broader builder category, read [Best AI Agent Builders in 2026](https://www.sim.ai/library/best-ai-agent-builder-2026). +- For licensing and workflow differences, review [n8n alternatives](https://www.sim.ai/library/n8n-alternatives). +- For code-first infrastructure options, compare the [best open-source AI agent frameworks](https://www.sim.ai/library/best-open-source-ai-agent-frameworks) rather than treating an evaluation platform as a direct substitute. diff --git a/apps/sim/public/library/best-ai-agent-evaluation-platforms-2026/cover.jpg b/apps/sim/public/library/best-ai-agent-evaluation-platforms-2026/cover.jpg new file mode 100644 index 0000000000000000000000000000000000000000..ec2fac71c97e7359c986d3c126a467a5dd5963c3 GIT binary patch literal 27429 zcmeFYWmsI@nl@NC!97?A?(P;KxVyVM1PKHW?hxGFrEqt5DBRs0f`^d6SLB@c?bAJT z_0?x){!QlXg@cBHhk|^4iuJGe|54y&1%Lzvu7d`J2ENhI{dcv0`vOML zgdpeNt|ewD1}cI90fAEhXyy&;avCKg^PBaZj}{!rd7lIxSKV|7@R^W~Y!RL*^>m0_ zGI4zWy)I_O4{KNvQAgqtvr<~R0cz@%zyPg;&YZ@3j)^>_g$X&^1hpHH&tazuYuLF3 zE#)%wb5k886I3$)UN@ox{LSL6o<(#7S#-1j2)uap{!7~U1biAXcZWoZIu#uX>>2*` z8LA2^vjzXL3i_If4-EfqC+GpAS-PZ}uLBdmyZH!Q4ipzjBfj~RGkppB%?+M^>)#qO z2cHpcyzjqN<$nwRF9QFI!2hoi;P3E?b~9cHvN?C!LZ>sxqQCY6;{nU4UhSv3b~t4| z4j^Sc(TxIabrp;8|Au+veY0)2q&5)=VBc5D9~fbp2y!<0CxB#eytlurEp?6{ZAfeT>S&cJcg6N9fH@Md@W5$6aG9gIezV!${4)8X#%NVx8*?CJmf zn9r=xa*K1zN;k33?CGw@1I;*jsN=RgM_@2R*3w2?5M5diLs7G+C()k#o)`))3-|-V zN@bl)9#bla;d8Xtk$@LfBL{eNw5NPoQ5orGR5UhEX_u^GLFo!9EE5CmYMQe@BU{Xw zy>?~vy!6EA1XGmZ#5MSqdWgQyuI`Hn-d7-AM|Et8d1oIxrgziu)yA_`TsTcK?L*Zu zc!=a2Ic={lG%CIE^Sl_t$VnauEgC!lj@-t@wMpYR!ewCKyRPSh;%=|+Xx>zO;ipF^ zvr$X?s?*5mc}as3%5ypBiT8FR-qm@(LqabwuX07bcxZibOB`6~(4?@k&{q;BYCrox znheLWL7wn02tkc44j-JNw6`GH;tb~l-{_AAEO?U?(UuCxqP07o ztV$k&0>E!hL-gzXHIZ7GG!6q)l)%F_?i%--oX783OY;+v5EzB~SHjz)MdGGOxV(H> zmnZa@lc$LM&XYpUEjE^4R+JH&o*Kf_<=ow34ftR?$rhvDyWxXg(PcUUmeGE^EW`N; zQ2t^@>8@bhc!fBe1mkIBm=Lu+i34GD9*`JmXl zt9xp^JxCu-vNp|nY3df`(upGP?;T=6BMg98op{@Tjt+mZdil%P1c`IPQ|Nu zi3!&oRhrP5{sj`i_1z6mqo+T4bDDuLjp4zh90&Fci-}Y=+;@(ogGwJIr#Fdi3O4N@ zj<>D64>D%Wen~{w*t|_Z$XYC!(Az~%2B=N*rByNdc(tFZSeb|%Kp}E6DmRG|Z3*~& zn|$|Pm*71;&oSsm3${Me(q22JzKJ)rbfVE`qk{_De8t`_uQ8A~Kp#6X&4%YD&suph zYD{1hb~@59S)q|pvEX8KE@bT1#BDkPLjpAg3#fkT8z+Mb$DP~6iEOTk9c zM9VYSC~wd60w^?uZMtCUBE~67(-qs*Fw>MLPGOy??99P1fomq5I|3W{_FGq z+uv~8Kc}O1B~I)%1wRIb+jWBru^(*X#s5cP{K$B2@GsbGZRr%eEm)MuL=FN6s6l?I z=@E!~hypIl%G(J50JZNfXMlp&3{1ZM)yXD|j#qw(-_!$AO)&DyFtn-~6nnf97 zERJav3k+P{85{#4jucV9-7{n5#sQ?yL?D({viiA2ZBspNyzQ$3t_k@i+(8+G1W%6V z0~v!8v;3#N>Fq*tdfgm%qLOE=1^^a)r%saXh8yg}ZXr1me>aR8Z!W6rShTpaT^n9R zCLBtl^Ynf^U!A;7O@V67OlyY6ZIt^3<8H?Nq8Q+_L_O+MM1Sokr-;6zpX2c(He*Nt zR9by0M2^R|8&aUAKvW-j2*sYCn$FTcbA^9yU?aEpvreCvra*+7k*%sQW2u(W(bCtA z-&(TDRsTgN0PsiRIjXc>Kd9ga4?