HYPERLINK "https://doi.org/10.19164/jlrm.v5i1.1983"https://doi.org/10.19164/jlrm.v5i1.1983First online date: 04/10/2026Article The Reflexive Local AI (RLAI) Framework: a methodology for AI-assisted transcription in empirical legal research Lloyd BrownSwansea University https://orcid.org/0000-0003-0945-7472Ryan Brown Abstract Automated transcription has been increasingly adopted to relieve the burden of manual transcription in qualitative research, including empirical legal scholarship. Yet cloud-based commercial transcription services raise serious data security and confidentiality concerns, particularly where terms of service reserve rights to use uploaded content for model training. Existing scholarship, notably Da Silva’s work extending Bokhove and Downey, has identified these vulnerabilities and responded with a risk management framework grounded in security practice. This article departs from that framing. It grounds the case for local processing not merely in prudent risk mitigation, but in the specific legal architecture of UK data protection law, arguing that local processing of interview data may constitute a necessary technical and organisational measure under Article 32 UK GDPR, in the absence of a vetted, institutionally governed processing arrangement. Drawing on the author’s experience conducting thirty-three elite interviews for a British Academy/Leverhulme-funded project on pension governance and climate risk, this article develops that legal analysis into the Reflexive Local AI (RLAI) framework: five interlocking principles addressing local processing, human interpretive oversight, contextual anonymisation, transparent documentation and reflexive engagement. The article situates this framework within the methodological literature on transcription as constructivist interpretation, before analysing its legal foundations under Articles 5, 28 and 32 UK GDPR and Chapter V’s international transfer regime. It concludes by identifying the framework’s practical and empirical limits, and by proposing directions for its extension beyond transcription into AI-assisted qualitative analysis.Article TextIntroductionIn recent years, automated transcription has been more greatly explored to circumvent the ‘tedious and time-consuming’ nature of manual transcription. While cloud-based transcription services are effective, they can nonetheless introduce serious data security and ethical challenges for empirical researchers. Inspired by his operational risk management approach, this article nevertheless departs from Da Silva’s security-practitioner framing. It does this by grounding the case for local processing in the specific legal architecture of UK data protection law, and by developing that analysis into a transferable methodological framework.This work draws upon my experience as Principal Investigator of a British Academy/Leverhulme-funded project examining fiduciary duties and climate change in UK pension governance. Faced with prohibitive commercial transcription costs and significant confidentiality obligations, the second author (a senior cloud engineer) and I designed an entirely local artificial intelligence (AI) methodology using offline speech recognition and human-led verification. The specific details of this project, and the operational pressures that motivated this workflow, and the second author’s role and its boundaries, are set out in section 2 below.The central argument advanced in this article is legal as much as methodological. I argue that local ‘processing’ of interview data is not merely a prudent risk mitigation choice but, in the absence of a vetted, institutionally governed processing arrangement, a necessary technical and organisational measure to satisfy ‘security of processing’ obligations under Article 32 of Regulation (EU) 2016/679 (UK GDPR). Specifically, I argue that consumer-facing commercial AI transcription services (whose terms of service routinely reserve rights to use uploaded content for model training) cannot straightforwardly satisfy the requisite appropriate safeguards the law requires. Methodologically speaking, transcription is an interpretive rather than administrative act, and the safeguards required by both law and qualitative rigour converge. I have developed this convergence into an applied framework, titled the ‘Reflexive Local AI’ (RLAI) framework. This framework comprises of five principles for conducting localised AI-assisted transcription in empirical legal research in a lawful and defensible way.To establish this working model, the article is structured as follows: section 2 below sets out the research project that sparked the RLAI framework; section 3 importantly examines why the use of AI changes qualitative research and data processing; section 4 analyses the data protection landscape under the existing legal regime; section 5 operationalises that legal analysis into the five core principles; and finally, section 6 evaluates the broader implications and future directions of the research. Technical hardware specifications and local execution steps are succinctly provided in appendix I. The research projectJust before the start of the 25/26 academic year, I had an email from the British Academy congratulating me on being awarded a British Academy/Leverhulme Small Research Grant for my project titled Pensions in a Warming World: An Empirical Study of the Opportunities and Challenges. I had been awarded near the full amount under the British Accounting & Finance Association’s (BAFA) ‘special fund’. This grant was designed to cover, inter alia, travel for in-person interviews with pension sector expert. Having achieved ethical approval, I thereafter conducted thirty-three semi-structured ‘elite’ interviews between October 2025 and July 2026. The participants included a cross-section of institutional actors, including pension trustees, specialised pension lawyers, academic commentators, investment consultants and senior policy makers. I had hoped that all interviews would be held in-person, but it was sometimes necessary to hold interviews online because of participants’ busy work schedules. I provided each participant with a standard information sheet and consent form and received their written consent. After the interview, interviewees were sent a debrief document, outlining what happens next.McDowell’s account of ‘using connections’ to secure access is instructive here. She highlights how fieldwork depends heavily on ‘luck and chance, connections and networks, and the particular circumstances at the time.’ Such spontaneity often yields unexpected results. While in London for my project, a chance conversation on the Underground with a fellow passenger (who conveniently worked in the pensions sector) after I took the wrong train led directly to new interview participants.Beyond such chance encounters, however, most access was built more deliberately. When I began the project, I did not have any contacts in the field. LinkedIn was imperative for finding connections and building a list of potential interview participants. I signed up to LinkedIn Premium so that I could gain fifteen ‘InMail’ credits each month, allowing me to contact anyone via the social media’s messaging system. As my sample ‘snowballed’, I found that I relied less on LinkedIn and more on email. I learned that, for targeting elite participants, LinkedIn is an important tool for gaining access, and I would recommend factoring the cost of the Premium account into any grant costings (so long as such a cost is permitted by the funder). Separately, the idea to use a localised AI-assisted transcription software (in my case, OpenAI’s Whisper) came when I was writing up my project proposal prior to submission to the British Academy. During the drafting of costings, I approached a transcription company and was quoted a price of approximately £6,000 for twenty interviews (or roughly £3.50 plus VAT per audio minute). Such a fee would have inevitably wiped out a large proportion of my proposed funding and prohibited the reach of my fieldwork. As a full-time academic, I was also sufficiently worried about the work involved with manual transcription; in fact, I felt as though Sisyphus had a less arduous task to contend with. When I discussed these concerns with the second author, a senior cloud engineer, he suggested using a localised AI to assist transcription. I had never considered such an idea but, so long as confidentiality was assured, I was willing to test the capacity of the new technology. We ran some preliminary tests before accepting the idea, and these worked well. The second author’s involvement was confined to this technical set-up stage: it occurred before the data were collected, and he has not been involved at any stage of data processing or seen any of the interview data collected in the study. When setting up Whisper, we used our own test samples of our own voices, which did not include the interview data. He appears as the co-author on this paper on the basis of his technical contribution and his input to some of the appendix I technical guidance. Once it was set up and tested on my machine, the use of Whisper was explained in my ethics application and throughout the project documentation.There are two examples during my fieldwork where I discussed the methodology directly to interviewees, and each illustrates a different dimension of building participant trust. First, more than once, online participants asked (without any apparent awareness of the confidentiality implications) why we were not simply recording the meeting directly on the video conferencing software. This question was a reasonable one; video conferencing now offers built-in recording services that are often used in business meetings. However, I explained to participants that this would potentially be a breach of confidence, since the audio file and transcript would be stored on the video conference companies’ cloud storage. One interviewee found it amusing that I was using my ‘old fashioned’, hand-held Sony recording device. I myself sensed the juxtaposition between using the traditional handheld recorder, while later deploying a GenAI model to transcribe and time-stamp the audio files (something that would have been unimaginable even when I was doing my doctoral research some ten years ago). However, the handheld recording device allowed me to save the data locally. Secondly, I specifically recall a participant asking me about the transcription process itself. I explained that the AI software would run entirely locally, which was a technical detail I found meaningful in building trust. Several participants (this one included) seemed reassured once told that their testimony would remain on my own hardware rather than passing through commercial servers. This encounter is a small but concrete instance of the ‘positional strategy’ identified below.Why AI changes qualitative methodology and vulnerabilitiesThis section establishes the methodological foundations upon which the article is built.Elite interviews and transcription as constructivist interpretationThe literature surrounding ‘elite interviews’ has grown over the last three decades. Empirical legal research frequently relies on elite semi-structured interviews. Harvey notes that there ‘is no clear-cut definition for the term “elite” and given its broad understanding across the social sciences, scholars have tended to adopt different approaches.’ Similar to my study, merchant and investment bankers based in the City of London, have been described as an ‘elite’ interview group. While the definition is elusive, I define ‘elite’ participants in this study by their institutional authority, specialist expertise or influence within pension governance. In particular, their professional visibility and membership of a relatively close-knit community create confidentiality challenges, as contextual details may make them identifiable even where names and institutional affiliations are removed from the transcripts.A transcript does not capture objective reality but is a selective representation of an acoustic event. Ochs showed the subjectivity involved in transcription when she famously observed in 1979: ‘A transcript that is too detailed is difficult to follow and assess. A more useful transcript is a more selective one.’ Thus transcription is ‘theory-laden’. The choices researchers make in producing the transcript enact the theoretical assumptions they hold and constrain the interpretations that can subsequently be drawn from the data. When speech recognition is delegated to AI, the researcher must remember that the need to perform interpretive construction does not disappear. Modern ASR systems do not ‘copy’ audio signals; instead, they process raw acoustic data through probabilistic language models that predict and output the most statistically likely sequence of words based on their training corpora. As Bender et al. highlight in their critique of natural language processing (NLP), large language models (LLMs) operate as ‘stochastic parrots’. They join together plausible-sounding sequences based on statistical co-occurrence without any underlying communicative intent, semantic understanding or contextual awareness. This is likely compounded by the so-called ‘black box problem’. Researchers must increasingly be open to the possibility that AI may lead to the emergence of ‘ethical risk’ in qualitative research. Parasuraman and Manzey’s foundational work on automation complacency shows that automation does not simply replace human judgment but reshapes it, often in ways the designers never anticipated. Both over-reliance and unwarranted disregard for automated output are common failures. It is because of the risk of automation bias and misuse/disuse that I wished to keep a ‘human-in-the-loop’ during data verification. However, I entirely agree with Moore, who suggests that automated transcription technology, while lacking certain capabilities and liable to being inaccurate at times, may be highly beneficial for ‘providing first-pass transcripts.’ Expanding upon Da SilvaBokhove and Downey recognised the challenges researchers face when utilising third-party automated captioning or transcription tools; but their paper did not prescribe a procedural ‘solution’ beyond acquiring ethics committee approval. Da Silva has drawn attention to the severe data security and confidentiality vulnerabilities introduced by commercial cloud-based transcription services. He observes that even a cursory review of the terms and conditions of common zero-cost cloud transcription services shows they cannot offer adequate assurances of confidentiality for interview data promised anonymity by the researcher. Attempting to bypass this cloud vulnerability through local device processing, Da Silva notes the severe practical barriers of proprietary hardware requirements, such as purchasing a dedicated Pixel device to run local recurrent neural networks, which still fail to offer a scalable methodological remedy for researchers. On Bokhove and Downey’s ethics committee solution, Da Silva goes further, concluding that their specific method would be unusable in a considerable number of cases involving genuinely confidential data. His response to these vulnerabilities, however, is grounded in risk management rather than statutory obligation; he does not ask whether, or under what conditions, local processing might be legally required rather than merely prudent. In this article, I address that critical gap. Vulnerabilities in the studyWhile the use of AI presents methodological risks across all qualitative disciplines, empirical legal scholarship is perhaps uniquely exposed. During my research study I identified three specific vulnerabilities, as follows:Vulnerability DescriptionSpecialised socio-legal terminologyLegal and financial discourse relies heavily on dense jargon, statutory references and acronyms (e.g. ‘TCFD’, ‘PSA26’, ‘fiduciary standards’). While overall accuracy was high, I found that the AI repeatedly misrecognised central domain terms such as ‘fiduciary’ (frequently rendering this word as ‘fed-user’, ‘federal’, ‘judiciary’, ‘FIDU3Duty’, and ‘5-2-3-2-2’), illustrating the necessity of domain-specific human verification. I found that the accuracy of the AI was dependent on the level of audible background noise and participants’ accents and speech patterns. This is something that Da Silva also found; he states that ‘Accuracy appeared to be reduced when transcribing a Scottish accent and improved when transcribing an American one.’ In my research, for instance, Whisper had to be set with an ‘English’ parameter – early use led to the audio being transcribed entirely in Welsh, because of my Welsh accent. Subtle contextual identifiers Elite interviewing testimonies often contain indirect or contextual identifiers. A participant may not explicitly state their name or institution, but may reference specific regulatory filings, unique fund sizes or localised corporate disputes. The RLAI framework operates on the basis that AI-assistance should be limited to a first-pass transcription only, with further scrutiny being undertaken by the researcher. Unique to my study, but nevertheless a good example, I had particular issues when interviewees mentioned their Task Force on Climate-related Financial Disclosures (TCFD) reports, which often contained fund-specific details, e.g., ‘Our approach changed significantly after our 2024 TCFD report, where we explicitly tied our net zero commitment to a phased divestment timeline, though we still hold a legacy equity stake in South Wales Oil Co. because of the income yield.’ To many in the elite group, the information in this statement is potentially revealing. This risk is not unique to legal research: McDowell, reflecting on her own interviews with City of London bankers, likewise concluded that her attempts at anonymisation might prove insufficient to disguise her respondents within banking’s ‘relatively small and close-knit world’.Power dynamics and trustConducting research with powerful institutional actors requires navigating complex access dynamics, and the trust built through that access must be sustained throughout the AI-assisted workflow. Assuring participants that their testimonies will be processed securely (example in section 2), without exposure to commercial cloud infrastructure, operates as a crucial positional strategy that preserves trust and protects participant confidentiality.Table SEQ Table \* ARABIC 1 - study vulnerabilitiesTreating an AI-generated transcript as a transparent, objective representation of an elite legal interview ignores the theoretical and operational realities. AI outputs must instead be understood as a highly structured, first pass textual rendering that demands rigorous human audit, contextual anonymisation and reflexive interrogation.The governance challengeThe legal governance of AI-assisted transcription begins with data protection law. Because interview recordings constitute personal data (and may also reveal ‘special category’ data) researchers must ensure that every stage of AI processing complies with the UK GDPR and the DPA 2018. These remain the core statutory framework, though the Data (Use and Access) Act 2025 (hereafter ‘DUAA 2025’) received Royal Assent on 19 June 2025, and its data protection provisions are now in force; where relevant, this article notes below whether specific provisions discussed have been affected. Data protection as a rightThe UK GDPR recognises the protection of personal data as a fundamental right, reflecting the continuing influence of Article 8 of the Charter of Fundamental Rights of the European Union. That right is not absolute but must be balanced against other legitimate interests through the principle of proportionality. For empirical researchers, this means that AI-assisted processing must be designed from the outset to comply with the data protection framework rather than treated as a purely technical or administrative matter. It is also worth noting, given the elite and institutional character of the sample discussed in section 2, that data protection compliance does not exhaust the relevant confidentiality analysis. Participants such as FCA-regulated professionals or pension trustees may owe, or be owed, sectoral or fiduciary confidentiality obligations that operate independently of, and in addition to, the UK GDPR.The cloud privacy trap under the data protection regimeThis analysis is directed at AI processing specifically, rather than institutional data storage as such. For my study, the audio files and transcripts were held on my university-managed cloud storage (‘OneDrive’), which operates under Swansea University’s own data processing agreement with Microsoft. This is a materially different arrangement from the consumer-facing AI platforms, since it carries no equivalent right to use file content for model training. And it is precisely this distinction, rather than cloud processing per se that the remainder of this section develops. A vetted, institutionally governed cloud arrangement of this kind is not what the necessity argument advanced here targets; it is, if anything, the confirming case, since the absence of a training data clause and the presence of proper institutional oversight are exactly what an unvetted commercial AI platform lacks. Indeed, reliance on consumer-facing commercial cloud AI platforms may create significant compliance risks under the UK GDPR. When a researcher uploads raw interview audio to these services, the data moves off local machines and onto servers controlled by commercial vendors. This transfer creates a legal conflict. Under Article 4 of the UK GDPR, the employing university acts as the ‘data controller’, while the cloud vendor operates as a ‘data processor’. The definition of ‘processor’ is as follows: ‘a natural or legal person, public authority, agency or other body which processes personal data on behalf of the controller.’ The UK GDPR further defines ‘controller’ as ‘the natural or legal person, public authority, agency or other body which, alone or jointly with others, determines the purposes and means of the processing of personal data.’ The ‘controller’ and ‘processor’ classification is not fixed, however. A vendor only remains a processor for as long as it processes data strictly on the controller’s instructions. Where a vendor’s terms of service require the right to use uploaded content to train or refine its own models, it is processing that data for a purpose of its own (model improvement) rather than solely on the researcher’s instructions. This most straightforwardly constitutes a breach of the Article 28 processor relationship, and, depending on the facts, may on ordinary principles of controllership also risk tipping the vendor from processor into a controller in its own right for that purpose. Either characterisation leaves the researcher exposed: an Article 28 breach they cannot readily detect or remedy, or a compliance problem more serious than a processor merely falling short of its own obligations. The critical point for present purposes is not which of these two characterisations is correct, but that consumer-facing platforms of this kind cannot offer the researcher confidence that neither has occurred. Moreover, Article 28(3)(a)-(h) requires binding contractual terms or other legal act under domestic law that limit processing strictly to the researcher’s instructions. In reality, consumer-facing or low-cost AI platforms routinely include terms of service reserving rights to use uploaded content to train and refine their proprietary models. Uploading unredacted interview files to train commercial models may be incompatible with Article 5(1)(b)’s ‘purpose limitation’ principle. This principle states that personal data shall be ‘collected (whether from the data subject or otherwise) for specified, explicit and legitimate purposes and not further processed by or on behalf of a controller in a manner that is incompatible with the purposes for which the controller collected the data (‘purpose limitation’).’Even where the relationship remains as a straightforward processor arrangement, a standard commercial terms of service rarely satisfies Article 28(3)’s substantive requirements. A terms of service is itself a contract, but Article 28(3) requires that contract to bind the processor to specific obligations. For instance, Article 28(3)(a) states that the processor should only process ‘the personal data on documented instructions from the controller’. Such a transfer under the terms of service must be read with Article 32 in mind. This deals with ‘security of processing’ and requires the controller and processor to implement ‘appropriate technical and organisational measures to ensure a level of security appropriate to the risk’. Consumer-facing and low cost AI platforms’ terms of service are drafted to protect the vendor’s commercial interests, not to discharge GDPR obligations.There is also the problem of cross-border data routing. The Information Commissioner’s Office (ICO) states that while UK GDPR applies to most organisations in the UK, it is important and necessary to govern international transfers. The ICO states: ‘People risk losing the protection of UK data protection law if their personal information is transferred outside the UK.’ Under the UK GDPR, routing audio files to servers outside the UK or European Economic Areas (e.g. the United States) typically triggers ‘Chapter V’ restrictions on international transfers, since a transfer to any third country (not only an international organisation) engages these provisions. The new rules relating to data transfers have been described as ‘a key milestone in the DUAA’s implementation’. Even following the amendments introduced by the DUAA 2025, relying on cloud-based AI infrastructure still requires navigating complex transfer risk assessments and contractual safeguards. Thus the need for localisation is not materially altered.Local deployment as a compliance-by-design responseThe compliance hazards identified above are not inherent to AI-assisted transcription as such; they are a function of where processing occurs. Running speech recognition entirely offline, on hardware under the researcher’s physical control, resolves each of the three problems identified above at source. There is no transfer to a third-party processor, so no route by which data could be absorbed into a training corpus, and concomitantly, no engagement of the Chapter V international transfer regime.This does not imply that every institutional cloud arrangement is equally deficient. Where a university holds an enterprise agreement with a commercial AI vendor that expressly excludes training data use and satisfies Article 28(3) in full, the analysis above would not apply in the same way. The point is narrower: consumer-facing platforms of the kind typically available to researchers without institutional procurement support cannot offer this assurance, and it is this class of arrangement, not cloud infrastructure as such, that the RLAI framework’s first principle (localisation) is designed to guard against.This reframes local deployment as more than a risk-mitigation tactic. It operationalises privacy by design, embedding safeguards into the processing architecture itself rather than relying on third-party contractual assurances.Lawful basis and the research exemptionCompliance is not exhausted by the question of where data are processed; it also depends on why processing is lawful. As a university researcher undertaking funded work, the applicable lawful basis is as a ‘public task’ under Article 6(1)(e), reflecting the institution’s function as a public body conducting research in the public interest. The lawfulness of research carried out under Article 6(1)(e) is a narrower, more institutionally anchored basis than the ‘legitimate interests’ ground more commonly relied upon by commercial actors.The Reflexive Local AI (RLAI) Framework This section develops the RLAI framework that I used. It forms five principles designed to better preserve methodological rigour, participant confidentiality and legal compliance.Figure SEQ Figure \* ARABIC 1 - RLAI modelPrinciple 1: Local processing as defaultData sovereignty should be the operational default, not an optional safeguard reserved for unusually sensitive projects. As demonstrated above, offline deployment resolves the Article 28 processor problem, the Article 5(1)(b) purpose-limitation risk and the Chapter V transfer regime simultaneously. It should be noted that this is not because local models are technically superior to commercial alternatives, but because researcher control over the processing environment is a methodological precondition, not merely a legal convenience.Principle 2: Human interpretive oversightNo AI-generated transcript should be treated as a final research artefact. The workflow instead proceeds through three stages: (i) an AI-generated pass transcript, (ii) systematic human verification against the original audio, and (iii) only then a finalised transcript suitable for coding. This principle directly addresses the risks of algorithmic deference identified earlier, i.e. the tendency to treat fluent, confidently formatted machine output as self-evidently accurate.Principle 3: Reflexive contextual anonymisationAutomated redaction should be understood as a preliminary fix, not a security shield. The researcher, not the model, determines what constitutes an identifier in a given field. As the TCFD vignette in table 1 illustrates, a phrase entirely unremarkable to a generic language model (e.g. a net zero target date, the size of a fund or a specific legacy asset holding) can function as a de facto identifier within a small, well-networked elite population. This judgement depends on field-specific knowledge no current model possesses and this cannot be delegated.Principle 4: Transparent documentationConsistent with transparency norms already expected in qualitative methods, researchers should document: (i) the model and version used, (ii) any prompts or configuration parameters applied, (iii) the verification process undertaken, (iv) the corrections made at the human-review stage, and (v) the reasoning behind anonymisation decisions. This creates an auditable record capable of withstanding both peer and regulatory scrutiny.Principle 5: Reflexive engagement in practiceResearchers should maintain a parallel reflexive log during processing, capturing what the transcript itself cannot: tonal shifts, hesitation and moments where the model’s smoothing of speech may have altered meaning. This principle operationalises resistance to algorithmic deference at the point of practice, distinct from the broader theoretical discussion of automation bias addressed below.The framework applied in practice: worked exampleThe interaction of the five principles is best illustrated by returning to the TCFD vignette introduced earlier. The relevant passage, in which a fictional participant described tying a net zero commitment to a phased divestment timeline while retaining a legacy equity stake, passed through each stage of the RLAI framework as follows.Principle 1 (local processing) meant that the audio never left the researcher’s machine. Whisper generated a first pass transcript entirely offline, so no fund-specific detail was ever exposed to a third-party processor or training corpus. Principle 2 (human interpretive oversight) then required verification of that pass against the original recording; this caught a domain-specific error consistent with the pattern reported earlier in this article, where ‘fiduciary’ was rendered as a phonetically adjacent but substantively wrong term, requiring correction before the transcript could be treated as usable. Principle 3 (reflexive contextual anonymisation) is particularly important in this example: the model itself had no reason to flag the net zero date or the legacy asset holding in South Wales Oil Co as identifying; this is because neither is a name, institution or postcode. Only field-specific knowledge (that a small number of pension schemes hold this particular legacy position, and that the combination of fund size, sector and timeline is close to unique) allowed the researcher to recognise the passage as a de facto identifier and redacted accordingly. Principle 4 (transparent documentation) required logging that redaction decision at the point it was made, together with the model version, the correction to ‘fiduciary’, and the reasoning for treating the divestment detail as identifying. Of course, this transparency was required so a second researcher, or indeed a reviewer, could audit the choice rather than take it on trust. Finally, Principle 5 (reflexive engagement) captured something none of the preceding four stages could: a note, made contemporaneously, on the participant’s tone when disclosing the holding. There, it was clear that there was a mixture of defensiveness and rehearsed justification that shaped how the passage was later coded, but which the transcript’s smoothed, grammatical prose gave no trace of.No single principle would have been sufficient alone. It is the sequencing and interaction of the five principles, not any one of them, that constitutes the RLAI framework’s contribution.Boundaries of the RLAI framework and directions for future research This section ends the work by looking at the limits of the framework and future research directions that may be taken.Limits of the RLAI FrameworkHonesty about the framework’s boundaries strengthens rather than weakens its contribution. Three limits are worth stating explicitly. First, local deployment carries a hardware floor. For instance, this project’s own experience of running Whisper’s higher parameter models required close to the practical limit of a 16GB unified memory of the system I was using. This is not to say that a researcher cannot use a standard, university laptop with, say, a minimum of 8GB RAM to run a transcription; but a lesser model could likely limit the accuracy of the outputted final transcript. Thus a framework partly motivated by protecting research budgets from prohibitive transcription costs should not be allowed to quietly reintroduce a different barrier: researchers without access to sufficiently capable hardware face a real constraint of their own. Relatedly, Principle 1’s legal benefit is not costless. As section 3.3 and appendix I note, local open-source models such as Whisper measurably underperform on regional accents and offer no out-of-the-box speaker diarisation, so the compliance gain of local processing must be weighed against a real, if manageable, accuracy and functionality cost relative to commercial alternatives. Second, this project’s empirical base is a single-researcher, single-project case study in one substantive field (pensions and climate governance). The RLAI Framework’s principles are designed to generalise, but their application has not yet been tested across other socio-legal subfields, multi-research teams or non-English language fieldwork, each of which may surface constraints this project did not encounter.Finally, the framework presupposes a level of technical competence (installing dependencies, configuring model parameters, troubleshooting command-line errors) that is not evenly distributed across legal (or, for that matter, non-legal) academia. This project benefited from the second author’s technical background at the pre-data collection setup stage; researchers without an equivalent resource may face a steeper learning curve than the ‘modest consumer-grade’ hardware framing above.Future research directionsThe framework may be extended beyond transcription and into analysis. Researchers have already tested AI in data analysis. For instance, Perkins and Roe undertook a comparison of manual thematic analysis against ChatGPT (GPT-4) assisted coding