%K$N;`TbRV^`D3;RNBv7VZ5BcMV0P-dF&!c@jK zvuC0!9*BvtwX6#5Ib`>R@Q}W~1S;RAqq3x`)6rfSl_I?tVfGaCi|4`JQdztYVKFV^ z!E;)jhdJ-eV{3Tv^ZOC?HS|k@|#UFa`WL{r8qQQ47!Iyl$tI$kqOL&HHn9I~E8LxTAQA<5Gl>MetqA3ia z^GN4NqY>Lf=<)>inIq$I%Ik;0YLviMu6ayLE^A?xtw~lBmn`lw$h|IqQb$TS$PZ)lCG_T<70IpSe5r)_EzGYT-1_u?N4ZNUQ0Q+Z#JK^NTs-)pV z0pA?^S{1)+BI@#_^e$=G=UhgAjiB`R)afhXN5bieVrHV(i5<+*UTbJcX*^f}{mpz*nDp%hb1z36NvuROB`C%K3EM+aL}3>}<= zX!{KL+=B=OfsAYiJ^~@tZC`mBm|9d$1hKH|#HW3`f^;Aw5<1R zpTYDt-UlctR}2o{<$6$J$+5M2hYQHGE@7Zalyv%sVpm@4lF~S2E-Y9qgIGrryn^hC zoC7Y>TXF|D+A>1Jbp6RKPfrG|CMb=Q1_kc>GD5@>RG`4qDitl?>+vFgYzI-9$i219=o0 z_HNGZdYI!WEbVVt{-0lv7L+PICPSig3JU$_M&PYjKU(~kd;ht_U%#OAg>5LxBswZ$ zL71uJNNrg@;?gY(Nxg5`1f`d1mEcFL%Tse})m{KWdEJxnlv>Q{0x)GW_B}|H*6GGt zNVYM4wqe?B2?T@mfLb+vV3Kci@0fPck~Qi-{NBI1`ES0_QtRQ%Joqv__yMSCw9^Gr z&}H5!&h;+|_M!ndDVx6S!R820XH;*W& z{AoD2g7@g*+nCFq>)xTw-jbTS9Rg3^%J-%ZnlJ5WxhygAai2OM15 z`*qXM<$t;7X)ACH9q-ybEB0paJ4Z}&&$ygUS*E}aOQxl(Vhp~b9ZNsb;K-ITOkf<{ zU&>f?z3d1lwGw#x@KH})k0R9hp<}s`{mH;cT%uJ$S@E%{sYXHfnP}@j`u-nWVSUaU z`5zSf|DecwaIA>~4FCxV2>}TO2L%NUjz|FjP|%Pt5U>C^Gz?4*bS!KfVm8jV>|)G{ zhNNWVLL?5pEZyK>6g;>j1U%pca8{1IF&8ySqo{pwP%kI*0uZ)K0;v?#Bs^|BWbE$xtW@TMGa;zyad|*2VbaXe zW;_Nnz50ZKTgJG=+!KXPPTy-gI&_r#A)Y;5q!Fv(;eHDO4+iAqWC)>kXFj-n)!%iq zZV+vqn^Q2r?Ns9Rutnv;@gWg2p*PMO#Az7gc^6a^@Wlqgx&>rO=jKQJHjez9!H(0y zl$b!L9dIqIp~a)xFJE!7PexjDn!I>P6p}@S=GvY@;1N9RmEjWo0?0aRqb?nYfqIrX zyFS#5Vqgv8j6Kc&`O}|fdk!%Zf1Sd&7jD0?LeFVV+qkII!^NrWiK%k=M(vkRQ)8Pd zZP~y!(I@O4`C)oE*>dclgi#;8McA2!7<6YTSAZqBdE396ztLU@lBXy9Ep*%8Qs8m) z+-6;9j}{V(}L5zTPqOPF}QqEPEEzSrRnWYYqD)$ou8j z(Bt!1ZFq(jKya^flB2>>b*E=?tV~hQSLlO$Q6gW8t)N=;qL%uys&UJAts=@r;$Ta2 zn6)xhpc{sU5dFUHR$cAS^K1@-p;J$|vxp`o!H`4w7eLAjKw#NbNhN|9?pN~iK|=*X zhUTkX+pdgT8c_ z{_A-2=KW#YhkGk*BA6Ed#l3rljc_pz%GVRVGNlgwIHhSCGWtV~+q*=!$yUd6){1-3 zvduKMwV|RQvC1_|KN;?BGW(ki&!+ZSMJkC)f%DZpgs(T{;V$-x!>A%aI;qrT!507* zOV@Xdn@!!r&X!4UXa9ye6vxWag4f8m_;`Kk?D?;?{TU} zN<9BFLuPpCVN++z1brX_A-TeN(G%r)`P7`CY#Y+A@~W$I%GB5~#Kux((^CQ=do!^C zv)_DG5{>6FM|Xnihenslm`LKOqxlTI9V0m>72TqKfvn#e7J)j>#a-sJ$6eJw%w!^W zUjS=YbD#inttiO8e#+Cs7l2BMU|rQg!)Zgd{P81{?i`(ReJwZPqoCg5$j?gwg++Hg z6$3T@0jA@#&Nd5BwhyT9XNT|;hu8d9;)=^_7(v>y#S#?+rr-6Ort9@xL-Xf6-3|$d z9$TAD-m#y5TlL7mnoWc_jTVr)L~4PJxC!6pmFs@LkF5kS{tNBxt@^PF+>Hbj1Ccjz zhJ%A%nj)qxRS8bK$YC};(%kj+bz_YrlO&(4xrI#TqFpT)Xr7k&5N-+VMKi{D;y$y- zJeo@<B%n3-*wfIA_FpwH3fM=b~ zb8h%^K0LO_&UH)&ZwzW=BGs)e(kML#QT;nw1p{E#hTU<^Cby`M8-MnOQYfZea~^ ziub`4ZY^m7c^J%9`M3Pfzx*mKiK-{3onUc+PKaLCy9o-vqzfsXCsyW$$*Xb7oK6M@ z#F(GP)TzZ&8&)f9srccj`0IA!G?Ip;d0<M^l!giN=+0QmN9+k%e^Y^YJ)_Ksq}jt=Qz+YGut7j;LB!ig2d$KQ)o~ zPg8>N(8D1`bCq{C-mvfvmj8dj!s<{MA#K{;249Yo89Ky=Y z7pp|t*xj>E@B*P%{P@4J&oFFQqcZ$|3UF67_`8L5Yfh}rtt~Es4@KZ%?AtAi1 z^Otm{TX3q;g$5K>o1Vu&6R+pJoj@1}D)5kx z%fk4{ZcL==T8VGAU1Y~ao-T5i8|0X&-M51^I;y)tUo?@05#1%j*Hlf!RH+2y7v@@V zlF$vQJn%a@yZO|RXr^ySp1^kQr>Jg&&CwI-Y*eYb6zq^*mWZ>Y+8PuIq#5vcLA+GE za=^|Hg{YV*Q$zB;w3wUG!}ws(f6XTyL_T<`rl((~q8tcezSB-aI2Z zg8k$7rhwp=QJr$<)kXOBWxD5g4qG*|4_)7{F%EVfI@hz%r_io4n+}`j*R|2zN=3@q z_k$&j5NI_;&<1xU8&nYSHO5H<+E(@cHzN;~!o+py?ucxl<#}8*`Jay%! zccu$fD0Plp8mglMmm4VNckS4D?SU{!K~kMMk`h?fDDceI}LWuyH1#Go@j>*x_43G z2wX&eFOMit-&*M~^)AucE}@M1f|8t>#7%J9SPQ9)_-J&9(ViwA&GlX?Lqo>R?9~vw zfd2hdXoc+%GSu|)p`uJiyRTb%WR6>P(}WJ3fb|mh7?s1GQFIojz^c+siGfE|S}+Sk zcpKjSV!n=R&aHH%3*5cSuZ0fRy5rZ$Piv5F;)Kib%gEGA`>v zcE1eKCHLxFLzgNlk{Yi|rK%$Nz7p7zl&*y)d)CqAkC%XM;oK?Mgl=`Hmqfl${s{BQ z?$suky$cKGEqRP8hDY4CHs>W>(aRthwu(fY(OIx3=+d2qC`Nz)b~QA1tV{W}Ic zPV=g?sy*y3@#G&r+t(vfH?w%AVDPlKb??{v(Dgx@Q*Y?CxdlEwM>o-#glXpDxpJ}u zyv3dQ0RMF<>RxVD=9jnr)317Y-B6@9KmK}9J9BF=c7&R_R6>>CQ@$YYaea(JU%p88%UJ?^vZZrO9U2pY*Sr^t`rqtU0;XwO=H4$a-e}*gOcwto?PpCh z=+W$?x-W+a$gKFq&m~4mK?{Dxy(rT9=v3wMO)KgKSM!JE-$~%l%it zpm#L>CypIE!~n2G>tj$f*Ju2hxKBwzI5-v9Vp!@^)!=rFam6Ic;*5wd4$a|2@^@7WH?t=AsdNjD|ICNHKu;9!$ z?iA4v=`=O#Ot%RIis-o2+$U|{s;k?@*eToJXe#QD9Wr?K?HtTjqkJFh$LEDMUPZoM z!C+8J*&fYYr%+TgFJo9})wZoz^{T$ff@GejdnWnZNU`Zt^&L+JUQyGFb(zz7tV*-- zP{C)S*z+xkeuUh3L29UQynNLIz9C32P1By@UxP^BDgRo&H%^vB2n@jCZ(WFKeF5mMPG9(CJiezlaBJtUnxXd7}Zuge&@H)>NB3ws!HP_}Joc&S$uI$n!i(&j9c zCtxONd(WyfReumqJTdCPd9qWxR}T#n^~+x~-R{}N?PUq{JZsHgz;U*FXJgP-TpJcX z8%}}BlBZ;|%89b>lkP4$0i* zM1~1gnxry0ZuP;sn30pHX;DYHj`|wqDq$@IexW3U{JHIU^vvepuuaA>xtGjz$a3Xva<&Su6(KQ!X`;i3azdaX zGO#(>faOl-HTzDLA%ySpSQu$$hs;fCRBcD}-3bI&kuzZ?d{XCJs{6t6lw_`Sy;!fN zX(B}KVOMVS@nrFPatH@6n|X$mg;C)O`a)4rq5L`Tc~l$zz8w}Ruq$7OiGN?xS_gXn zY+oP(LI$?pQ?wS{BBB1L#_^R=GX1GWgKQb|MwHVVi{a?Z8X1A{iZ^c@Pz=B$*Zmki z8bdqD#6AF|;#3wW^m$`JXrS0-tT|}efK6+GgE~=J+UBuUnWt^H^iFa1Opf+ml)MJG z$fltIlllTcUbh~vh}IjR*j0_(%foq);Wk!jTfbg{Z*U5KcOY7OSJK-6qZ0BQV`jDn z0T>AOW5O!dW1xNJ0WntwnQNJBSRE-A5;=4|KY<(xEqET2BC=Vhoe#7b5q>5t50c=C zJND-f(JNJTUoSD0E;IMOKkoRdq(|SA*tnv;g3$8bbEI_ zjyrf*bto`iXk}m&xi5dAMPfi`)_&pf%yzb;(SXC+uYTbQsJ>Q7c32#sAM#zL#!U`E z%XotedbhT8tdwME2`=hSVde%=-^hyt)dF zAAtAdq$7Uj_}NJ)NYeeB9W6P8c73jmUhHcw&<5AA^Fi6M;WV}z;+(x z9;)3`noScYOXW^CXPrW1X-JTJn+go3-+XCWoNL2j1QmXHra$kpzFlxs;psN>Gxa8C zIADcZB_eG;p@J5N{h^rB#*$$|I}hhxGb|S#0p0iP6uUv)hP9ir4i5Sjwm|v;C+WtD zLJ32-ypp55TTyMkOsxWD{$RaAS{yNPrLo3BbM2kQ@unRGA!PMuHS(5f{AoFk80fbe z++d|+VKE;C*DKwjqU)1UXq26Ac1tI)vY4y>#cD4$ntXA>;|0*PJYXPobRhj;H8&GY zxj3Qy$~sB#9g+L`s~USay5JDcyKp5)43e;VMHGEJ22c6D8nmzLT+3W`nA237?rsd) zNTjm^q-XGlf6r{=a(xm7gNvKLtJKz;Qy_LoRQM?u?NZo0(96}5e(OTbhI3)|92-Bw zkvh|yt*D3}O$^1+U&xfcooo@y`fXxuW<}oJo7nu7wGIc{*u*kmMo5%7mKD#V?Px|q`l6|r>BjQyiZ%A@>}M|-*kg}P9)iIstqD~)}H)K6{&g&gV`ZML@iZDXVs@@XzED9fCOiXY;~mES5F^w%8B zp4~0tb?4rmbp~0iYTwd=R^=zYhiN71Kcn2~%FvjF9>d!4Z$EX9Le@Xx6(jct$;s|5ju3*?