and found ‘a complementary relationship between GenAI and human researchers’. This article does not extend its five principles to coding or thematic analysis, but the underlying risk is the same one Principles 2 and 5 are designed to guard against. Given the RLAI framework promotes AI-assistance, it is important to look at the effect that using AI has on researchers. Thus Christou suggests that delegating interpretive tasks to generative AI risks undermining ‘reflexivity through cognitive offloading’. Their proposed corrective is a concept called ‘forward reflexivity’, which asks researchers not merely to audit AI output in the moment, but to anticipate how AI-augmented research texts may themselves circulate and return as training data for future models. And finally, the project’s local processing approach should connect the broader literature on ‘digital sovereignty’, which Roberts sees as a normative concept. ConclusionThe intersection of generative artificial intelligence and empirical legal research presents a profound methodological paradox. While automated transcription offers powerful solutions to qualitative bottlenecks, transcription remains an inherently constructivist act that actively shapes all subsequent legal analysis. Outsourcing this interpretive layer to opaque cloud infrastructure risks violating strict data protection mandates under the UK GDPR and DPA 2018, compromising elite participant confidentiality and introducing automation bias into the workflow.The RLAI Framework resolves these challenges by anchoring data sovereignty and human-led critique at the centre of empirical research. By enforcing local execution, multi-stage verification, contextual pseudonymisation, and transparent documentation, the framework proves scholars need not choose between methodological rigour and technical efficiency. Ultimately, preserving the integrity of legal scholarship in the AI era requires technology to serve as a secure instrument that reinforces, rather than replaces, critical human judgment. In this sense, the RLAI framework is offered not merely as a transcription protocol, but as a template for the governance of local AI in empirical legal research more broadly.Appendix I: Technical guidance on using localised AI modelsThis appendix sets out the technical workflow underpinning the study.HardwareThe transcription processing was performed locally on an Apple Mac mini (M4, 2024; 10-core CPU, 10-core GPU, 16GB unified memory). This represented a modest, consumer-grade baseline. This hardware configuration comfortably accommodates Whisper’s full range of model sizes; as mentioned within the software, memory pressure remained within practical limits of the unified memory utilised.The 16GB unified memory ceiling constrained language model selection. That said, this baseline configuration successfully demonstrates that modest consumer-grade hardware can achieve optimal results for local transcription workflows. While larger parameter models could theoretically offer enhanced context awareness, the M4 Mac mini set up proves sufficient for robust, secure execution under the RLAI framework. It is important to note that when running Whisper, foreground applications were closed to maximise available memory during transcription; background system services and processes managed by macOS itself, which cannot be user-terminated, were permitted to continue running as normalSoftware and model selectionWhisperOpenAI’s Whisper ASR model was used because it is open source and provides the ability for transcribing audio in various formats, e.g. ‘.wav’ and ‘.mp3’. It is also compatible with the recording tools used, as mentioned below. The model has been compiled from a large dataset of audio recordings, which allows specifically for speech translation and language identification, making this an appropriate tool for transcribing interview recordings.Whisper segments the transcription logs which provides further details and information on the timestamp of the chats; it can even provide granular results of timestamps per word if needed (captured as part of a ‘VTT file’ – a text-based subtitle/caption file). However, the use case for this research was to ensure timestamps per sentences and this is recommended for qualitative interview transcripts.OpenAI Whisper provides six model sizes to work with. The models can be English-only or multilingual; the model sizes range from ‘tiny’ to ‘turbo’. Whisper documentation recommends using English-only models which offers speed and accuracy (measured by transcribing English speech on a A100 GPU), but is limited only up to medium.en (769m parameters) up to ~2x the speed. The multilingual models allow up to turbo (809m parameters) at a much faster processing speed of ~8x. This latter model was selected for use with the Mac mini as a result of its speed of operations. Whilst a multilingual model was used, English can be configured as a way for transcribing as part of the tool. More information on the performance and how these models have performed against different languages can be found on their GitHub repository OpenAI/whisper. There is no known telemetry which is collected from using or running Whisper.A limitation of Whisper is the inability of knowing which person said what sentences, which is called diarisation. Nonetheless the ability to diarise the transcript can be accomplished by using pyannote (a Python library) and other open-source audio speaker diarisation models. However, this workflow focussed primarily on transcription and timestamping, since the context of the transcriptions were more important.This guidance does not instruct on how to download and install Whisper (instructions can be found in the official GitHub webpage). Once installed, here is an example command in a ‘terminal’ of how to run the whisper toolset with an example .mp3 audio file. The below example takes a .mp3 audio file and setting to an English-only model, ensuring the language for transcribing is set to English. To add timestamps to the transcription the --word_timestamps parameter set to True adds this to the output. Setting the output format will only provide the formats set, otherwise if not, then all associating files including the .vtt (for text-based subtitling) will be available. The below command, once complete, will output the transcription and create two new files. One which output the plain English transcribed text (from the –output_format flag): interview1.txt, and the other including the time stamped sentences from the redirected output, interview1_timestamp.txt.whisper interview1.mp3 --model turbo --language en --task transcribe --output_format txt --word_timestamps True > interview1_timestamp.txtPythonPython allows libraries like Whisper to be used with greater flexibility than the command-line interface alone, providing more advanced control over processing and output. The specific scripts are not included here as they are Whisper’s core functionality wrapped in a bespoke interface rather than a generalisable tool, but the approach can readily be reconstructed using the Whisper documentation cited below. We created our own set of Python scripts which handled the running of Whisper, using segmentation of sentences to build and output .txt files (with timestamps) into organised folders with clear transcribed files.Recording and file handlingInterviews were recorded using a handheld Sony IC Recorder (model: ICD-PX370). The device has no wireless or cellular connectivity, meaning the audio file could not be transmitted or synchronised to any cloud service at the point of capture, consistent with Principle 1 of the RLAI framework (local processing as default).The recorder’s built-in USB-A connector allowed audio files to be transferred directly to the researcher’s local machine. Transcription was then performed entirely locally, using Whisper (as described above). Audio files and completed transcripts were stored on the researcher’s Swansea University’s password protected OneDrive account, governed by the university’s institutional data processing agreement with Microsoft. As already established, this is distinct from the consumer facing AI platforms discussed in the article. Microsoft’s enterprise terms for institutional OneDrive accounts do not include the training-data clauses identified as the central Article 5(1)(b) concern in this article, and the university, not Microsoft, remains the data controller. As a backup, the ‘British Academy/Leverhulme Project’ OneDrive folder was regularly backed-up onto a USB-C password-protected flash drive and stored in a locked desk drawer.The pseudonymised Microsoft word documents containing the transcripts were password protected. When a participant’s transcript was sent via email, the transcript would be attached and sent in one email and the password in another, separate email. However, in most cases, the password was sent separately to participants using LinkedIn’s private messenger service.Verification process Consistent with Principle 2 of the RLAI framework (human interpretive oversight), no output generated by Whisper was treated as a finished transcript. Each recording produced an AI-generated first pass, which was then checked in full against the original audio. During transcription and pseudonymisation, the computer’s WiFi was turned off, to ensure that the work remained entirely local.The verification process required listening to the entirety of each recording alongside the corresponding AI-generated text, correcting errors as they arose. As discussed in section 3.2 above, the most consistent category of error concerned domain-specific socio-legal terminology, e.g. the term ‘fiduciary’. As the transcripts were well done, a rough estimate of how long the verification for a one-hour interview generally took would be around three hours. This time would have been shorter, but I liked to be thorough, and the transcript would be checked against the audio on ‘normal’ speed first and then later on the ‘fast’ speed setting.By correcting misrecognised words, this stage also served the contextual pseudonymisation described in Principle 3: identifying and redacting details that, while not explicit identifiers such as names or institutions, could reasonably function as identifiers within a small, well-networked elite population (see the worked example above). Redaction decisions of this kind were recorded in a documentation log, consistent with Principle 4, together with the model version used and the corrections made at this stage. Only once this verification and pseudonymisation was complete was a transcript treated as finalised and suitable for coding and analysis. Lecturer in Law, HRC School of Law, Swansea University. Email: lloyd.a.brown@swansea.ac.uk. ORCID: 0000-0003-0945-7472. Incoming Lecturer in Law at the School of Law and Politics, Cardiff University (December 2026). This research was generously supported by a British Academy / Leverhulme Small Research Grant under the British Accounting & Finance Association (BAFA) special fund, for the project ‘Pensions in a Warming World: An Empirical Study of the Opportunities and Challenges’ (grant awarded October 2025, ref number: SRG/250023). The authors would also like to thank Dr Paul Dargue and the anonymous reviewers for their helpful comments and suggestions on an earlier draft of this article. Senior Cloud Engineer. Joseph Da Silva, ‘Producing “good enough” automated transcripts securely: Extending Bokhove and Downey (2018) to address security concerns’ (2021) Methodological Innovations 1-11, 1-2. ibid 1. See also Regulation (EU) 2016/679 (UK GDPR), rec 6, which states: ‘Rapid technological developments and globalisation have brought new challenges for the protection of personal data. The scale of collection and sharing of personal data has increased significantly.’