`5|EyvuEZmhrd0Ve zQLB9OMF{&ZfaXuN_VaBNnbMD12AP*`3#%S$tUM+d88wksx6nWc0-Has%3|e$lP1W7 z-l#J>dzFTvYF0QEB7%WAMuNKWM>mxY{!grZS``ycpLC;FW4+A}5Ts-v z&f?H^(^)owjWmHDM60XNo)@O?lZh3VZ~2`5hDT{H{Ste@W2B!TG|@KyY> zA6^#<4)ls4VsvF8Z(LVg zJh%sx7S%_{Y770DB>KzmPPD2xR&p0q3$NaOxc{f@9Y(%>9G~eEh6)RZ5DaOtW0-4 z3iEP;M7f}nm~aECGcR=v{IOo|ERfS_Du~j#A8{KQrN?j3Zs}-&7DY^(HRUi&LcKL{ z;A{P+_}G!UcSDsha9ait8kAmw5L${164*9y zOd2at4o37+Tw6O9oguHrIHWB`gis3QEF15${^dWG$@grMeqj-=_(Pdk_#}|mRqb*J zPiaoATHmHYqO4e6`Mux+^1a+jgvyT zFe9BmKhuyu6tGnBUiK6=OmvE&nELQ8OY5gy$g5RRJEugMUlT9DZvk->qWSzCJl znSQOQ7`#=xUf1<${G==9dNOI8p|`Vt)||~$IgfY7VA3M@K9&iBHc%p|Ci0Ztgppe6 zS1gU|;%?}SnlXP(poRV|#e}w2B8g3r_XW%y(fyqnYNF^`!T^xPvT_dN^=RwbZw$5*A7S z7e;L^r>N3O`b@P2w=C}kCVkYo$r%U7z99-iH;PJo^3M!a=>aca-SF#xg_FO%w(~!} z6gcR|Jw~iZ&WGJ-tz8j}ZJt#+X|GUg&2gObB`$D}r`G|ijwRat{BMH(h-!5YyRUYA z*L*~uFCV|<8?X4OgaPZ2n2!%kPFu|__M2>Q2hk%<#@Mfn^mkG-C~3x>XDV{`T`{s5 zZ3k5JU5ja?of7toND@kOq-uvTAy5Vf8i-ce^sHNx?Ut_93UcIQ1bSRIW5)@Uj*iu# zq&G*I?C!2I8!(CD>c4rAZjA#Q$>M*>H>H3mA}!uEeP(UwYBv$CMkBCdbzRRUNLoUx zhbiR4S^nv2AY6A|;el;IWUw(~XD`hfScqqJ*)ys(6m7@UgP;Lc4B%v1Pi_}itEWvRjXN= z74B&)J?wHaP$__mz%Ht~07-4W!F~-#kG$7EqKxP@!12o7o3H7MeYlEw*VbE4`#GX8 zwNWh;Bn3Z*|B^yqDC+NFKh7UR>i-6I!oMMK)9UQ)MgBUGmu$}Du zYSPSSr5mg|{E8lWl{=j?yYMsYW_92!14zX&5EbX`W1q<(h~T||!_p7no>+3m!v_yX z`g1J3_!7%^ZlAt5ybs{y{HC1@5b7HDt72WL%p1(g z8|d9nOiE;bd`N9*a^3pDrtJ;!&e9dh*?DmoGj7^?m5Vl|582XUSx**R^MG9<|xU)5b}90bf--hmVW5#)pd+C!tE8jKYUL}LFAK7HL4vu zQ=BMqt3L&Mk0{#kQ@P=ya*gK`P+eDsg1psT@a<2Gxn0GAt_2&9hlj4r??ewT+%Y^m z{Jm2VUz)KMz($W{QkS%P$NXxD%_xgLJR~iza7lBW(yq;R1HQIF@yHdnakvc~k*1aO z5^tY#UJf!AUT86V9Bbg_46RN=(HpQ4W$_!~^KUm93_JEjA#})&v)cjNETfUB)1KUi zD%u|ICGy_*-vSB|aEO%vlcZLsg{sWue>|jX=&)sor8NaV%`P6fxqTD>qcAzU_XB;4 z6s;!1gd4&J?J6GmPS!0-V$xB8?b6Q@95uV_=^Vo} zyZLLEetu?ARfI+5xLTg`4RZmQY2)`j9IK>UNXBf`1MWW$Ka7RS5b<{GbhL=dn{X$wWPG&($MUA@ z0Vn}>#yf+wn7@)8mMQV1Yp8`yGmN0?b;L=+szCvU+~zX1HFNCUgXz`du{VmKk zau8U@mA@&{*D&3wUCe_(kQU%5TI6p0$yj4w6~qHZz@{sf-q4P_4;cLXEztu zMFcKa7F{ZZoe{22V3M1D)3`w5B^aMX0lyRXuGOwT6Fx~h_t#DsbvQ(Ctl)A8{uD@YU^vkQ}F@%&=- zg|LeEC)i5)mF&eSAy2Y{HbMn23gN#EZ*(i~7-+_G_BOcHNoAj?@qELK`)zUa_03+F z{=5mBRt4pc_~hQy?ar&_AkU;oF&C*^9Z{_eX;$cXbT+p15gfH?I}jmSVt5NaOq?OfocUngd$G)H~{3&WhV7{OXZV+L2j9S;(Sp*ThpEKVFV!PfK5f z>e=(jdV&2OP|H=3^(vJmXv{uAoDwku;Vq1kTv%BO!gcGADrz zcJL;F6IizIbA^QUKBUe*NEWI!&4G1raY;O{T7LqD8oP-wn=MiXFS2qb z%B60!uXy}GW~n?$Now2^FZ}00t1(OPmo&vEuZnLbQTI5}916p9p0?o3_ylPQ6Tj=X ziIMhlQ7>^KR6>5^jT2#((61%f?aTIcLT)FBtnrp~6I$?^dbF&r8yNlz zSIoEH^xi)*ZWQU{+4PMPz%S2J5AtaV;P*#w$E=jT|31h!$i_=AO{V~5h?#6R>KNR9 z0M4@V3^H>ckP$llBPJ`bx>?be1_xwY^}LmIWZIxv{bNJ(C-JvwKCB+%oiY9Ja3|bb zh&x!X%OpK$L1WwxAOe);AS}2J_S#>+ z1)Q^sLC&|2AZ4rZc`e}C!;(1{aWcEsAXn<3rwbhS!^y9IvL^^V!|ctbDn?qjFQ}~= zKPt>qsRdX=L|La!O@TC#nQ!