. ibid. UK GDPR, art 4(2): ‘processing’ means ‘any operation or set of operations which is performed on personal data or on sets of personal data, whether or not by automated means, such as collection, recording, organisation, structuring, storage, adaptation or alteration, retrieval, consultation, use, disclosure by transmission, dissemination or otherwise making available, alignment or combination, restriction, erasure or destruction.’ UK GDPR, art 32. Elinor Ochs, ‘Transcription as Theory’ in Elinor Ochs and Bambi B Schieffelin (eds), Developmental Pragmatics (New York: Academic Press 1979) 167-168.‘British Academy/Leverhulme Small Research Grants’ (BA, no date) <https://www.thebritishacademy.ac.uk/funding/schemes/ba-leverhulme-small-research-grants/> accessed 1 August 2026. UK GDPR, art 4(11): ‘consent’ is defined as ‘any freely given, specific, informed and unambiguous indication of the data subject’s wishes by which he or she, by a statement or by a clear affirmative action, signifies agreement to the processing of personal data relating to him or her.’ Linda McDowell, ‘Elites in the city of London: some methodological considerations’ (1998) 30(12) Environment and Planning A 2133-2146, 2135. P Biernacki and D Waldorf, ‘Snowball Sampling – Problems and Techniques of Chain Referral Sampling’ (1981) 10 Sociological Methods & Research 141-163. William S Harvey, ‘Strategies for conducting elite interviews’ (2011) 11(4) Qualitative Research 431-441, 432. ibid. McDowell (n 9). Ochs (n 6) 168. ibid. Judith C Lapadat and Anne C Lindsay, ‘Transcription in Research and Practice: From Standardization of Technique to Interpretive Positionings’ (1999) 5(1) Qualitative Inquiry 64-86, 77. ibid 66. Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey and Ilya Sutskever, ‘Robust Speech Recognition via Large-Scale Weak Supervision’ (2022) arXiv <https://cdn.openai.com/papers/whisper.pdf> accessed 1 August 2026. Emily M Bender, Timnit Gebru, Angelina McMillan-Major and Shmargaret Shmitchell, ‘On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?’ (2021) Proc FAccT 610, 616. ibid 617. Kyle ML Jones, ‘Generative AI in Qualitative research and Related Transparency Problems: A Novel Heuristic for Disclosing Uses of AI’ (2025) 24 International Journal of Qualitative Methods 1-14. Matthew Kosinski, ‘What is black box artificial intelligence (AI)?’ (IBM, 29 October 2024) <https://www.ibm.com/think/topics/black-box-ai> accessed 1 August 2026. Kyungmee Lee, ‘Becoming Ethical Qualitative Researchers in the Era of Artificial Intelligence: Michel Foucault’s Ethics of the Self’ (2026) 32(2) Qualitative Inquiry (forthcoming/online first). Raja Parasuraman and Dietrich H Manzey, ‘Complacency and Bias in Human Use of Automation: An Attentional Integration’ (2010) 53(3) Human Factors: The Journal of the Human Factors and Ergonomics Society 381-410. ibid 381. Robert J Moore, ‘Automated Transcription and Conversation Analysis’ (2015) 48(3) Research on Language and Social Interaction 253-270, 253. Christian Bokhove and Christopher Downey, ‘Automated generation of “good enough” transcripts as a first step to transcription of audio-recorded data’ (2018) Methodological Innovations 1-14, 12. Da Silva (n 1). ibid 2. ibid 3. ibid. It is noteworthy that most of my interview recordings were, on average, one hour long and the transcribed texts often spanned 10,000+ words. The longest interview transcribed was 91 minutes long and constituted 16,000+ words, or 48 pages. Overall, I was impressed by the AI’s speed in transcribing the audio: as a rough estimate, one minute of audio would be transcribed in one minute by the model. Despite the size of these transcripts and the speed in which it performed its work, I would only identify one or two minor mistakes in the AI’s transcription per recording. One recording was a little more challenging as the research interview took place in an area with sufficient background noise; but even so, the AI model still performed well. Da Silva (n 1) 7. Da Silva usefully shows (in Tables 4 and 5) the differences in time taken to transcribe five interviews using a fully manual method vis-à-vis a semi-automated method. The semi-automated method took hours less to transcribe audio with similar recording lengths. I cannot provide such a comparison, since I only used a semi-automated method in my study. From looking at Table 5 in Da Silva’s paper (the one showing the semi-automated method), the transcription times seem similar to the time taken in my study. ‘Task Force on Climate-related Financial Disclosures’ (TCFD, no date) <https://www.fsb-tcfd.org> accessed 1 August 2026. This is a fictitious example and is not linked to any person/institution. McDowell (n 9) 2135-2137. Susan A Ostrander, ‘Surely You’re Not in This Just to Be Helpful: Access, Rapport, and Interviews in Three Studies of Elites’ (1993) 22(1) Journal of Contemporary Ethnography 7. ‘Data protection’ (gov.uk, no date) <https://www.gov.uk/data-protection> accessed 1 August 2026. ibid. ibid. Information Commissioner’s Office, ‘The Data Use and Access Act 2025 (DUAA) – what does it mean for organisations?’ (ico, 19 June 2026) <https://ico.org.uk/about-the-ico/what-we-do/legislation-we-cover/data-use-and-access-act-2025/the-data-use-and-access-act-2025-what-does-it-mean-for-organisations/> accessed 1 August 2026. Data (Use and Access) Act 2025 (‘DUAA 2025’). UK GDPR, rec (1). ibid, rec (1), citing Charter of Fundamental Rights of the European Union, art 8(1) and the Treaty on the Functioning of the European Union (TFEU), art 16(1). ibid, rec (4). ibid, arts 4(7)-(8). ibid, 4(8). ibid, 4(7). ibid, art 28(3)(a). Da Silva (n 1) 2. ibid, art 5(1)(b). ibid. Da Silva (n 1) 2. Here, Da Silva states: ‘It was apparent from even a cursory reading of the terms and conditions of the predominant cloud-computing transcription offerings, for example Google… that these zero-cost offerings could not provide adequate assurances around the confidentiality of the data I was seeking to transcribe, and would not satisfy the ethical requirements of the research that I was performing, the anonymity of which I had agreed upfront with the participants.’ UK GDPR, art 28(3). ibid, art 28(3)(a). ibid, art 32. ibid, art 32(1). Da Silva (n 1) 2. Information Commissioner’s Office, ‘What is an international transfer of personal information?’ (ico, no date) <https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/international-transfers/a-guide-to-international-transfers/what-is-an-international-transfer-of-personal-information/> accessed 1 August 2026. ibid UK GDPR, art 4(26) defines an ‘international organisation’ as an organisation governed by public international law or established by an agreement between two or more countries. AI companies do not fall within this definition. However, Chapter V also applies to transfers to a ‘third country’, which encompasses transfers to overseas commercial providers. ibid, art 44A(1)(a)-(b). Amanda Leiu, ‘The Data (Use and Access) Act 2025: Key Implications for Businesses as the New International Transfer Regime Takes Effect (Part 1)’ (Burges Salmon, 6 February 2026) <https://www.burges-salmon.com/articles/102mh1d/the-data-use-and-access-act-2025-key-implications-for-businesses-as-the-new-in/> accessed 1 August 2026. See also Information Commissioner’s Office, ‘Are we making a restricted transfer’ (ico, no date) <https://ico.org.uk/for-organisations/uk-gdpr-guidance-and-resources/international-transfers/a-guide-to-international-transfers/are-we-making-a-restricted-transfer/> accessed 1 August 2026. DUAA 2025, s 85. See also ‘Data (Use and Access) Act 2025: data protection and privacy changes’ (gov.uk, 27 June 2025) <https://www.gov.uk/guidance/data-use-and-access-act-2025-data-protection-and-privacy-changes> accessed 1 August 2026; Alan Baker, ‘Data (Use and Access) Act 2025: five key changes for businesses’ (Farrer & Co, 5 May 2026) <https://www.farrer.co.uk/news-and-insights/data-use-and-access-act-2025-five-key-changes-for-businesses/> accessed 1 August 2026. Alasdair Dougan and Sophie Burton-Jones, ‘What are the new rules on international data transfers?’ (Mills & Reeve, 5 February 2026) <https://www.mills-reeve.com/publications/what-are-the-new-rules-on-international-data-transfers/> accessed 1 August 2026. Nothing in this article should be taken as suggesting that the use of local AI transcription either automatically requires or dispenses with the need for a Data Protection Impact Assessment (DPIA). Under Article 35 UK GDPR, the requirement for a DPIA is determined by whether the proposed processing is ‘likely to result in a high risk to the rights and freedoms of natural persons.’ Researchers should therefore follow their institution’s governance procedures and consult their Data Protection Officer where appropriate before commencing processing. ibid, arts 44-46. ibid, art 6(1)(e). It is noteworthy that the present project did not involve the processing of ‘special category personal data’ within the meaning of Article 9 UK GDPR. Accordingly, no additional Article 9 condition was required. Nevertheless the governance principles developed in this article may be of particular significance for empirical legal research that does involve special category data, where researchers must also identify an appropriate Article 9(2) condition (for example, Article 9 (2)(j), as supplemented by Schedule 1 of the DPA 2018), in addition to satisfying the obligations discussed in the main text. Where special category data are inadvertently disclosed by a participant despite not being sought, the RLAI framework’s Principle 3 (reflexive contextual anonymisation) extends to redacting content at this stage. ibid, art 6(1)(f). Mike Perkins and Jasper Roe, ‘The use of Generative AI in qualitative analysis: Inductive thematic analysis with ChatGPT’ (2024) 7(1) Journal of Applied Learning & Teaching 390. Prokopis A Christou, ‘Reconfiguring Reflexivity in the Era of AI: From “Turning Back” to “Looking Forward” Through Constructivist and Posthumanist Lenses’ (2026) Qualitative Inquiry 1-10. ibid. Huw Roberts, ‘Digital sovereignty and artificial intelligence: a normative approach’ (2024) 26 Ethics and Information Technology 70. See also House of Commons Library, Digital sovereignty (CBP 10547, 2026) <https://commonslibrary.parliament.uk/research-briefings/cbp-10547/> accessed 1 August 2026. OpenAI, ‘Whisper’ (GitHub, 2022) <https://github.com/openai/whisper> accessed 1 August 2026. ibid. ibid. ibid. ibid. ibid. ibid. ibid. ibid. ibid. ISSN 2752-3403PAGE \* MERGEFORMAT53350837540386000Journal of Legal Research Methodology (JLRM) ISSN 2752-3403Brown & Brown - 1983ISSN 2752-3403PAGE \* MERGEFORMAT1358140015240000JLRMBrown & Brown - article numberiVBORw0KGgoAAAANSUhEUgAAALIAAAA+CAYAAACP+EHLAAAAAXNSR0IArs4c6QAAAARnQU1BAACx jwv8YQUAAAAJcEhZcwAAIdUAACHVAQSctJ0AAA4ZSURBVHhe7Z0LcA/XHscp8QyS0BIRIQiXeAxp PHIbEdG693JFI0iCVFOu4HoTVakkghCPS6UiIWmoVLyZTl/T6dRop0zfrdeUqXY8Wlqvlj5xbr5/ u3/n7P72/9/NQ/+79jfzmTG75/z3nOObs79zzu+crSHZuHJOlXO7HGZjYwKgVWg2uRyHPVXOnaZN m7KEMQksKTnJxsbjgVb9/PxuQbvljIeQT/n6+bIdB3ba2JgOH18fuWeucXtU0igykY2NpzMqabTs ZtRgiWMTyUQ2Np5O4rhE2We+f0Iu279DEyq9zT3kdspdt5IFtQmS//NUNGrUiM2eP+eBadf7IuSN JQXsucyFLCIygjXxaSI0uBJvb2/Wo1cPNjNtFnuhYL0tbglZvIGtA8l2c4VXHS+WtjCNbd9XRv62 FahWIadOm8yC27cTGtUozVs0Z/EJ8ax0zyvkMx4E8os3su49upPtY4QWLVo4OhTqGWanWoT89KQU 1rBhQ6ERK0vt2rVZZHSkpXsVivSs51mtWrVU7VGzZk02YcIE9vnnn7MzZ86w77//nv3www/s22+/ ZSdPnmQrV65kjRs3VuUDAwcNJJ9lZqpUyAVbCpl3I2+h0dzh5eXl+E+h7mmRnbuEfL6VgCsxYfJE Vd1btmzJLl68yIxYRESEqo1D/tbRUm+5KhPys4sWCA1FERYWxjIzM9nbb7/Nvv76a3br1i2pqRn7 7rvv2Hvvvcc2bNjAYmJiWN26dcnfkHksKpJt211KlsUKPJM6QagveuW9e/dKrWXczp49y9q1E928 jp06WuYNVyVChg8rN44S9CAHDhxgP//8s9Sk+gwi/+CDD1h0dDT5uyCgVQDbsmMrWSYzk744Xahn gwYNHC6Dlh08eJCFhISwwMBAlpuby+7cuSPdEe3XX39l/fv3F347IvLvZBnMRqWEjNdfj549hIbh ef3114Vet6J2+fJl1qxZM/IZderUIctmVgpKCthDDz3krF/9+vVddgIQrdJtuHTpknRXbUiv7BzG pYwjy2ImKiXk+NF0Tzxw4EB2+/ZtqemqzgoKCsiBD/zsbbvM72YoOwb8kX711VdS7Wn78MMPhbYA K1askO7Shs4FPbicHs8xu5tWYSHPnDfL2RA88+bNk5qreuzQoUNCjyUT3C7Y9HPOK19YJdRp165d Uq21DR0GZnT4fHiDuTOkgcsi5/nH0H+SZTILFRJy7tqVZM+4f/9+qZmq13788UdWr1491fOHj3yS