`Ie5-Yyqp;Bn0cD{UWwzS)@C}M`;GO4`RvrNJHF$RD z(t$XOfpF?FTic}Q9rX}3K{ge#>Ur1fEOZy_`N<~?h2m*|&+b)sYqbn;gruk~>(z*k zt*%qaw~xA_h)0_v7LZWTvbvuMFhy&-wR&^3r$gMYpD(zbQa4sn_deIL9*mtfdL3Ub z*PnX@+n(rZ3!V@@R&7=|0l~fF4mxLk`xG}0jjg>gWm6~TbU7>;UUEyf&fr|mNqfp( z0f9-oT9$+NuKZMYO$)Emil|la+QvGO5qtDZ$HvtKGf8TtF>P=z#lN1NPN^zIBwfCg z%7W|kfQ?bNJJ5Sqgp9#kXCb&dAtxUOzf;%kzni|qTaY&6aY7`heLwV&C&@YEbfr$c z5N|1vlv^BajiV8d(>5+}6w+9cBQXMP9Hpqvenr9wuYX9;+yydWPI|oUPaqJy{iAcb zR4xwv=dn-@r3owlvC)29Yw4+qp|p~+e=I*N5+jbUBL0JpQ%YSD8IK5e0?Q?G63alQ zj2e$R@*nyFW>xDlhI?}hQ5;jmSppq#AI!+t2ARiQPPk4h`LVZ=ip4rj@bq#ekW!(-kv9T0>!2K#-is z;fm2~Q#{nb{Zv^ac<)XPS1{k1Aoc0z2exbC4pNIloxf@-R&9Khu>lU7X_@;1U04vS zBz<$?&iluW8Pd_`G4A6tD%6M{T|d*od)Z{bd)Xi$A)(fQ8bMVdEA$Jb=Ylqt# zLGTW@_EHRL>`<{`-0F?~8ECHoawv1N^%uZLX;}AK!ZB}k%B=Tw#UW=Wl0M_Ar=*fq zCb6pP8eCLoEYsytC5#i_wX=k{QdVk|Mq$^AHBSk|BUNRfYIsbWX{P9rkH?ncE?mq_ z>gP(YuzFMx9yZ~fKq2sw} z>=yQc_Eawb{DDbE8lW&nK)S2SlHf?-L~we-`ZRV9|V_XN7q74$4S zcZyET)9vq>*76u+XM#@4Wem6F=q+4KrO|0fw-Pi=qYmHwO5|kf z)uf5_Qt9gJz=mn6{s9==twHYw}7xqu{5R5qt1!K6S07HTcu!^&A{e12ZI zQCA#3c4CH@dW%F|WKpG>Xfelyq8C6{!GG->YM0Q4Pt@?ST)*yz92g{K@wWV^lGOoG z7M5#vPwx|JTXo!TEOpe6a^iJEqoa`ArBFkp>vA3@x`UVR2nWul?_L1@o%vVKFMyaI zL-@hdxB|}V?!~D;ms6%tBOQU~=m7_nVWvWH7BCF0YYkzUOJodL3ya$Yzl_aVRH5B^ z@6sc@7RKqll*Pm=i><4n@rA;8aaf|l)uZOAU_96mZUFn6mL-6_ibD zqmdcCm)!kWZLg%hn;98*%d+IZ8p|A%+M`%peU8isky*So|mG#4jYUaRGaKsEF@M7Ohfu< zHFnsb*YWR$Y5=lM%Sl_@IYsy)g{lKF0}`7XSG*u5t5Pe<;_Ou?pyBdy@(Jp!dN zc>&is3c6(43nk|lz>V__evPx&!Cz;y1V)E)!`Ew)CGBp>`t99V5jqsbua>Ei%uUZ>g}Ledyp;V4_1w_G4=a~}@2p_KSt$|-&%)f^<# zZonP@DlKti@?9MZ-!FR62rP?mA)rvLI;%~9<8E5hDH%*FuC>wC3ji?~<0@KxpmwFV z`2(8tqw||LO#z{?s^lJtp+u5E-A$flv|Fe24CG-MU5l@ z&?(#|zr>8M=Pkaslh0$NmtIJZ4YzE~lU<3WY%N74@4#_G?ykuRZ~8B$BKy(XzMccI zb!=2T{O#24@v&W4r3T4LlxN%Qb=`NrmIYgz`5tS(x)}YaA%Xd2ZnL^Vb)S{8SfoF> zauUTOkz7VH3DZW2m8$6qyOm6?otfnZ#!U;f8(VpJFlKmouQ$$(L($h+j*Xj+y+yp;=&9Ky?)UqyBq-NF^%Q@LoQ+i1p*cy!SU7S&T2}Yse zwp(n*f{=_;z=ihU1W8xML5+&I1Z@jeckoJc@lp9`d-hM!V31C%W^NK8^5rH2fyV`CRc)e;iPO2P*fm{lCe zV>3{rB!^Gzfd!>-IpN#37OP!;;}NZNbDN!=&@&QTINF?1;f-dCHc@xIOOs6=HZ75K zPc|h-nY(FLOsTl#(DG&5wx3SPAwo0n7#H}g&hb%dhhmWT<|cvzV#zo^d8&fsuLo;p zeEA1?;g4CVo&IN6O^FF zL7noj)`%9{0Uix#>WdMGnNX!L#ScnO+FW!8f3SjeG6;BrRZ`!R~6FtSQ6YHbtWde-E_cJn&rhpjB#+0NjwV)gn+gsegWh< z6V!=w47c~%c0p+V*wTrTE6vVzr)7>EY*fQqNOK+_S@$kV zx|ZJ7050J395j8S563^ud6yUv&Plu<#{oa;J-^?-FSXe`t1rvvD4JR5T7j8_t4Mv{ zZX#KWn~zLrR~_dPM+X%`*F0FD>gpLeWuSDdGmT-qZaX=g@3a6(!rV7O?4$cA)ftTv zmifK9j)g!i35mY?R0 z5d&Uxph$g)-VRL=d|(h96`maB{ul6yn&#z_I7oT*TMCP}CB4A1*s1 zxvlz%GXFQJF~2!^5L^yI2)`43M=^k_#hO?-$7$dCP$O^Gd15OyEo8#H%sMGPf$b% zuTkOBz*s=2>?v1K2KeEYkdfO>F-ZekJu8AFlX^p=`;>sc1*GuZ50mYo_}=m^mG6}~sOit3lMM@U_OSiah%c~))sONM)~qX0 zg5}O>-{3y2R39Vuk7X)qX=$#S5%Akrc36K>0>*MfmoC}b(;>vZb%!$Bug1o8TVG<$ z3>g4J9IOn_>wf(Z#eystMODr7;h?<~e{m1WimI$ScKG>oT!