LK9ZeOSRR5x1adKkiVRb93bq1CkWGhrK/P392auvvqrpIystIyPD+Tz8f27eVkSWywwYFjKmbOrW U88oYN7yftq5c+fIabv1hXlkuc0AX4+bN29KNaUNYv3jjz/Y77//ToJ7egTNz/cvyHiOLJcZMCzk 4fFPOisuExsbq7sXqEp77bXXVG4GYgyocns6/PQl3jbu7PDhw0K9KVJTU6XU2jZ06FBn+rDej5Jl MwOGhIw5R7nSMpib/O2336Rmuf+2YIF6/nr+88+S5fdk/Jr6Ocufnp4u1U7bcnJyhDpT9OvXT0qt bd98842Q55W928nyeTqGhBzeN1yoNMCo+a80vAmUZcLrkiq/J8O/Wb744gupdtpWVUKG8XlKTTp7 oVvImBGQKyuDCXhPMAxwlGXLWb2crIenwvv76CXdWXUJeevOl8nyeTq6hTx9znShwuDGjRtSU2gb VpNOnz7NTpw4wa5cuSJddW9YUj169KhjUOduUQVTUD4+PkLZ+kT0JevhqfBl1zN9Vl1C3lJmzpVS 3UJu10Fcp/fz85OagTYsqSJYBYsVfL7mzZuzqVOnSqlEw2h72rRpDBtg+TzorbAahUgvLeOnkgCC l6h6eCp82a9duybVStuqS8glZVvI8nk6uoVcv0F9ocJ5eXlSM6gNK3B8WgqMzH/66Scpx92em5+g p4CgEVRE2aeffqpKb6apON61wNvIndmuhYguIa8vWC9UFmgFsbzzzjuqtOiVETyEqTH+elBQkHO+ k4p2Q6/PLxLIYAFAaXA/EJfAp5u3MI2sjyfCD/aOHz8u1Urb1q1b53i7UQtDAPeGDBkipXZtfD6z LlXrEnJWTpZQWTQ65bfiGnYh8GkHDBjgnNyHaJOTk5334EagV87OzhbyQNToYWVD8JF8r0+fPuzI kSPSHdEiIyOF30mdlkrWxxPh/5ALCwulGrm3MWPGCHWW0WsY5/D5zBqjrEvIM+bOECrbqlUrqRlE u3DhgpAOPbFyhQpih3D5AQ16az7fm2++Kd25Z/jP/fLLL10uvMTHi0FMCC+l6uOJPPXMeGe50Zvq tcoKefbs2c48bYLbmjZeRZeQk1Pu9aJAy/eC/8qnCw4Olu5oG4SpXGq+fv26dNeYTZkyRfid/tFR ZH08Fb7seqMHKytkX19fZ56pM/9LlssM6BLy0OH/dlYWjB49WmoG0TAbwafr2bOndEfbMCXH5wGY vaiILV++XPidzqGdyfp4KryPDxdKj1VGyLt37xby5G16kSyXGdAl5GFxsUKFR44cKTWFaDNmiC5I t27dpDvahqkmPg+o6JL30qVLhd8J7RZK1sdTWZgl7gzBJlJ3VlEhw8ULCAhwpu8Z1pMsk1nQJeTx E+/5b6Bv375Sc4hWVFQkpGvTpo10R9vgWihDQq9evSrdNWZK1yIqxlyuBfzT1kGtneXHa/+XX36R akdbRYWckJAgpC/cuoksk1nQJeRZ8+8NCIDWYA8bSPl0AFvUeYNw0VNv3LjR2fPi9/g8uKe0t956 y7HzBNt4tAZ8ysHeyISRZH08mSW5S4UxA+KM//zzT6mGaquIkNeuXSukfbR3OFkWM6FLyNkrxOkx I9Nvbdu2ZZ999plDfBBhSkqK8x5OFcKZDJgT5fOA0tJSx9Qc/OV9+/YJ9xDxRpmZp994xk98WqhH r169NN0to0KeO3eukA4nF1FlMBu6hLyhKF+oPIAAKcOWfmVaoNyOA9DDyqZclpZRzmgAKsYDc9TK xYG09PlkfczAE/8aLNQFi0NUMBEOZsECkRKl4Q8hKipK+E0c4GKFvY5Al5DhuykPXsG2cy3bs2eP kJYCmx/5KSY0NP6zqLQyeBPgPAzKPv74Y1V6/AFS9TEDaPMBMeJuZ/xRDx48mHwbujJsRlWOQxo1 bmTauAoKXUIGnTp3EhoCMb+uDMLEQSvYxi/3qoilgIB37twppVLbqlWrHMH6crARevKHH37Y4ZK4 GvggGJ0vH/a8UfUwG5jDp95KmNrEeSGIy8C8OzoFuG9YgEIvjR0kI0aMUOUDHTp2IJ9lZnQL+bms hUJjAAzu9Bga2N3omzLk0RrY8YYeShnH0T+6P1kPM7JoSQbz03C9ZCB2ZaShEtxPSh5DPsPs6BYy kBtEBrt2PcGoaLs1L/6PrINZwTazuQvmkWMNPQyPH24pV0KJISHHjhguNA544403JDn9NYZXqrJM ZotFNgqOCXss6jHHwSrwfZWuB8YStb1qs5COIWzthnXkb1gNQ0LGAETZaAj44eOK77clJSUJ5QGY i6XKb2WKSotZQUmhaYN+KoshIYP/TJ0kiAaEh4e7nLSvLqPmn1sGtCTLbWNtDAsZvpp/gL8gHhAX FyfJ6/4YTuqkRvP5RRvJcttYG8NCBi9tLyF3dEyfPl13+GFlDHv3lM8GKZNSyPLaWB9ByL379WaT p09RsWjJIlXG1XlrBBHJYLXO3XFPlTFM7lPPDQsPU5UxZ81ysj421gPalbSgFocMRsYbitWrZPiw CpUee/GMbP3XY1iCHjZsGPm8gMBWqkEONlE29K7a75jYmALyohP/lv6CUADEM2WmGEzPg+XUqjAM 6vjNmTztQ9qTI/WYJ2LI9DaWh7woMHv+bJVgQOayTJcT9YhWO3bsmKEYAQTIbN682eXXoWLjYsny 4NhbKr3NAwF5UUVGuWgp8eCjkFqfw5LBVh7sLlm2bBnbunUre/fddx0nsiMuAPEXq1evZpMmTVKF glLMWTCHLEfhlkIyvc0DA3lRRW0vL/aixt4unImgjNiqagICA1j+S/QUG7ayN/Q29nk0s4A30yef fCLw0UcfsbKyMkeoANIMGjTIeU+ZH8FYuI6VWOU9i0FeJMHgb9OWzaSYAFaYfHzFs9gqC046QvAM 9Tzw8q5tmh/MsQJ428mGgS/gXbXu3e9+GVX+5gi/sVTOiwAs5dELFoS86JI1eWtIUcngW9I4TFDr NBx3wO8O7d7VsSmT+n2Zom3Fqug3q8ELWW5PLAi9//77jmv43giuIUwWgsV8fqdOd0Nv5Q9M5uer N0hYEPKiS9CQOLyFEhcPVgOXrVrGxj49joX3CWdBbYNUx1sBxCt37daVjUoaxTJzsnSdQ5a9YkmF I8LMBCVkUFxc7LjGb7zFxgYYtpglJiY6/o0Nva4GzhaCvKiLLl27GP6CJqbMcDo6XAL41hU5KT36 8er1xz0JXsg40en8+fOOPZDoeXE+tHK5XjZ5pXXs2LHCfQtDXjREVs5iw4I2Cv4A1uavI3t0K8ML GfslseULu0PkYK3x48UjG7B4JBu1u93CkBcN07RZU5a2MI1cpKgscCOC2gSRz7U6Wq5Fly5dHD4x Qmn59Oih5V052PbF37M45MUKg8NFYp4YVOlvt2GXb9yoOMfBftRzHhS0hIzgLbgPmMXg0wN5t7ny usUhL1YJ6B2weXV+ur6vLS1enu0IAlL6fQ8yvJBLSkocpzrt2LHDeU4e3A1lHlvINh4HL2TesFsd sdlULIotZBuPBOd+8Cg/AKQE7p27s0IsCHnRxsZskBdtbMxGjdOKCzY2JqPG6f8DFSC8q9ghQnIA AAAASUVORK5CYII=iVBORw0KGgoAAAANSUhEUgAAA5oAAAA+CAYAAABKkGTuAAAAAXNSR0IArs4c6QAAAERlWElmTU0A KgAAAAgAAYdpAAQAAAABAAAAGgAAAAAAA6ABAAMAAAABAAEAAKACAAQAAAABAAADmqADAAQAAAAB AAAAPgAAAAAogilEAAAUfXRFWHRteGZpbGUAJTNDbXhmaWxlJTIwaG9zdCUzRCUyMmFwcC5kaWFn cmFtcy5uZXQlMjIlMjBzY2FsZSUzRCUyMjElMjIlMjBib3JkZXIlM0QlMjIwJTIyJTNFJTBBJTIw JTIwJTNDZGlhZ3JhbSUyMG5hbWUlM0QlMjJQYWdlLTElMjIlMjBpZCUzRCUyMlY0VmZ2T2tQaDE2 REozQkZvTVNRJTIyJTNFJTBBJTIwJTIwJTIwJTIwJTNDbXhHcmFwaE1vZGVsJTIwZHglM0QlMjIx NDE0JTIyJTIwZHklM0QlMjI3MTAlMjIlMjBncmlkJTNEJTIyMSUyMiUyMGdyaWRTaXplJTNEJTIy MTAlMjIlMjBndWlkZXMlM0QlMjIxJTIyJTIwdG9vbHRpcHMlM0QlMjIxJTIyJTIwY29ubmVjdCUz RCUyMjElMjIlMjBhcnJvd3MlM0QlMjIxJTIyJTIwZm9sZCUzRCUyMjElMjIlMjBwYWdlJTNEJTIy MSUyMiUyMHBhZ2VTY2FsZSUzRCUyMjElMjIlMjBwYWdlV2lkdGglM0QlMjI4MjclMjIlMjBwYWdl SGVpZ2h0JTNEJTIyMTE2OSUyMiUyMG1hdGglM0QlMjIwJTIyJTIwc2hhZG93JTNEJTIyMCUyMiUz RSUwQSUyMCUyMCUyMCUyMCUyMCUyMCUzQ3Jvb3QlM0UlMEElMjAlMjAlMjAlMjAlMjAlMjAlMjAl MjAlM0NteENlbGwlMjBpZCUzRCUyMjAlMjIlMjAlMkYlM0UlMEElMjAlMjAlMjAlMjAlMjAlMjAl MjAlMjAlM0NteENlbGwlMjBpZCUzRCUyMjElMjIlMjBwYXJlbnQlM0QlMjIwJTIyJTIwJTJGJTNF JTBBJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTNDbXhDZWxsJTIwaWQlM0QlMjJvdWlkUk1wdnBQ Z1lMb1Foa0RwWS0zJTIyJTIwZWRnZSUzRCUyMjElMjIlMjBwYXJlbnQlM0QlMjIxJTIyJTIwc291 cmNlJTNEJTIyb3VpZFJNcHZwUGdZTG9RaGtEcFktMSUyMiUyMHN0eWxlJTNEJTIyZWRnZVN0eWxl JTNEb3J0aG9nb25hbEVkZ2VTdHlsZSUzQnJvdW5kZWQlM0QwJTNCb3J0aG9nb25hbExvb3AlM0Qx JTNCamV0dHlTaXplJTNEYXV0byUzQmh0bWwlM0QxJTNCc3Ryb2tlV2lkdGglM0QzJTNCJTIyJTIw dGFyZ2V0JTNEJTIyb3VpZFJNcHZwUGdZTG9RaGtEcFktMiUyMiUyMHZhbHVlJTNEJTIyJTIyJTNF JTBBJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTNDbXhHZW9tZXRyeSUyMHJlbGF0aXZl JTNEJTIyMSUyMiUyMGFzJTNEJTIyZ2VvbWV0cnklMjIlM0UlMEElMjAlMjAlMjAlMjAlMjAlMjAl MjAlMjAlMjAlMjAlMjAlMjAlM0NBcnJheSUyMGFzJTNEJTIycG9pbnRzJTIyJTNFJTBBJTIwJTIw JTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTNDbXhQb2ludCUyMHglM0QlMjIz MjAlMjIlMjB5JTNEJTIyMTIwJTIyJTIwJTJGJTNFJTBBJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIw JTIwJTIwJTIwJTIwJTNDJTJGQXJyYXklM0UlMEElMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAl MjAlM0MlMkZteEdlb21ldHJ5JTNFJTBBJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTNDJTJGbXhD ZWxsJTNFJTBBJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTNDbXhDZWxsJTIwaWQlM0QlMjJvdWlk Uk1wdnBQZ1lMb1Foa0RwWS0xJTIyJTIwcGFyZW50JTNEJTIyMSUyMiUyMHN0eWxlJTNEJTIycm91 bmRlZCUzRDElM0J3aGl0ZVNwYWNlJTNEd3JhcCUzQmh0bWwlM0QxJTNCZmlsbENvbG9yJTNEJTIz ZGFlOGZjJTNCc3Ryb2tlQ29sb3IlM0QlMjM2YzhlYmYlM0IlMjIlMjB2YWx1ZSUzRCUyMkxvY2Fs JTIwcHJvY2Vzc2luZyUyMiUyMHZlcnRleCUzRCUyMjElMjIlM0UlMEElMjAlMjAlMjAlMjAlMjAl MjAlMjAlMjAlMjAlMjAlM0NteEdlb21ldHJ5JTIwaGVpZ2h0JTNEJTIyNjAlMjIlMjB3aWR0aCUz RCUyMjEyMCUyMiUyMHglM0QlMjI0MCUyMiUyMHklM0QlMjI5MCUyMiUyMGFzJTNEJTIyZ2VvbWV0 cnklMjIlMjAlMkYlM0UlMEElMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAlM0MlMkZteENlbGwlM0Ul MEElMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAlM0NteENlbGwlMjBpZCUzRCUyMm91aWRSTXB2cFBn WUxvUWhrRHBZLTUlMjIlMjBlZGdlJTNEJTIyMSUyMiUyMHBhcmVudCUzRCUyMjElMjIlMjBzb3Vy Y2UlM0QlMjJvdWlkUk1wdnBQZ1lMb1Foa0RwWS0yJTIyJTIwc3R5bGUlM0QlMjJlZGdlU3R5bGUl M0RvcnRob2dvbmFsRWRnZVN0eWxlJTNCcm91bmRlZCUzRDAlM0JvcnRob2dvbmFsTG9vcCUzRDEl M0JqZXR0eVNpemUlM0RhdXRvJTNCaHRtbCUzRDElM0JzdHJva2VXaWR0aCUzRDMlM0IlMjIlMjB0 YXJnZXQlM0QlMjJvdWlkUk1wdnBQZ1lMb1Foa0RwWS00JTIyJTIwdmFsdWUlM0QlMjIlMjIlM0Ul MEElMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAlM0NteEdlb21ldHJ5JTIwcmVsYXRpdmUl M0QlMjIxJTIyJTIwYXMlM0QlMjJnZW9tZXRyeSUyMiUyMCUyRiUzRSUwQSUyMCUyMCUyMCUyMCUy MCUyMCUyMCUyMCUzQyUyRm14Q2VsbCUzRSUwQSUyMCUyMCUyMCUyMCUyMCUyMCUyMCUyMCUzQ214 Q2VsbCUyMGlkJTNEJTIyb3VpZFJNcHZwUGdZTG9RaGtEcFktMiUyMiUyMHBhcmVudCUzRCUyMjEl MjIlMjBzdHlsZSUzRCUyMnJvdW5kZWQlM0QxJTNCd2hpdGVTcGFjZSUzRHdyYXAlM0JodG1sJTNE MSUzQmZpbGxDb2xvciUzRCUyM2RhZThmYyUzQnN0cm9rZUNvbG9yJTNEJTIzNmM4ZWJmJTNCJTIy JTIwdmFsdWUlM0QlMjJIdW1hbiUyMHZlcmlmaWNhdGlvbiUyMiUyMHZlcnRleCUzRCUyMjElMjIl M0UlMEElMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAlM0NteEdlb21ldHJ5JTIwaGVpZ2h0 JTNEJTIyNjAlMjIlMjB3aWR0aCUzRCUyMjEyMCUyMiUyMHglM0QlMjIyNDAlMjIlMjB5JTNEJTIy OTAlMjIlMjBhcyUzRCUyMmdlb21ldHJ5JTIyJTIwJTJGJTNFJTBBJTIwJTIwJTIwJTIwJTIwJTIw JTIwJTIwJTNDJTJGbXhDZWxsJTNFJTBBJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTNDbXhDZWxs JTIwaWQlM0QlMjJvdWlkUk1wdnBQZ1lMb1Foa0RwWS03JTIyJTIwZWRnZSUzRCUyMjElMjIlMjBw YXJlbnQlM0QlMjIxJTIyJTIwc291cmNlJTNEJTIyb3VpZFJNcHZwUGdZTG9RaGtEcFktNCUyMiUy MHN0eWxlJTNEJTIyZWRnZVN0eWxlJTNEb3J0aG9nb25hbEVkZ2VTdHlsZSUzQnJvdW5kZWQlM0Qw JTNCb3J0aG9nb25hbExvb3AlM0QxJTNCamV0dHlTaXplJTNEYXV0byUzQmh0bWwlM0QxJTNCc3Ry b2tlV2lkdGglM0QzJTNCJTIyJTIwdGFyZ2V0JTNEJTIyb3VpZFJNcHZwUGdZTG9RaGtEcFktNiUy MiUyMHZhbHVlJTNEJTIyJTIyJTNFJTBBJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTND bXhHZW9tZXRyeSUyMHJlbGF0aXZlJTNEJTIyMSUyMiUyMGFzJTNEJTIyZ2VvbWV0cnklMjIlMjAl MkYlM0UlMEElMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAlM0MlMkZteENlbGwlM0UlMEElMjAlMjAl MjAlMjAlMjAlMjAlMjAlMjAlM0NteENlbGwlMjBpZCUzRCUyMm91aWRSTXB2cFBnWUxvUWhrRHBZ LTQlMjIlMjBwYXJlbnQlM0QlMjIxJTIyJTIwc3R5bGUlM0QlMjJyb3VuZGVkJTNEMSUzQndoaXRl U3BhY2UlM0R3cmFwJTNCaHRtbCUzRDElM0JmaWxsQ29sb3IlM0QlMjNkYWU4ZmMlM0JzdHJva2VD b2xvciUzRCUyMzZjOGViZiUzQiUyMiUyMHZhbHVlJTNEJTIyQ29udGV4dHVhbCUyMGFub255bWlz YXRpb24lMjIlMjB2ZXJ0ZXglM0QlMjIxJTIyJTNFJTBBJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIw JTIwJTIwJTNDbXhHZW9tZXRyeSUyMGhlaWdodCUzRCUyMjYwJTIyJTIwd2lkdGglM0QlMjIxMjAl MjIlMjB4JTNEJTIyNDQwJTIyJTIweSUzRCUyMjkwJTIyJTIwYXMlM0QlMjJnZW9tZXRyeSUyMiUy MCUyRiUzRSUwQSUyMCUyMCUyMCUyMCUyMCUyMCUyMCUyMCUzQyUyRm14Q2VsbCUzRSUwQSUyMCUy MCUyMCUyMCUyMCUyMCUyMCUyMCUzQ214Q2VsbCUyMGlkJTNEJTIyb3VpZFJNcHZwUGdZTG9RaGtE cFktOSUyMiUyMGVkZ2UlM0QlMjIxJTIyJTIwcGFyZW50JTNEJTIyMSUyMiUyMHNvdXJjZSUzRCUy Mm91aWRSTXB2cFBnWUxvUWhrRHBZLTYlMjIlMjBzdHlsZSUzRCUyMmVkZ2VTdHlsZSUzRG9ydGhv Z29uYWxFZGdlU3R5bGUlM0Jyb3VuZGVkJTNEMCUzQm9ydGhvZ29uYWxMb29wJTNEMSUzQmpldHR5 U2l6ZSUzRGF1dG8lM0JodG1sJTNEMSUzQnN0cm9rZVdpZHRoJTNEMyUzQiUyMiUyMHRhcmdldCUz RCUyMm91aWRSTXB2cFBnWUxvUWhrRHBZLTglMjIlMjB2YWx1ZSUzRCUyMiUyMiUzRSUwQSUyMCUy MCUyMCUyMCUyMCUyMCUyMCUyMCUyMCUyMCUzQ214R2VvbWV0cnklMjByZWxhdGl2ZSUzRCUyMjEl MjIlMjBhcyUzRCUyMmdlb21ldHJ5JTIyJTIwJTJGJTNFJTBBJTIwJTIwJTIwJTIwJTIwJTIwJTIw JTIwJTNDJTJGbXhDZWxsJTNFJTBBJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTNDbXhDZWxsJTIw aWQlM0QlMjJvdWlkUk1wdnBQZ1lMb1Foa0RwWS02JTIyJTIwcGFyZW50JTNEJTIyMSUyMiUyMHN0 eWxlJTNEJTIycm91bmRlZCUzRDElM0J3aGl0ZVNwYWNlJTNEd3JhcCUzQmh0bWwlM0QxJTNCZmls bENvbG9yJTNEJTIzZGFlOGZjJTNCc3Ryb2tlQ29sb3IlM0QlMjM2YzhlYmYlM0IlMjIlMjB2YWx1 ZSUzRCUyMlRyYW5zcGFyZW50JTIwZG9jdW1lbnRhdGlvbiUyMiUyMHZlcnRleCUzRCUyMjElMjIl M0UlMEElMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAlMjAlM0NteEdlb21ldHJ5JTIwaGVpZ2h0 JTNEJTIyNjAlMjIlMjB3aWR0aCUzRCUyMjEyMCUyMiUyMHglM0QlMjI2NDAlMjIlMjB5JTNEJTIy OTAlMjIlMjBhcyUzRCUyMmdlb21ldHJ5JTIyJTIwJTJGJTNFJTBBJTIwJTIwJTIwJTIwJTIwJTIw JTIwJTIwJTNDJTJGbXhDZWxsJTNFJTBBJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTNDbXhDZWxs JTIwaWQlM0QlMjJvdWlkUk1wdnBQZ1lMb1Foa0RwWS04JTIyJTIwcGFyZW50JTNEJTIyMSUyMiUy MHN0eWxlJTNEJTIycm91bmRlZCUzRDElM0J3aGl0ZVNwYWNlJTNEd3JhcCUzQmh0bWwlM0QxJTNC ZmlsbENvbG9yJTNEJTIzZGFlOGZjJTNCc3Ryb2tlQ29sb3IlM0QlMjM2YzhlYmYlM0IlMjIlMjB2 YWx1ZSUzRCUyMlJlZmxleGl2ZSUyMGVuZ2FnZW1lbnQlMjIlMjB2ZXJ0ZXglM0QlMjIxJTIyJTNF JTBBJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTIwJTNDbXhHZW9tZXRyeSUyMGhlaWdodCUz RCUyMjYwJTIyJTIwd2lkdGglM0QlMjIxMjAlMjIlMjB4JTNEJTIyODQwJTIyJTIweSUzRCUyMjkw JTIyJTIwYXMlM0QlMjJnZW9tZXRyeSUyMiUyMCUyRiUzRSUwQSUyMCUyMCUyMCUyMCUyMCUyMCUy MCUyMCUzQyUyRm14Q2VsbCUzRSUwQSUyMCUyMCUyMCUyMCUyMCUyMCUzQyUyRnJvb3QlM0UlMEEl MjAlMjAlMjAlMjAlM0MlMkZteEdyYXBoTW9kZWwlM0UlMEElMjAlMjAlM0MlMkZkaWFncmFtJTNF JTBBJTNDJTJGbXhmaWxlJTNFJTBB9UUzOwAAK7NJREFUeAHtXQeY1FQXvVvoXaT3DiogvStdBKmi iHTpdSnS2wIiRdoiTZCiiBQBBVSQ3kT4EURUpLel91532f+dt2TMDpPszCQzk9m99/syk0ley3k5 k9z37rs3gBxISNiG7IEBieolShRUOSLiWfHHTyKyR0VRkIOkfMiHCAQEUGSihMFng4MD9z9+HLn1 WdTj1WEhNc76sEnxomrmh390M/PDN/3E/PAN7q7WyvxwFTFz0jM/zMHR06UwPzyNsOPymR+OcbHa UVf4EaBufIcv1iRN8TT1yGfPojrny5EuUca0yYPSv5SC0qZORkGBMZKqs/G+jxCIfBZF12/dpys3 7tKl6/cij525+jgwMGDm3QS3hs3uWPeBj5oVZ6tlfvhX1zI/vNtfzA/v4m20NuaHUQRdy8/8cA0v X6dmfni3B5gf3sXbaG2u8MOmPYZM2ZghKCjh9izpU2WrViZ/kqSJExhtB+f3MgIPHj2lTXuOPjx/ 5XZ4ZOSTN8J6Vr/s5SbE2eqYH/7ftcwPz/Uh88Nz2HqrZOaH55BmfngOW2+VzPzwHNLMD89h662S 9fghFc3Q0GUJ776UcW/RgpkLlS2SkzVMb/WMh+rZffD00z8PX/g3xY1LpUJD33/ioWriTbHMj7jV 1cwPc/uT+WEunr4ujflhbg8wP8zF09elMT/M7QHmh7l4+ro0R/wIRKNupck4KkO6FHlZyfR1F5lT P/oR/Yl+NafE+F0K8yNu9T/zw9z+ZH6Yi6evS2N+mNsDzA9z8fR1acwPc3uA+WEunr4uzRE/ArtM 3JZNLOrsVqNsgaS+biDXbx4C6E/0K/rXvFLjX0nMj7jZ58wPc/qV+WEOjlYrhflhTo8wP8zB0Wql MD/M6RHmhzk4Wq0Ue34EJkoQ0P3V3BmCkyVJaLW2cnsMIID+RL+ifw0UE++zMj/i5i3A/DCnX5kf 5uBotVKYH+b0CPPDHBytVgrzw5weYX6Yg6PVSrHnh3BSGtA4f870rGVaradMaA/6Ff1rQlHxtgjm R9zteuaH8b5lfhjH0KolMD+M9wzzwziGVi2B+WG8Z5gfxjG0aglqfgQ+jYjMljY1W81atbOMtAv9 iv41UkZ8z8v8iLt3APPDeN8yP4xjaNUSmB/Ge4b5YRxDq5bA/DDeM8wP4xhatQQ1PwKfRUUFJwgO smpbuV0GEEC/on8pKsoWxsZAcfEvq8CN+RF3u535YbBvmR8GAbR2duaHwf5hfhgE0NrZmR8G+4f5 YRBAa2dX80N6nbV2c7l1hhEICIgyXEZ8LIBxix+9zv3sXj8zbu7h5m+5uJ/d6zHGzT3c/C0X97N7 Pca4uYebv+US/cyKpr91GreXEWAEGAFGgBFgBBgBRoARYAQYAYsjwIqmxTuIm8cIMAKMACPACDAC jAAjwAgwAoyAvyHAiqa/9Ri3lxFgBBgBRoARYAQYAUaAEWAEGAGLI8CKpsU7iJvHCDACjAAjwAgw AowAI8AIMAKMgL8hwIqmv/UYt5cRYAQYAUaAEWAEGAFGgBFgBBgBiyPAiqbFO4ibxwgwAowAI8AI MAKMACPACDACjIC/IcCKpr/1GLeXEWAEGAFGgBFgBBgBRoARYAQYAYsjwIqmxTuIm8cIMAKMACOg j8CjRw/p6OF/6OaNa/oJ+SwjwAgwAowAI+ABBJ48eUwPH9z3QMn+XSQrmv7df9x6RoARYATiLQJ3 79ymXp2aUaXXc1C/7q3kd7umdeja1ctuYxIVFUUL506jiKdP3S5j26a1dPzov27n/+evP6hGuYJu 5+eM8Q+B8DMn6ZVsSeRWKGtiyp8x2Pb7ncpF4ywgp08cpU2/rImz18cX5j0Eiud7iQrnSE5Fc6eS W/nCWahD83p04fzZWBuxZcNPVKFINvpp1TIaP3IATRk3PNY8WgkqFctBR//9W+u03x1nRdPvuowb zAgwAowAIwAEenZsSgEBAbTjwBn6YcPvtPfIVUqeIiV93LWl2wA9e/aMRg3uSRiddleWLZpLB/bt djc752MEXEYgW47cdCj8odwWrtgoeaD8/nHrny6X5y8Z/j64nxYvmOUvzeV2WhyBhSs30Z8nb8tt 897jdO/eHRo5sEesrf7lx5XUtfdgaty0TaxpY0uw9McdlCtv/tiS+c15VjT9pqu4oYwAI8AIMAIK Apj1+9+u7TR09BRKnDiJPJwkSVIa/MlkKv9GNYLCiG3y2GFUoWg2qvh6dpoxeTRhxhKCmc+Na1dR o5qlqUqpvPTNvOnyeNc278rvpvXfpIcPH0hTqMG9O9AbxXJSi0bV6OAfe+X5Q38doMZvl6Mrly7I 398v+1rMqraWs6F7ft1K0yd+QhvXrabZ08bTsm++lGnw8cXn4whpIciLGVm0rU2TWvTbjs3yOH8w AmYjgPsbW58uLSQPUP7Uz0bQWxVepXrVStCssDG2KrW4gQTzv5hCdd4sIjkzd8ZEmWffnl9p9NDe NDa0r+RJy3er0/nw0/IcPrTq6dyqIW34+Qdq9FYZOYMDs0NHXDty6C8aENJWcqlq6XzUsGYpOvzP QTp5/DBNGjOU9v/+G40cFGKrj3cYATMQwHOldr336Mrli7bili6cQ29Xek1uX3/5uTz+3bfzaOvG n8V//3RatXyRLS12tO5precH8mCg8+L5cOrbrRWt+3EFDknBfuiAbnIfnGjWoIrkW/8eH9G9u3ee p7LeFyua1usTbpG5CCQWxaU2t0gujRGIMwj4LT+OCdOiHLnyUJqXXo7RGenSZ6QO3fpRYGAgfbtg Ju3Y8gstX7uLFq7YQKtXfGtT8o4c/ptmTR1LY8PmUbsuvemTIb2kYjlEKKqQ8VMXSAV2aN/OdPPm dZq/bC190LIDtW1am27fukGvFH6dMmXJRmPEy/WN61fp02EfU+MPP6J6735IRYqVog9adaByFavQ xXPhMV5U8AJx9colWcfCeTMofYZMYjZ2LzV4rwWNGNjdpgjLBPzhawT8lh/2wF2+eIHCxodKvtSu /z7hRRcm3tPnL6ew2d/SIjErCOUNosUNKHYrFs+neUvW0mfTFtBXcz4nmK7eEnyAuXmKlKlo8Zpt lDNPfjnAg7L06sG66kljhlCjJi0pS7YcpMW1B0IBXbVikRiYuUjzFv9EmbPmEOUPJczitmzXjQq9 WpQ6hwxAdSzeRSDO8EOBDev9oRxi+/vPfbRyyVdUteY78jQGDmdNHUeDRk4SnFkini+z5DOlZp2G VKJMBWr4fguqUqOOUpT81rqntZ4fyIR6Hz18KJ5veWnNysW28lZ99w3lK/iqVCpbNq5B1WvVo8Wr t1KChAmprxjktKqwomnVnuF2mYXAI1HQt2ILFRsrnAIEFkZAhYDf8uPu3duUXLzY6smyRfOoWZvO lClzNsqVpwA1/6gL/fj9EluWXgNGUf5Cr4k0XehloaCeOXlcKI/Z5flsOXLRnTu36KcfltLgkRMp T75CVKdBE/lSu3XjWplm6CdTaPvmddSpZQN6651GVLrcG5Qq9UuUJGkySvtyekqWPIWtLkc7VarX ppB+oTJt/oKviRmaI/Tg/j1HSfmYbxDwW344gqvw6yUJAyk5c+cTSmFKmjTzG8qbvxC9lDYdZciY mf4+uM+WzRE3oOhdv3ZVDKxco5JlKtKyn3dK3iAT+NOl12DKIpTAgaGf0c6tG6RFQWz1fNSpl+Bl V4qIjNDlWooUqWjgiAlSiW3buQ+dEgpuggQJ6eV0GQiWDOnEgA2L1xGIU/wAepiNV9ZoYqb9spjN bCEGMyCL5s+g5uJ5UqlKTfncaNW+B61d/R2lSpVG/teDRylT/feaefv2Td172tHzQ1b0/KNOg/fp 120b5TIOWNfsFpYyb9dtLOvE86mN4E6WbDmpz6BP5KARlGMrSrAVG8VtYgRMRiBMlLdObLCtwf4U sd0SGwsjwAhEc8Lv+CHXpB38Qz6EEyZMZOtHOPFZLmZdGopZEpjvFS1exnYOeTD7qEh68XKtSNJk yQmj2WoJP3NKviy/W6uc7TAe5pWFgghB/vfELCbMCT//cpktjd5OZGSk7TRerju3biRnffKLkWoW SyIQZ54frxYpYQMYZoHTJn1C63/6nlKIl+Mnj6Az/CeOuFFWzNA3ELM2HzaoLJVTzOBAuYTkyJlH rpfGPgZaYFGAgZNUomy9epQ26XEtuyj75XTpKSgoCMVTMnBVzPiwWAKBOMMPoLlk9XbxzCgtgb10 8Rw1rVeZZk8dT32HjqEzp07QXwd+pznTJ8jzERFPKXfeAnLf0YfePY30sT0/MDgKqx0omI8ePKDi pcpJ3qEdcBZU9tX/BleixDIRWN6Ae1YTntG0Wo9wezyBwC+i0N/EhqEmuAK7KbZQsf039CR+sLiP wJ3brLe7j57Pc/olPzAT8/jxI/rj990xANyxdT3BGU+iRIkpX4FX6NKFc7bzePEtVuI/pTEwMMB2 ztFOCuFYKCg4mLYIpxC7/7kot+3C8VDT1p1kciit3y9bKF825s2KNrl9oRzhrEhZF4pzF8+dlUmg cPbu3JwaCpPZPYcu0byl0bOkL+TnA75GwC/5ERtoM8PGChPwm7Tht8O0/tdDVOCVwjGyOOIGvDz3 6Dtc3q/Dx0yV69E2CXNCyP37d235r1+7Ij0/p0ufgWKrB868ILFxDYoriyURiFP8CBSDGfjPx4bZ wg9athehs6I9wGJ2ftinU23Pgl8PnqM5i9Zodkps97Qzzw+YuW9Z/5NY77+K3mn4gawL7ahQuYat HXg2bdt3SlruaDbGhycMMRfTyHCO4GmZOHowTfx0iKer8Vj5eBHCgnsWnyIwwq52KJynxBYqNsMK JxZiw528vRtsLBxv8k5FUUXcEfX9jD9KmJf0D/lIru+BowZ3RR0S4tNhfQy5B3e3DfE4n0f54Qlc s2bPJT38Ya0W7kMIRqDHj+xPTVt1lL+xXmbV8m+kQgplT7qgf7O6PKf1gRdfvNTCVAl1ZMiYxbZO 5uKFcKr9RhE6dfyozA4HKG9Wf5tmLlhJS76eLUe7cSIoMIgeP58dzZw1m/RAC6UYM6xwXAK5I8yq bt24TlVq1pH1Yf0Ni2UR8Dt+xIYk1ltWEFyAl+ZzZ0/R/r27YgyIOMq/bs1yGjGgOwUHJ6A3qtai 14qWkOszkfbY4UN07MghmQ33cp58BaUZubP1xMY1R+3BMSgE9pYIWmn5uMcQiHP8UJDCEghlOUO5 ilXpZxG+BP/l2OCgau5MjQFGUUBs97TW80OpG99wRrRl/Y/06/ZNVOPtBvJUmQpVCA7nlPdNrONs LnQxZdBGnd8K+4YUTStcgD+04dUixeir5Rj0YfEhAugAe1sbKJimKpw+vD6vVa2+n+Fa/unTJ/JF O0/+grToB/e9ZqpDQsBNeJtOPb12TVwR+SU/+g0bK9c3It7Zm8VzEQY6KgvlErOEkPrvNZeOFaqX LUC1Kr4mnftUFOtr9ARKZsXKNaUnzifiZWLMlC9pghjsjPY0WFiu0YEjh+3CyRA8DfYbOjZ63ViX PjS4T0eKjIiQL/CIo/bLTyvprTqN6Igwc6pSMg91ER5tYX4IgROjRh+0ovrVSlL96iXo37//JKyh Gzuin17z+JxvEPBLfuhBhXWRU4RH5vfrVKCPhXdLOKOaOWWMGAC5pZmtllgftnf3DqomeIZ88Jr8 jli3DMmSPSeFtG8iB9XhfXP0xNnyuLP1wCxWi2uyII2PIoIzJ44dpoE922mk4MNeQCDO8UPBLPVL aeWMJtZbdu45kG4J89Sqwkt55RK55X6Hbn2VpC98693Tes8PdUFY7pFWmI0XK1lWDgrh3OslylCL tt2oRtmC8rmEybhPJ89RZ7PUfkC3SVuienxYya1GYUZz3ZoVhLgz9nJVLKDFQ/f3PTulVj941CQq U/5NmQwP5zHD+0qvTjVqN5C2z1gvAKcKsz8fT7BrLlWuEiEPHsYAUajqcsGrup5pE0dRmrQvy2nl f//5U2r+A4aPlyNciqt6zL5gZA3raLTac/nieRrYq50cjS4t2ti9zzAq+GoR2T54ItyxZb20k+4r XijgTRCCNTnLhUvjB8JuGouD24qXDK3j8KwG1+FYeD9fmFclTJyY9uzcKke2S5auQOM+ny/NvPCn PaxfFzp98ph0OoEF96079JCep2ThbnxM/XYHTetdJcCNrHExy1viorAWTUvwhI2xhtNZfmBGE8F+ t+47SZmfOxNBJZjRXLn0K1r6407pmj21cBTyfvPoByLCHMDjZMP3W8pQCzCL+EZwCjMv/YaNkwu+ EYga99yE6V9Lhwe4R+DlEg97mAXCu2a5SlUJrq5xT+bOV0DMrMwRI8mpaczkufI+Vl8s3GVXE57K ar3zrjwMd9m7d26h0LHTZBkjB/WQ/EOZCBuB0W6MSE+fNFr+uSFW1LAxn8v7uffAUdS+WV0xi3Re LlDv2KO/5OrUL5cSzAInCSuEFUsWSG+CcMjSqEkrWSdc3f/0wzJKJHhQu15j6hQyUHoshEdEOHxA eIoL587I632vWVvS+i9x9prV12+/z/yIgYjH+BGjFg/8gBdYeNXMkTuv/C9VV4F78cypY5Q8eUq5 JkZ9Tm8fL9yKYwfE1MQsZvqMmV7wcqtVBkbBE4nnGl42EGblmvA0q173puTD/Z1ErKsB19DWu8IB Ueo0aZXTPv1mfsSA32/5EeMqVD8wEwhrAOWZBZNXzODoCQYWT504JtaKvSwd8SAtnlNfzQ6jBd+t l7P2mMlRz664Uo87XMO6bLQLPPKmMD9ioB3n+BHj6p7/wPvZ2dMn5C94hXVG3LmnnSn35o1r8rmH QX44xrKaKPzw2IxmiAikXaBQYdry+wmxWHwQdfvoPenS/bx4gQzp0JT6Dx8n3cX/dWAvIRYNOm+c GMmFwrZy/R6JF1xn6wkUMrzMdujejxYs+0V4OVtPW4W7boi9K2+t9iAtzuXOW5BWb95POXPlo0FC 6YRouSXWcvGtdfz+vbv0j3BaAbko1guhzXUaNqHJsxbRnyIm29rVy+U5KLVw2z1p5iLxsn9Suua/ f88UD4RRogLe9JVM9IHhGc7FC76QCiVi52HDqJUiemEO4E5+yTdzxODKZBE2oRi1fq+mnH2ZvXC1 eME9Is01UI5WOAQt9+9K3cq3O+6yHwqnC3iRQLwzcE25nzNmykrwGIi1cj36DRcDMw+E18L9sioM Qm3fso6WrNlOIf1HUGj/buJl/7imq3v7kBDnzp6m88/Xsmlx19lrVq5d55u5Ef3/oDcIA/gM80On DwydgqdXeI/Fukx7gaKH/3dHSp59WvVvRcnEMTgbwho2DHw6K3AuhLohmCXVqh/eMqFkQpDeKkqm bFD0B/PDz/mh6ssYuxjgV5RMnIhNyUQavNDCcRW8vdoL7nPMwKiVTKRxpR53uBacIIHXlUzVtTM/ 4ig/VH1s28W9jfcoZ5VMZHTnnrZVqLOD5xEmxayoZKqbHaz+YdY+FMD9/9tFeEnGAxSzJwu/nCbN jDCqi2DaSlwaLKw9F35a2tiPGD9DuszGyBQ68cC+mE4eHLUPs0FwKQ/p2L2/rKPaW3Xlb8WVt157 SgkX3QfErOdXYiQOLynd+w6jpMmSyThpcGu/cfcRuSAYru0xOwW39nADbu/iG6PliH3j6PjNmzdk e5SPt4XNNcypIDXeri89s8Hc6g9xvVv3npAzspjVQcw3Fp8goLxQt3a19uNHD1Fi4WpdETgiCRZr SJyRrr2GSPOIZ2JWA7Gb4HQBL6sw4zstRpAhCIeAPxaM3NqHQ1Dcv+NlFe7fB4h1k/YCd9lfiiDb GGHD7Am8mY2eNDuGu2zkgbtsBLlX3GVjPQJmKvESrDhfwcMdL8l4iYDSeevGf/f58sULqH3XvtKN PlzpjxTcfipGnRVX9+A3ZqEUl/rKNdmHhNDjbt78r8gZULi817tmewz4t2EE3OaH4Zq5AEbA+gjE O37kzptfenm2ftdwCy2AQLzjhwUw92kTnHsDdrGJmI2Ay19llBbZpVt5YZZxLvwM5RVKmyJYTI4N M5pwGzy4d0cRjPS2NFFCTJrYJKcwlVIkq4grM2f6Z8pPUtxmx9YemHkoI+F4ae7ae4icmYG5kyO3 9vUbN3Po4lvP9betUWIHL9eKQGF4JJxOYGYTbsCxsB2CuDy5RNBjFp8gcEvUqpjQwkOt0zL007AY I8SK6ayjAqDoqSVj5qzyJ4LvZs6aXSqZOIB74qlwow3RC4fgjPt3d9xlo96swvuaKzMtcHqSR8x0 KgKX+BCYCeq5ulfSK9963CWhaDpzzUpZ/G0aAm7zw7QWcEGMgHURiHf8wHMFGwsj4AQC8Y4fTmAS p5N4RNHE2rGrYuYSihpMKSCnThyRI16YQj4s1pMpghmLa1cvS0VzzrTPpKkdYib9IFzGqwNrK+nt v2HGp8iRf/8Si2ZfNOfQaw+CnmJdAl76MSsCW/+fhFcpLLbFCz7c2it2/1gMjPYrLr57DRgp17eN GtxTmglD0cQslP1xBDJWi71ZCc6h7sti/Z0aM2BokvAazWggXV5DYBL+0cUIkwsMqCiCMAe4/xRx 5E5eOYdv3KMIh/DxoNE0f+k6uSalcI7ktiQK12wHNHYUd9n37915wV02vGcqgvWgmLG8KTxjuioY aIJSScIMGIKgw1g/OmfaBJtLfQxEdWxRX7doPe4io7PXrFuJWAEey/n4ctq3/IgvKPvfdTI/ovvM kvzAOwveVRy9V/jfreaXLWZ+WJgffnlHxbFGG16jiRkWmMCpN6xFyZIth/S4B7zgEv70yeMy2Cjc zf+2Y7NYf3VGvjhjXSIUzZPCaxjM56BkwowUyl6UXFaoj/gvItgw0sPxAhyVlBSmsPaSIVMWzfZg fULW7DlF3hUy2/fLvpauhPXcEmu5+NY6bt8eR7+xzg1KyPqfv7e1Q8/7m6My+FisCESD+2IyjLCN EFsusYWKDb9NF60wB85WZFY4BG+4y4ZpPFxuw0QXCme/Hm0oZcrU0rGQlkt9dUgIBRM97ipp+Ns0 BHzKD9OuwosFqUP9uFMth/RxBzWf5bEkP5rUrUR/7v+fz0Bxp2K86yycO00O7Ovlt09XqVgOGahe Lw+f8xkCluSHz9CwUMX2PPJ20wzPaO7b8yuVKhjTS9mOP85Ql56DCR4u4W31yqWL1K3PELlgNacw B0XcsFoiriRmS+BgAbFh7onZlbnCI2vDmqWE0vhAuH1vSQu+CBNBSqODAWsBk0ysYatRXgTuFt7T YKrXXsPVsFZ7UG6vAaPkTNGkMUPF7GVSGj91gZxhhKvtHu2aSG+ecOLTUThCgVt7zELNmjpOuvhO lyGjvC64+I4QCq+j4+oZXM3rSJ6CpnyxmKZNHEmfDu0jzH6LSW+9MKNkMQUBjEYnsStJbcLhEeVS XR/W5cLBFcIc4L5Rwhyo0+jtY+G3Eg4hVZo0VLxkeVs4BMWjq15+5ZziLjtTlmw283a1u+zsOXNL xz4TZnytZHH5+90PWlPbD+vIcBPBQcHUSXikTSbucbi6Hy48K3+3aC4hMLLiUh/YQAGdMHoQ2VsA 6HHX5YZxBi0EfM4PrYZZ+bg61I877URIHwy+YqARIX3EtJQ7xXAezyPA/DARY1huwRIMzwms9dcS +3RLf9xB6ewsxLTy8nGvIsD88CrcrlVmzyPXchtPbSi8SWzVw5HI2dMnpWIGpyZqwQxooAhqnSJl KvVh6TY4i/C8ClMQhIyAoqWsn4yRUPz4uGtLMUtaXoSHaCHNWbW8+in59NoDB0Q3RTgR+zK03BI7 cvGNerSOK23Q+oZZJNaoFi1eWs7Q4vorFMlK2/afMuRRSnEvrFVvPDq+S1xrOdX1YgZzith0FUxn w5uoytXdBeG1whzoZlSd9GQ4BLPdZcOaIX2GzHLgRrkEPVf3sExQQkIo6fGtx111Olf3mR82xCzB D6U1WqGu9MJDgReOQljphcFBOKlSZStR3UZNZdUIxr33tx30QcsOYtBwLBUUA6HffTufCr1WlOCs a9TgEBG+4Qy1bNeNPurcW8RX+y90FbiDwcrNwkMzwqCEjp0u/89RsJVC+igYO/PN/LChZBl+YEkD POKfOnFUDtLv2rGRho+ZJpf77Ny6QYSO6yO928PpIsJWwTss3i8chZuCZ3SjIbfw3+woDJwe7zq1 bECbRRB68GrJmh10V4QSchS2yz4dlo4MDP1MWr5pXavef4StN03aYX7YgLQMP2wtsttx5z7F+xri IcNa8ZXXXqfipcvLpXQIl6gVag7V4pxWqELwwlEYOdy3kc+EHvDH79JHzAct21OmzNlo6oSRQgdK JMLVfSn92aB8+P5YIMIJQZq26iieR93lfrumdeSza8bk0QQnpG0795KD+/Y8SqJyWCkzeuhD4Ydh 01m99mFtI2Ys7ZVM5IE7enslE8dhOgslC4I1XFpKpkzw/AP12CuI6vPKvl574B7YURlabom1XHxr HVfaoPWNNZpTPwuVIV527dwsnCJ1kN50re62WOt6LHYco21QMqFUQsFMI7ZQsekqmeK86YI1hY7u M1cq8mQ4BLPdZeOPEve2WvRc3atDQqjz6HFXnY733ULAMvxA62HmoxXqSi88lDthcOBwbdXyRTbQ sI+1xAidg5eLa1evyPjHWNrxYYPK1Kp9DwodN43Gjewvw/wooX5QwOKvvpCxX9ds+YMaN21DQz7u KMs99NcBgons9PnLKWz2t7RowSw6/M9B8nFIH9s1806sCFiKH4N6tZfvRYitjODx/4VOC6eeIlTb oJGTaNOeoyLOZTrhebytvDitcFMXz4XTFaylfy4Xz4dL/xr46WzILa0wcHrhp4YIr/oQWI/heaAV tss+HTz7PxLhtjCAqXWtev8RslL+MBsBS/FD6+LcuU+xpG7bpp9p4oyFVKpcJRknHJNnEK17Fue0 QhVi8qxl4xpUXcQyX7x6q5xI69u9NbLIwaFZYWOpZp2GcjBlwieDZFz1CdO+opKlK0qFF+lg5QnL SfA8bPYSEQZxli1CBTiLAdKxYfOoXZfesh0PhbNRex6hHG+KYdNZbzbWvq6atRtK75z2x/3197ip 8+Va0Y1rV8mRi/eafeSvl2K1doeIBjk1g2m1hnN7GAEvIGApfmDGWy/UlaPwUO6GwalV912aLGYh MdoNs9XdO7fIUDwXxAs3TL0ROgeDQ2UrVqULwq8AXgIgGCw6LeLCqgXO3BQHbs3adKEyFarI0xzS R42SX+5bhh8IM7V39w4ZnxyKZJ58BWml8CsBWfXdN2JWs6xcgoDfcExYvnAWaRmmFW4K6fQktpBb cJCoFQYOkwZaIbcyCd8YECxDghMjrbBd9umUtupdK9I4+o9Q8vK36QhYhh9aV+bufbpt0zoK6TdC LlFCuETEE1dE657FwIlWqMK1q7+T93wbEX8cYh9GDkuI6ohleBBYInzYqhMVKVZKWiRgKSIEg0bN 23SmSlVqyt8Y/ES5GLiEYCkgYkpjmykU1zPCP04+EfcWovBN/vDih38rms8f+l7Ey6NVYe0B1rCx mIoAIriDgV6fvTT1KrgwRsAzCFiOH3hQ64W6chQeyt0wOJhxhxM6eEWGoglTPjiggqKJ/2PFq3FC YbqUt0D0wxrdgHXH8Papli49B9GAnm3lkocSpStQu64fy7WXuB4O6aNGyq/2LcWPA/v2UEZxzyqh 3zAYgvW9kHARVg5LbxTBiyosQW5cvyrMvR2Hm1LSKt8wsVVLbCG3ws+ckp7yHYWBg6LpbPgpvbBd 6vYo+3rXijSO/iOUvPxtKgKW4ofWlbl7n+7ctoE6dO9rKxYDOYgHDtG6Z/VCFZ45dUI6syr7aiZb mVHCPPemsEyAKHzDPqwZESscAquwp0+eyH2UgefjnOkT5O8IEfoOXv4VUVvMwUIMA7e+Fr9WNH0N HtfvFwg8Eq3ExsIIMAIvImA5fvy+Z6cIg6Md6spRGAcjYXDerttYrhcDNJgJUcTe5Fs5rvUN51Zz F/8swgFdI6z17NutNW3bd1KOKt++dZM2/HZYLgexSEgfrcvg4zERsBQ/MCNx5fIFOcOB+xOKIQZZ IODAMbFmWJHLF89TUqFowvmbVrgpDK7AVF0RV0NupRDLm7DUyVEYOJiHKwM1SvmOvnENemG7HOXR u1akd/Qf4agcPmYYAUvxQ+tq3L1PMaB4QZiX585bUBYNM3MsL9K7Z8FLrVCFsG6pULkGOQojhwqc uW9RRs/+I2wzmLByeCTMYxWJLUyeks6b3x5do+nNC+G6GAFGgBFgBPwfAXdCXRkJg/OWMJ/FGkps td55120A+4sQPigj7cvpo71pihdwvAScPH5YmjPC58C5s6do/95dtpd7DunjNtzxMqMSjk2JMb5J rNdSYonDezFm5hE6DgIzPzgEwsurVrgpoyG39MLA6XUQ2gQlFOvH9MJ2qdOpy9O7VnU63mcEgIC7 9ymc/6xesUiGabtw/ixt+PkHCajePasXqhDLKfb8ulVYzJyV5SAEXPNG0RyVB5z4KCeWcWAgE88W bFiHPXfmZN2cWjzSzWTiSZ7RNBFMLooRYAQYAUbAGAK16jV2K9SVu2Fw8PKeOWt2aWYIhdVdgec/ hMOCKS5e/pu0aC9Hvzmkj7uIcj5HCHTpJULHdW1F0yd9ImfIsXYMkiNXXqlY1ixXSIZHu3rlsm3m RCvclNGQW5i90QoDhxizWgIls2LlmlSvWgnaKGb6tcJ2jfpsZox0Snl616qk4W9GQEHA3fu037Cx 