|~yrl#dq+6H-SiJmvH z9af8J)M)iF)?z$@uz|;W$__`qs9dBvUVoDogiQR#G~Yg68dC%*4YHVkob?!XpFMKa zTSgOC%WRCwv#mfg#GQ1fQlLEBp~5-v^U<#ibB*^7+OoAIr_k>LRf7t`*hCMlM=yuX;=* zreO?{*mR|Zp#j)tiAUwHb|xYUm?4pl_DsWK!$l;C0W^ex8jm&p;s;9-%|~}5PX{em zuJ0i#D9NuxmRIKNW_8JS2JnlgrG`JGZNJ}_`rSb7M(%aulvrbDs#V&qE3D^TKRp(n z+-@^!E2J`4UZAD&mNh!(pu_PgLPEJ)=fLKxbaF%j*rj;$_Oue?YvNbhZK_9&>r%wg zJse{|M^FphiB5;^h^H8emO?M*^dAz`ztwiIxoQ*FI!Z!cWHH$PbFIdb7VkJ;v%PR9ODp0hDkV!Mm2R%&CzF==8p0R9ge>Cpj``8Tcj#tznd z6#$@xN)e!H7v^{f6)V8eFT}%IWby(?Jblkrx)Ky>kYGJMu0MJ>wnPVRW6OE}5?@dP z%US@h+}h|&aBW3mriUQw^0p!>h~8s9$YbXVeo#-pDEz5Wan}4XIPxeAwrirjI(7>p zcTsm;a5yW1%#i0}rsmY9vht9H<0NQn0jp4U*gY|eup*lD)CrYFcmga0=bEaa%@DQM zE17CvYN{7Cu~1xF-J%wp7I-y5#{kC%?QquNHxhEebBE#mbaYJ8{@shElWV*m%XY96D<(n*K}2L+LS(4ScgbwGP@?bt>fql5djy zIoZqdt*C{n7-_JFnx|VO!%sYL_1ljl1IiSLQeZu0-b>-9iv&e;(Dz?M2oGxM3xCM> zxEs1)aA&}z>X19Zilq=2L#`H>*2P11Fy-*?y;^p5{Bl*%IUgh$;Y+~8P@c+USyk}@ zh^qI<_}!O3pMO>*yPkL)67ZLLMj0`oMcEzp9!v;Ur|rH*S8#X4FMI)9tzNM=iP^I% zm~k9A_iv{RMvSCbo2WP{iVB%S?uYGR2U(@K&$I|SR0G+SAg@nE5nEf%*LnX#1O+&l zC=1{t&+Qb{Jb;b3QKvl7U5ewk)mS}JFsYycrTVF4w6(%j+Fv-;ilR=$<7nm!yU7wP zo^;tvhLxUQ6#T$;osDUAKI)eUiQ^m|4=v=|7*WaCr@=ZMOcczGGt})WUHM%N;J^mH z`}wMVn!+EU(KCD&dB|D~HON(ontp)Y)1wnx;OTdPP2x_to-%`$isZ1F^rY>(t1%R@ zsy!Yz_*)iOT(4?h>kE+xGv6v%nB>k-Vm91)4c@G+*{{9g?^lhGrG=EmPJOVih^!6z zsUAe?1p7W#a0!Wi0E%Z>Ng&L$t7wKIytAP6W*lMW~J za3}(yNeAg53W6d{iYOL5U+{Ond+&Fj``o|wKYQ(G*1Klb%)B#e?dgj^J{d`mzz2IW z`rtF5n`&uUaPAduQ+wO;RQ#0sK&&)8#rVpLctNaz4~_VTzb z)BHIhIHKiYuH3b2Z)FPYhTy=H^~H7}Cuye+bqi>h2y+vJK;x5t$K=u_D)1Jv@!5Ts z&+@|`HqiWz`i<))owxl5FiYV7{tw`X6>4U_CmN4v4rpK=5HKOC*i$!R_ zz5RyfVtzh%T!eO#95svH#{AqJ_>!*}>&Y`uB(aN1LN5!Ls@kFZ!Gjgsy~NlC8Om?* z)}xG-2U}OjeR|xK@Mcqy`m4R8X=9a;FI*FbUCJ(V6u_G5rT4~(=TSY(l8tcNFAkLD zvHXEb`Q4IGCW7ie7+tLe4V;EmzI4-`*&w<*ubX3w31Mc^5PX~2n@+J;HufgA^oDEZ ziDF&e5Imtb0!aJWR6c}3lCbk}Y+d&8bEZ&$kE^ndq7Bq>^gfHhkpT_HO(LmgE)!NK zrz~MR_dcxFCyYo_Dj{9ekM=UILy(L}Y=>$U#xgqCXZvZMpin~!Em50Q=6 zBzqP_$0^e6OuLzcuiVqDi3Y0U0Uo|kHotQ6jqVb(Wr1!HID_b7$2T^+U%TR|U9fn2 zphsO5b}-Fg9^)hvDz#n0As9|CWq#!O?$x7RtkkE`8m(?^7S2uMlVF1c?v%|tgJRPa z1Qot&mtUXAJDqKM8Net@kV}$m(+%v9cQ@7h?SQQ((P{5&m9M`S#xA0(HhvlajIr?# zz;TaFWwEZ0*~IGKV)L;E(IX?)V(v>WivPV10X`ds=a`z#1(9&{6K6 zA>^M?`lg3=_X<^AyFPy>;_#sYkaGqcVAK6%U0z!+xA|d3()hu64T^) zqPrK#{sh^>Pi+n(UByb5w}}R65BJOIGx%GIZw%5y(Q9_}G_W|2m8;V_vgVb7jM(mt zS^$lHp4wzr)rPijtEp_cySxSYXP0XA%Ndi#{rh;K?&>OrYxbs5cR;`;IxsO-E0+uQ zax|S{c2)md5lvbVA_KK_Kb2TMHi;Q_K`*d4^Lz~9jV~PF7$D4xW^&w^hCAi@S+_^b z73f91=D6~$q!8X z@!VVkH&(_BV#>2*w!L=Avv*edz!sQzl%Zi{o*;M~L{Q1d3#Mo*w_$?8BYRHGbY&ep z+($RdEZ>LKMN9cX}+NOO)$ zKz+*e9_2xOZsvT)H}5*KfRT~W-USBj%1LZ5Ma%2sb8KUBlW>zPl>`A7l)q|f#h$gD z;Tqa}0)NdTFK#kxiW0!C4k~b{bu1S$Ac^f_m|AreKPQ;3siQKZgdiAaKw;YD`L5RB zw`Ds(ROUo)fgCfSN2WYE`|#uM9^MJ9$s~8I-;F0qyJO6{(P^ZXoACvqs6|!8ZTFFvZ7#6;0i%g7(0PO&PLrMadoX6 zhs3(;MuOp7qHky^pC3OJWYtR?uk?^h)ep2qI2f*~-OWtDxsrC@X@TP`EnwC;Vn|2g zh$&rZ=q&{&kM$DQo~mY}Z-ll9dS`iZKmmEK5F>+(^HQb)VkC5>xl%W%w0?PxAYkT+ zW)fWBxOXn^C*^DGN4|S;ye}U(OUp2oPw;W6d(^vRO)Kd1ic9iMor|PCW{-uZx|<6e z#x=4TSQ6Xz{@>~NFaDXC3l!I%C)$~GAbku=sm?tB{x4alJ21=dHo0}qN)|8+@IZrj z@oK_nUh0mMuk>GtL@y-`D11#b`#<^ObJUHj>fqKE)?u zyc9RWr7Oi+4Q&fMfz;c;m}izBVQxtS7OdK9i)NLdD<$a=DzNj)+JK4)o>XolDf;4U z<6bId-2_kysL-J&HCF#lVJiA!zK@Wg(VFhw|H5PIGOYIy-E_v|ki?Fp_zkFz_;@G) zUMMagj~*($kVM~eb83O)wf7Sc!5Xcp;H%B%GD?R_$W28ZKvVUYsITA1r!(2l?FqYn zZ||%GE&eUGtH%F>Q@T?&V(cp;m&dl5$~MhAj~s@yXUYc@PPY-_$A2pj?(*WM9^oo{O1%^8*@*>s@RZK98V9`FqYXXWM`5v{~)iCl1T z6pP` zXJLFnc$6LNxtP^8MGGI1-qpVhUx66w_Sqwi{Bs|@n^+Qy7j#$Vl($km+N0rYd0vMw9yJkh_>akU8K`JWY8U;AJh>6 z2TH+0G6BmcuI7Y}u~O+DhE*Wz?T*uk#N>Y1RuT$-^ViPz8^wq*FbDxdHMg6k&2)}# zXKtftj+0}+?)86_3P=aQ-JDu`PmSZk&5mCk2=TKqG0*=psVn4+1-^&z9E|A!Eo>~V z0N`YlVfCY!F6+na>6nFzq0H%V?g+}*{G@@L=l=leqm<_C_I!0vEpr+cjCUu4x--(n zZWw$Q%8)zF7_QD1%&OH?aqjD&J1Xo;0xN8=8x(VIKk}qcaKVI_<*@Tt)vpL4ZH?Ca1d4U9`1gnr_kCc$R5Nd?Eos4D4wHBw-MW{Kvwbgh z)1cU?Yg?xB#?q_Blyd*kMIm?=uF!JDM6_P}MDY}LOZwc3hK2R?nLO$kQXv|-eQM|H zq}i|20nYb?;AgilB!Zp@wl)`W-P41-i-&^pdFaFsW6Nw0rb5lT+$=L#>*fi6Q@R>)~FzF50Uq+KLU;Im-;pIMY zSQa&5MvVr&k72xtWzIqw3Y3gvIY+ZHj}y>#L099P9JKROcq0l@fQ?%in%m@2K<@-r z>7$7iPaSQHnbg6^)n+v0op-oV5B(o<&7C@iIJwF)LF64>k(Ok+pC#{LOJ*l52X5Am zFWSNgyADau^QV~AIei%6IuNApLj$a_+de(pb(9R9xlO-=hu{}%2Z&eON8NNYJIM4( z_Ps#kMlO}^5a^g#QQa+RxxTQvwwE)l%7dc3#W*)e@&)Wl**+`Tc?){;jb9hCbHdbvk)Jnv)i^RDzUh!kL-k&u^|dOoL8c_pM#%FC)Z^s6^#!|%=?o^ zHCXY0xJ6?=V8MVlW1sDAS#^Lf&1gmUZHG@YyE7!-%WuGI`c+Anp$t>eS)pwo{l~1J z_Uj^+C`SMuU}nd%*)FOe$S!`YH}b~Q$LKy$jYAbvYxn&D75r`}eQ{t;6@7vicJP;2D1gY5iDxAi-C;(zcKbfT}#bB^zAAAUJtj@L^GD7tRA6 zg7>4_FLTB;+^1JsHTw%vwN#ot@))gT>*RSha46?3PIja)_-&|lx3WuxP7)f{Ojs(z zQ3-~_9BayQ3k7mHLq94m7)6Y|Zs#1j`)UN8=J3K!v~AE${P$U%nKSJv{jqmME7rhr zE^AaL8S=E{++W3|br%223G?)2dac(_ZEKn+d9}k5cXSrZEOk!c-JEWnH8|wvP`7tD zU|TN5Q10tB2k9h-(E@4F(rwMHDF0H++%=Y}SdqBo`1r?_L1Bpaoi7j!j=4F6ECQj@ zfc_F}Mj;-b?Ds)r&ctShi;d{K75?ex>^)8k#Y!E%RHsdQ5m7@Y-F)7JtHa`VRm+?mNLFIIv=6*k$!FZUso>ANdEc zPC#0Lapb>461&QNEnjsf)txSVJygHk-bgGQ@P(|6fOKRkOiH^GyMk{X*I%ic(zBdN zZ7M2Gz2s{mTDo6fvM)2cZ<7~Ga&#-9Yhw4pl6;>9@`tN?IiWLsg0}bBC4U?R|8U>` z%OR#U;`LzgUbh0a67G)UGMVD%TX&abxV?K0DzND^NfUlY< zeZP7?n4jWV_skvC`QCI9Up;5A$B)0x7f2=Jb77S_HEa4JpYQY!P9Tg1asKQe6#-SG zLqa)ia)~9f7*O)XeCVS3yoJ|HyFn@cj&Yx~NzPzNwTt$5%ok<=Cv2{;4IvOJ{1MfX z-O*IFVB{8Qi}4wW3)}&_GMz@x6*ur9q$sAAWUHt{+;Yk*h{%2&EO=YECCtJ1n#SRT z0bPux%g){H*%E6OzEFy1j2M0n!yAznnKUFp64&H=6y8%o|Z*voPf@+3S z@(!3bx>06ytDGIC{>U75h~GK$=q%Hnj~Z${13LqQxgiU0}SIfW>mx@=3H@~XPSjBjJ!pm@j3L=LHDvC`JAILU=%eF;o zaho(ZdRk`4a=Lp?a4_>e0hxG~b$7KZ;iykM?iM7J=VR^``+}?ou{Ybl=H*ik2yPWi?mKX^B%FZ1gPGzvbd0O9h@;5fPyAY_Iv3EPrWTi&oQ*{h+o< z!YDY#^eFfNS(UG28?AGm8j$ZLmIR3ATq(kVUY6lH0 zzb&Fxk^3`(!sV&yxjsRq2j)}Nw-wpn*cf3XxTjRJ-X02AY>%8E2OWyxwEf85kzb$R zu>$~TvJol+Ln+tI*I69u13QD_0T~+XQLG04jD^Wu#gRlJ87F+?A8C zTI$`q#G5e8@w9^OJ*0^aP_s;Nlb0M(`0@nU@fuE27bWHJbBDQyu^Kw<6Vr>aLOn(KL7BlEcnR z2iyQ}~^3OarGtFzEntt24+!n#gK>-MKuyJRD%k1eZ(8M$(SqqA48m zM~(bmWST$twn$gPKg`XWNYHp16D}y8dbz^YO7x@+eu4=g{Z`+Xs3J zK=9;o@)};0Y!&?pJqRI~l@6_NRTT=SV_RrA8knDKvv0?vYN(ZsT3p<$aHl|eT3~Yc z&Drt~0Nd*#ZjKt1cFZ0<+*1%QliWqRAzec^FTRtju3kBxVWf^^m#omDkG(^q7&v8> zsXD&Zna@~R=oD~0`yW7&q9sly5BvwM3YadaSopC>Te@r2tlQrSlngZ-jZDJ4+25l% zUPVnA)}?M6yKy4m-Ivf_$^`6Zi!3E*`sG^F8l9vn(8l=9T`z6}X|)SbA#QKn9Q`bv zB3-|G#g~rlgK)>< z@^Egpb+~3t8h|RALY*LckPy@&!0`OKf3D5LICF$h;vf`amSD+2PZbT4ZYnd-5M^f; zzm87I$#@mnd&TwTM|EbMWLYr7$C>Z^@CLT@jQr6>TICE&OyI?bGxDLQZ7aqnQJ>?V zYVWnOIiIyU;b$5)xhHfI1NK>#JZ~fhHKt7P4`*MdcWSfv#t}a8twv~WYCkd+^+YDy z_^ne z)@UGq;)>%WGmQY6V;fR|es5nlA0wg2XEpYXzRxLnXB>VNgj_;@FnMA^=?#0R_8Z5Y z^5%^hmG}PgBVcrl_4%0e&w8{5DCbLKw)!4*qbM|_D|F$`GEZyY4qMjfK+fUqcju#l z`T5oS;JJ?6U8oCuji^f<@DxR%#+&yIus`a_l90(XxF`9v@xd-RF~pHm7;3OG)!{`( zAt|9bwi*%GRMn~dBma?$0gqvA6;59LVNIV?^rvwzvfRKfqbS1M7|f?JF@s@h5MlQ# z)b?jnVZ4(I5jSpF@cGG_7~n)w^{wz4Srw(W!K++(pJ)qG1$OiSmE(=8y14pp^t~IS zhuks6{-;8$C~WTJA+7P)RO_P>MQyOBNO_X2%_ z8g=v;VEekT_MW+DR*AZbU1I5zGjt-0(wFw^VAAFZ=YF$9WyuL!kiW_GJ+;e;H)2p} zQ_$YqlzCJES52=Ew%4n-SsZfkV$?RrWxGv48B0}(4LP!zM*yE|1WT<$=4^ye0??5t zqqOg2^rK5tnj)yUC)4oWq~=EI2`g1h9Ywq{_lWVCNP69r;BgGMx7->ELJ8A+@;J(- zlG$l;*eP&*q=BvTgeIf+XJ0v7+{c~$dnM847H?gWQ!O`oE1566Qv7hRAY<0ZfZ*y# zpGfsIqh_E2D$--v!FA?StYM}>lSPW{zdd#w_*z|`CdxL)6z@H|og5;)XseGemKo%7 zXdZp|129EdeCM36ZBX>(sK=VMYNtoo$@p@nrYD8^4>WcvG~qm-&+=;nd*}{@Y*-fA zPKaEK3HQ1DW!^1vjos2X_Q&N`;5DKzQAAIa;-}Cx)616p*+XwJ^PL|SqE6Zr%c#=CAIVA&r%yonbm}H`Xp&I7A#0+G=HP-u}xoC z_3U*9vS?P3g^;1eO@7N!Vw;Asr>pjup;4uZgz2A&QJ-vC*@P#3&hPnw9IV0p_e`lF z0tW|iE`O{wG`@~_(fR0r@%{*_agx2Ro)8pFJ*$r=Z*&$$xULJH@7FszL!G^xN6ODA zD!p0^5dI!^3)97*kL^AY_hciQLJQ9nJ~w3A7!w@|lrMEcWN(7y;I)t4?VtVhORR#& z*Pg>-Gzw9zevMY8+2K!YT9?iW!vw$u%?f>99PZ&f)yFb1PNXRo{U>KC_8u%z6QlQk z>RnTi(U-sSfuPBsyGW~bQn*3;$Q(UVP6`ahh|={^IC_Wk`6_p>PPf2%3J{pww6Aii+t@xag0Q;E*TultKOE zK7u=gOnqcabXhqBN9et7n{n$?@)oD&GFa%bOc(ibx-hcRq+XkXHuV1l@LIHI$g_^^ zm&{E0u6&vGICDLaBhk(Fd`(4(-7=v&AO8xr9A|CNkD!5Cyn@INnFpA`vxn1?`MD>+ zT^WedNAW^@1!5Gr5c3wurLq@9poViXT^2~a@M{_&-wt0V zId?Vj>UI0*s!aLH+h0(WX=dEq%e!9i!@2_d9(xt9VMX28{CCi8B&7JRN`0gCimcZ7 zSD9B`hrATJf9joO^EoYx)h{0gM~sGozPws3BW1L=itrzlhg2$<9u`?uhAn}a>CClr zx@Kk3>Kfu9+)5m|qE%$-OMt4hM$>DUkhc1C__~65RkdjEBCWHlgV*Rdf^va77V-v| zP-x<6DmogG3@N|sd}5!U9U@2SuRia*noEsb6jymaI!5XyOhOCP!t3k~pDu)wgj{@& z7`WpxGXY(Gk@NIzRyiY2s%7p;oZhUH{;?3nT`gbPLE|HDuk<{mJgPo_`UmKo3eu+57@Ex9!uBE0JDt{u_zJ zELNqIBPl6^wU6m+#|>ABWr0!{jhTc*t?p?BA7oXTa}|J|s68H<9i4-M-0`)f=