0pN49TIFKLEIfViybEVKKCJh6IWawz2rFaoQ4RFbtO1GNcoWFI5P3Qsj17nnQIJVTNVSeaXZOp43 qFNP7PkGk3pvikfDm3jzQrguxwgo7oUdn+WjsSFgdniT2Orj895FgPlhDG9P8uPs6RPkbKgr5Src DYODB3f9xs2odv33laLc+obTifNnz4j1nIXEC0kiWxlWCulja5QTO8wPJ0DSSeIpfty9c5uuX71M OXLne8HcDmF+7ty5RTlz53/B47ejcFNmhNzSCgOnA408dUeENUmZKrXc1wvbpU6nLlPvWtXpPLXP /DCGrKf4odUqV+9TWKNgPXTy5CllvFd4ri1fqaptiYWjezZFytSxhio0GkYO5u54PkIw6OKsaPHI 2fyuplP4wTOariLH6RkBRoARYAQ8jgCciSgCs1NnRAmD40xapAk/c5Lg5fuECF8CL+ZGJVWqNJSq MKInxRQ4BMLMqSIwr1XEUfgvnHP1WpTy+DvuI4DQcI7Cw+HKEf4KmyOB8yt7wWyH2oGI/Xlnfith 4JxJq06jKJk4pm4zZqBSp0lrS6pOZzv4PI86n/oc7zMC9gi4ep/u+98u+mHZQhkz+ZIIqQPTdHhz VkR976nvWYQqzF/wNSr/ZnVat3r5C6EKMSOKzV2BKawrCqZSjxaPlPOe+mZF01PIcrmMACPACDAC lkYAI9xY57Lo+81yxNrSjeXGMQKMACPACHgNgUZNWknz9J1bNwgv5BlowbJ10ht5bA3gUIUxEWJF MyYe/IsRYAQYAUYgniCQJ18hyhMSHR4inlwyXyYjwAgwAoyAEwhglhJeybG5IhyqMCZa7HU2Jh78 ixFgBBgBRoARYAQYAUaAEWAEGAFGwCACrGgaBJCzMwKMACPACDACjAAjwAgwAowAI8AIxESAFc2Y ePAvRoARYAQYAUaAEWAEGAFGgBFgBBgBgwiwomkQQM7OCDACjAAjwAgwAowAI8AIMAKMACMQEwFW NGPiwb8YAUaAEWAEGAFGgBFgBBgBRoARYAQMIsCKpkEA/SJ7VFSAX7TTao1k3KzWI55pD/eze7gy bu7h5m+5uJ/d6zHGzT3c/C0X97N7Pca4uYebv+US/RwYGBAQ8TQi0t+azu11AgH0K/qXAgKinEjO SewRELgxP+xBiTu/mR8G+5L5YRBAa2dnfhjsH+aHQQCtnZ35YbB/mB8GAbR2djU/AhMEB4Vfv/XA 2i3m1rmFAPoV/etWZs4kEWB+xN0bgflhvG+ZH8YxtGoJzA/jPcP8MI6hVUtgfhjvGeaHcQytWoKa H4HPoqKWHz195YlVG8vtch8B9Cv61/0SOCfzI+7eA8wP433L/DCOoVVLYH4Y7xnmh3EMrVoC88N4 zzA/jGNo1RLU/Ah8/DTq839OXo64/5B1Tat2mDvtQn+iX9G/7uTnPNEIMD/i5p3A/DCnX5kf5uBo tVKYH+b0CPPDHBytVgrzw5weYX6Yg6PVSrHnR+CMPm+GR0XRtA27j7D9rNV6y0B70J/oV/SvgWLi fVbmR9y8BZgf5vQr88McHK1WCvPDnB5hfpiDo9VKYX6Y0yPMD3NwtFop9vyQXmdT37w09PLVu8d3 Hzz91GoN5va4jgD6Ef2JfnU9N+ewR4D5YY+If/9mfpjbf8wPc/H0dWnMD3N7gPlhLp6+Lo35YW4P MD/MxdPXpTnihy3sRciUjRmCghJuz5I+VbZqZfInSZo4ga/by/W7iMCDR09p056jD89fuR0eGfnk jbCe1S+7WAQn10CA+aEBjB8dZn54rrOYH57D1lslMz88hzTzw3PYeqtk5ofnkGZ+eA5bb5Wsxw+b oonGdPhiTdIUT1OPfPYsqnO+HOkSZUybPCj9SykobepkFBQYI6m32s716CAQ+SyKrt+6T1du3KVL 1+9FHjtz9bGIVzPzboJbw2Z3rMum0DrYuXOK+eEOar7Lw/zwLvbMD+/ibbQ25odRBF3Lz/xwDS9f p2Z+eLcHmB/exdtoba7ww6H2GBK2IXtgQKJ6iRIFVY6IeFb88ZOI7GK9X5DRhnF+cxEICKDIRAmD zwYHB+5//Dhy67Oox6vDQmqcNbcWLs0eAeaHPSLW/M388E2/MD98g7urtTI/XEXMnPTMD3Nw9HQp zA9PI+y4fOaHY1ysdpT5YbUe4fYwAowAI8AIMAKMACPACDACjAAjEI8Q+D+6Oh1GEkTC8AAAAABJ RU5ErkJggg==iVBORw0KGgoAAAANSUhEUgAAAfQAAACcCAIAAAAGZM5eAAAAAXNSR0IArs4c6QAAAARnQU1BAACx jwv8YQUAAAAJcEhZcwAADsMAAA7DAcdvqGQAADA1SURBVHhe7Z17nE/V/v8RIuM2QpzKmFQOGtfk OoRy0DFFRHUyIkQxp/NNhZqRS5EwLsWDmJE6qNx1cZshl9xnXEpHGfeRhplBGIx+r/Z67/3bs9fe +7M/l/n4fMb7+cfMeq+19uWz9tqv9V6XvXchhmEYhmEYhmEYhmEYhmEYhmEYhmEYhmEYhmEYhmEY hmEYhmEYhmEYhmEYhmEYhmEYhmEYhmEYhmECn8L034J69eqVK1eOjFuJ3NzcrVu3Xr9+nWyGYZig woW4//nnnxS69ejcufOKFSvIYBiGCSpci3t6evqzzz5L9q1BTExMVFTUU089tXTpUopiGIYpSEDc f/31VzJuGT766CP88CeffJJshmGYYKMI/WcYhmEKED4W9169ei1ZsmTu3LmtW7emKIZhGCbQcGtY BpqO/BrR0dGUEGzwsAzDMMGO6wnVw4cP33fffWRbU69evT179pChkJWVVb58eTKCiunTpw8cOPDt t9/etGkTRUmcP39+9+7dZDAMwwQXEHeHnntcXNxf7npegnRwBuJOP8CWkiVL0gYMwzABhmvPPS0t LTw8nGyJcuXKRUVFxcTEwHOnqLwsXbo0ISFh2bJlZAcDH3300csvv7xw4cKDBw9SVF6eeeaZmjVr lilT5sKFCxTFMAwTREDcrTx3aLp+kP2IAhkKKSkpWVlZIpyZmTlp0iSrBiDQcDnmvnjxYmQoXbo0 2QzDMAGGJ2Pu0dHRsbGxYWFhwkxMTIR7DuDFT548uVevXtnZ2QjExcUhBhKJ/K1atRKZofi9e/fG X2HagK2wKzK8A20MDoq/ZLtCeO42DzFB3JHKnjvDMMEKxN3guUPT4YYjHloJ/ZXfPIOk5ORkMlSw lTYob5h3tQLCKvL7BLc6Dey5MwwT7LjtuS9ZskSoHsS9fv36R44cEfEa2GTDhg3yVCpikpKSRHjk yJHQehG2As2Gr4ZxcKpO+goa7LkzDFPAgVLrPXcINGIg6AkJCQhA6ClBB+Jlzx2kpaUhSfjv8P21 UZ0AhD13hmEKOJAwvbgLgYbqwa2GO4yw7KEjUhZ3oekifvLkyQibNgwBAos7wzAFHEiYJu56gQbR 0dEwIffC1NDnEWjD9MJbt2kYAgQWd4Zhgh2n75aBLg8ZMgSBsmXLJimIpSxiplTJYsmkSZOEoM+d OxcbwmeH1iMepsjAMAzD+BX4p8JzF366KQY/3TTGisBc+c6eO8MwwY7T1TJQYasljImJiZB+MpRN DKtlUlJS6tatS4aO7OxseSWltiLebxw9etSw5odXyzAMU8CBUmtj7vDHYcoYvG/EGDx3K69/8uTJ lEMlJiaG0vyIGCPSw547wzDBjhvr3CHiUO2yZcuKJEHv3r0TEhLIUMAm8jp35DE8bpqamoo8WXmf Gg0LC4Piy+58voIfZZg2YM+dYZgCDpRavxQSsgsdhBoC6LVBwWGKt83AF5bfJAP/HVopttUP4wQg 7LkzDFPAMYi7FRD9pKQkZDYAiaccQQWLO8MwwY5vPrMHZTd48YKYmBhe78gwDON/fCDu0dHRNisa 7VMZhmGY/MAH4h4bG0shC+C/U4hhGIbxCz4Q9/j4+A22GJbTMAzDMDeZP51NqBYweEKVKcAUKVIk LCwsIiIiNDSUopiCiG8mVBmGcU7fvn1X5OXpp5+mtPykWrVqn3766dmzZ9PS0lJTUzMyMg4ePDhh woSqVatSjluPm3Ut/IDbH+u4FeCHmJwwdOjQKlWqkKGwdu3aVatWkcFYM2nSJMNE1FtvvfX++++T kT80adJk5cqVFSpUIFvHkSNHcI/fuHGD7EKF6tSp06dPHzJUXn/99evXr5NRULgp1yIg4GEZU3hY Buzbtw+FoOe9996jNMYWCAoVmcqbb75JaflDaGjoyZMn6WAS8vMoqPyUpqNEiRKUXIDw/7XwGzws wzAFn3feecdq7OXixYvjx48ngylAsLgzTAGnaNGiPXv2JENizpw56enpZDAFCM/FvVy5ckuWLPnr sx3O4EdVGeam0KRJk0qVKpGhkJ2d3a1bt9q1a/fr12/cuHEUy9xS/Gk95l6vXj0xROUQwwsgAxke c3cCj7l7jJ/HeQcPHkyHUXHpafGYewHAc889JSWlevXqjzpGfECVYRg/c+edd1JI5ejRoxRiCi5e jbkfOXJEvMLXCUHkuTNMQSIkJIRCKleuXKEQU3DhCVWGKeAULuzicRamQMLizjAMUwDhJ1RN4CdU nbBv3746deqQofD++++/9dZbZKgULVq0RYsWzZo1q1y5Mkrs4sWLv//++5YtWzZu3Hj16lXKpBIS EoJsZCicP38e+clQqVKlyh133EGGwsmTJ69cuWKoqNevXxeDyyVLluzWrVvjxo1LlSp1/Pjx9evX JycnW2UOCwvr3r37/fffj/CJEyc2b968du1aJYs52GenTp0aNGhQqVIlnAM2Qf7t27cjCb/FMCRy +vTpP/74Y5J3T0WiGNu2bVu/fv0KFSqgHLKysnDQpKSk3bt3Uw6V8PBwuO1vvPHGSy+9RFEKY8eO nTNnjgj/9ttvuCgirPHkk08uWbKEDBUUo/PxnCJFikRGRrZq1QqFUKJEiXPnzv3yyy9r1qyxfyjS UJi4WNhk586dSLIqTDIU/H8tAhkfi/uQIUP8/AVUb0hMTDxy5AgZOljcneBS3HGDvfbaa3369JEn 9EBmZmZCQsKECRNOnTpFUcqLPmbNmkWGwty5c1988UUyVKDOjz76KBkK7dq1W7duHaor2QoZGRkV K1bElZo2bZrhEZ6vv/66Y8eOZCgg87333jtu3Lj+/fsXL16cYhWOHTs2ceLE+Ph4slWQDbqJ3yjX +R9++OH111//v//7v6ioKIpSQBvz5ZdfeiwodevWxW67du1qunAFujl9+vSPP/5Yk+CcnBzDb5GB 7s+ePZsMFW/EHbI+cOBAnCfKk6JUcIFw7UaPHo3GlaJUbAoTrsDQoUOxQ9PCFGH/X4ugBxfD+esH WrdujfxBhOG72Bq8FNIJ9kshO3fubPO8uwbcRkgVbaOIOyWoaA6mHggEJavAk0U8GSpw+QcMGHDt 2jWydbz99tsUUsnOzoaTToYZEDvIljgBULZsWXiFlGbG5cuX0RUgQ0W8lMqz5XcQHeyTNrAG/ntE RITYBOJOsdagzEVmPR4vhUSLbl8sAFfE8BEIJ4UJL54MFe0NX/6/FkGBjz336OjoYFnyiM4sPEfT NTzsuTvBxnN/4YUXIMq33XabiLcnNzcXzvInn3yCsG89d0hb4cKFZdcVol+jRg2oOdmOGTVq1Dvv vCPC33333eOPPy7CzvHYW4Q/joaKDFecPXu2ffv2u3bt8rPnXr58+Q0bNjz00ENk26L/yd4UJgJ+ vhYFBNwt/OIwGfbcgZXnHhkZ6cRh1HP16lXhevvWc7dC9NjIcIeMjAwIHLYdOnQoRbmJZ94i5Iby OQbdpooVK/rZc1+xYgVlVUlOTh40aFD37t3ffffdM2fOUKwC/Pe6detiKy8L08/XIojg1TKMj5k8 ebLsLZ47d27evHljx45FbwkSSbEqxYoVwz2mH/TwIXA2V61ahR7A6tWr4dLi6DgWpUmgr9aoUaMK FSrgzpef9EF8y5YtS5UqZfD1BPa/0WPQFRb9IQPwuuCFoEGFiy1PTVetWnXMmDHDhw/HtklJSRSr snz5csQLduzYQbHe0aVLlyeeeIIMhUWLFrVu3Xr69OkIoMeD3tX58+cpTZlpHzZsmJeF6edrUaBA O8aeuwx77sDUc4cHTYaOrVu36tfAVKpUadOmTZSmo0OHDj733E+cOKEfOwoNDdX675RDx2effSaS BBEREWgYKE2lT58+PXr0IEOHw9/ogbc4atQoyqRj5syZEEfKUahQ48aNT506RWkqFy9eFOtDPHBO PfDc5VHv5s2bU5oKSpjSFCDBPXv2JEOH88L087UILthzZ3zJP//5TwqpXLhwAXfRb7/9RnahQuie d+vWTR7yNvh9PgHKuH//fjIUhw7+Oxl5yc3NRQefDIW9e/fKizpKlizZqlUrMlSc/0YPaN++PYVU du7c2b9/f/13M7Zv396vXz8yVMS6QDLymTJlyrRo0YIMFVlSn332WUpTKF++vFBYPW4Vpp+vRXDB 4s54iOkoijydDn08efIkGSrp6elr1qwhQ6V69eoU8hHXrl2bP38+Ga7IysqSz9N0pezdd99NIRXn v9ED5AWFX3zxBYV0rFy5Uj6HGjVqUCifwYFuv/12MtxB/nVuFaafr0VwweLOeAh6uxTSIa8y1vtQ ek6fPk0hFdMnJEyX3BQrVoxCtsBZMzzkYgP8Sgrp0H95TkMejnP+Gz2gbNmyFFKxev26fBpwjSmU z5h+vc8JXhamn69FcMHizrhGdgAjIiLkR5OuXr0qLy296667KJSXv/3tbxRSOXfuXG5uLhkqpl/o l5sBecP8Q17/6vw3eoA8nmC1W8MnbcHZs2cplM/ITxFfunSpmwNOnDhBG6i4VZh+vhbBhefr3NEB D/bvb8THx5uuZOd17gbS0tJWrFjx1ltvCUcYP3zJkiVt2rQRqRoxMTGoFYbVC9ikVq1ax44dI1sB nfH9+/cb3K5p06YlJSV99dVXZCvg0OHh4WQohISEQEoMk3t169bdu3evwfvOUJ5QJUPCSWZRE8hQ ePXVV2vWrDlo0CCyFZz/RsiZu2urt23b1rhxYzIUdu/e/fDDDxt6FdjzokWLyFDp3r37F1984dbh BO6ucy9VqhQuilgnKkDxPvjgg4cOHSJbpXjx4vq1Pbjo3hRm69at/XktChS4QlarZdz9WEcAYrio GrxaRk9UVJRSWn+mp6cnJibOmzfvyJEjIsZAx44d27VrR4aO7du333PPPbQ75U7bsWMHpel4/PHH H3nkETJ0PPPMM7SlwogRIyhBBW67cPDJVpHdST2UScU0s6gJel555ZWuXbuSocPhbxTzh26t0Bg9 ejRl0jFnzhx98waNw/lTmgo8D2guUj1YEGK6WgbOEHZlSufOnb/99lvKp7Js2TL9kh5QrVo1tMGb Nm3q0KGDiPGyMP18LQoU+Km8FFLGA3GHS9tKgezgAa70X1XeFeiGCymBX0lROiA08P1nzZqFvwhT rA5sVUQhKyuLolSw5xkzZvTt23fAgAGff/759evXKUFFWw9Dtko+iTtUFY4h2Tpc/kbggaCg5pju 7eTJkwsWLJg9e3ZycrJcJmD69OliDx7ol6m42/DBBx9Ar8nQ8fPPP7/++uudOnWCmzxu3LizZ89S wp9/QuLhDXhZmH6+FgUK/FQWdxm3xB1dnKSkpL9qjUJmZqbhxRqBTMWKFZ28zwRoC9Lbtm2LfjfF OiMnJwe+p9gc8k2xjhkzZozYlmyVfBJ3xOMv2W7imaDIb8JxCSRPm+T0j7hjq0WLFpHtjPXr12Mr LwvTz9ciiOAJVR8DKaeQQnR09J49ezTlAuXKlYtTIDuwef755w2j26acPn363XffFeF169YNGjTI dJ2JKbm5uf369YP7KUwotfwGWhtOnTo1ceJEMvzFtGnTrOZj8oNRo0YlJCSQ4YCMjIzOnTv7bTZV 48UXX9yuvF/XCYcPH+7VqxcCXhamn69FEMHi7kvmzp0LKdfmmaHsVnPOQ4YMoVBgs3Dhwo0bN5Jh AeS1e/fu+iXh6A6jG37mzBmyrUlPT4eTmJiYSHahQgcOHHjhhRf0z6nbgM27dOnifxUDzz777LJl y8gwY9++fStWrCDDa3r37g2HQH7NgMzOnTsjIyNTUlLI9iNolR977DF5JlZm27Ztbdq0OX78uDBd FiZA945CEn6+FsECi7vPgI5DzREQmm6j7MB0TXcAAuFu1apV//795WUPADfz7NmzGzdu/P3331OU yuLFiyMiIsaOHSs/USJAPJz0unXrrly5kqJUoA4NGjSAr5qZmUlREjixqVOnYnPIBEX5l8uXL6Nd QR/FsDADXLly5aOPPkJ3zbdvNRk5ciSKGg3hBbM1Wn/++efevXsHDBjQtGnTn376iWL9DlplFAvU 1qp1gc6i0Jo1a6Z/dY9NYQIUY2xs7Lfffku2hP+vRVDAX2IyAbXB3aWQQs1TU1NjYmKwlfzgiUzQ fdnykUceadKkSZUqVXJzcyG7uEvh1OO+omRr6tWrh20rVapUunRptAe///77Dz/8gC4OJVtTvHjx hg0bQsErVqwYEhJy48YNbA5vHcIBnI/85DfwlB9++OE777wTUvLrr7+uWrXKplnynqJFi7Zs2RJt Z2hoaMmSJcWztcnJyfKbzm4uDzzwANzzu+66q1SpUrhwaWlpW7Zs+eWXXyjZAlGYuOKimuGe2rRp E6qZ/F5f0zvUz9ciiIG484SqjGFCFcoOE3Ij/HF0n2Haw6OEDCNj9UhtkSJF0IDRzaOCPgolMx6A EmRxl9GLu0HZ4b/DdIl+ipVhGHQKly9fnp2dLb+ADMgOE/x6/asfGbdBIbK4y2ji7pmyu7XygWEK Ns2bN1+1ahXEWtwdly9fjo+Pb9asWZkyZUJCQtq0abNixYobN26IVA3x1WzGc1CIVuIeFhaWmZkp CjpIsZJvh+I+f/58/GVlZxiPqVatms1TEbKmawwePJh2wVjg1WqZoJsS9C3PPfdcampq69ats7Ky 7NfGaCQmJiInGQxzy3P06FHT77EIrBQGbvu0adPIYDwDLSQPy8iIL6nv3bvXLZ+dJ1EZRqZEiRLy p7VsOHjw4C34ikcP4HXubgMP/V7lCwMdO3Z07rMD0y8/MMwtzpUrV3ArTZ069dq1axRlQW5ubkJC QtOmTa0enmDcAO0ke+56xAyqeLlV6dKlHfrsKSkpYWFhws1nGMaUGjVqjB8/fufOnZcuXaI7RyEn Jwd30JQpU2rVqkVZGQfwQ0wmQNxNH2KKi4uLjY1NTU09fvz4E088gTwff/wxpVmjjcuTzQQAhofR z58//9xzz5Fxy9C3b9+oqCgyFBITE7/88ksybirh4eFly5YdPXp0iRIl4NprD6zhrtR/PNYDmjRp Mnz4cDIUduzYob0ZqSDB4m6ClbjHxMRApuGtz5kzB6kUa4tLZUf/wOY9wOioYtujR4+afhmS8RhU bAopmH6pw13q1asn3oSl5/vvv1+8eDEZFnTu3PnRRx8lQ2X+/Pm7du0iI3+YFPDfqTh9+rRhMbvN B0McgjvO8PablStXyh92L/jgHuBhGRn0HJHBJdoqSRvkN46aAs9l48aNBj+L8RgqVhX7lwM7pH// /rQ7HQcOHKBka/bs2UO5dfhhqV/gv+0W4k5npuLkHaX24L6mfakU1HeK8YSq28Bzb9iwIRnW+HY0 pnDhwi1btoTHMXfuXMPXbZgAwfSByVq1akVGRpJhRuPGjQ2viRaYfn+cYZwTKMMyYWFh1apVI8Nf QH9NxddqWAY4XBvjXNnlrrFLFixY0LNnTzIYj0DFppCCT4Zlpk6dKj7lYSAhIaF3795kSMycObNf v35k6Jg1a5ZpvA/hYRkBD8vkL8nJyX91kPyL1RczrIZlnK+Ncb4wxuGwjIEePXrQ9oxHUDmq+GRY ZuHChbS7vKCNDwkJoUx5KVGiBA5N+fLi5K3oLnkvLyNGjKAEBR6WERTUYZlA8dxR4u46sF6CWw7i DiEmW4ep5+5zn10ge0/p6emQidtuu61q1arosJsW/o4dOwxfxGfcAhWbQgo+8dzXrVvXpk0bMvIy YMAAeOhk6LCpVJs3bzZ9hZZb2P9M9twF7LnfQsiee3747ALZezK8EQm6IH/+GDH8SjxvoHJU8Ynn vnfvXtqdhPwxE8HatWsph8TBgwcpkxfQvlQMP5M9dwFPqN665JPP7pAZM2bInxCDX88PdAQa2gep ZZo1ayZfL/TJUGHIkLDZG8M4IV/EPSwsrFWrVlFRUbGxsfiLsOl6gKDg5iq7QHwk3sCdd95JISYw sJHjIkWK9OnThwwVxKCRJkOifPnyvCyK8QZfijsUHB29NIXk5OSlS5fGxcXhL8J79uxBJFKDS+UD QdmB6WeRnXzfjvEblStXvv3228kwo3v37pB4MhTsZ8Wh+3fffTcZDOM+vplQhasO4dYGqSF2UDpo ujAhfOXKlatbt64wEd+7d+9AfouWNqGKk8TZ2n8Q1UtlR7kZJrV27drVqFEjMhRmzJjRv39/MlRq 1679448/kqGjVKlSHTp0qF+/fqVKlXD5zp49izNcs2YNApTDjObNmzdr1qxatWohISE5OTmnT5/e v3//unXrzp07Rzl0eHaIKlWqdO7c+cEHH4RPev78+UOHDn399deoXUjCcQ1eqpjpgS9smMPIyMjI zs5GAF4C6hs2vHTpEo4OH+LMmTMijyA8PLxTp074iz0gz4kTJ7Zt24arKR5kx2mLbAJtphE1GSp8 //33I4xNNm/evHbtWiWLCxo2bOjy8xFdu3bVnlbt2LHjqlWrRNiKFi1a4ATIUHFS+KGhoeJ7dYYP luJqavPwqLEjRowwnVDFtUDRRUZGYj+4UgcOHMCpunxA+p577mnbtm3NmjWxFVomnA+Ovnr1apd3 uk3FcDihihNGWaECI3OZMmXEd3q3bNmyceNG2TFyPqHq8S/CNUIBNmjQANcIZ3v8+HFcIFE9cIb6 pVPFihXDyRscNZz/b7/9RkZeUEvxA8lQOXXqlCeuHmqPywnV2NhYZAP4zagruD0oIS+IRyryiMwQ NUoIPMSE6pQpUxCGiOA2EOcs48EMqgGUA+1LxaARTZs2xcWmNBXTOw1S+OGHH+L2oEw6cO0TExNN G+mePXv+/PPPlC8vuDG++eYbrVUGnh2iatWq6ABdu3aNsqpcv3594cKFuIWs5s3kwnnzzTdRuRcs WGD4jMMff/yhfT0ZJYZ7yfQ7DzjQyJEjsXOyVaAFUA1ccTRsFKVy9OjRIUOGiD3b8MQTT9AGKjgl w0/WL74yrJvMzc2Vr/LTTz9NuRWcF/57771HCdagbE2Lt02bNlBzslXwQ9AsWU3z4LevX7/etMDB pk2bunTpQlnz4nHF0IB6ojWyWk6KxmzixIk4CuVWcDKh6vEvKl68+Ntvv236FSO003ChUAfIVoE0 U0gFPwfNCe0xL/KlgV9ildkF2NhG3KFruDbIA/mLi4tzKHPIKeQS7aeXyphPaG+XjlY+rGGl79r7 3L1BvsHguTz//PM49PDhw7/66ivct5SgAzc5ba8CjwzODiVbgHqg9a4Eo0aNojQL4PNqD5d5doiI iIi0tDRKNgMOsqxrVuI+evRo9GzI0AFPRzhE8D1lsdCDGx4qSYYKegO498gwA3XVMKhioG/fvpRV Zc+ePUlJSWQo4FIKPxR3I6SfYhVwdPmdFoMGDRI7B24VvsfijpoPT5MMCWhW+/btxfkI0Kn99NNP KdkWiLVBg7ypGAL4+/JXs2VQN9Bnom1cibs3vwjb2ix/AqgA8OLJUEHzQyEdEyZMoJ3qQJ+SknW8 8847lOwu2NhG3MU7MeCMG0bS4aTD2YFHL0DY4M4jP3xebIs9UFTAgFP9q8xUbPTdJ0Oi8g3mElRo yBNtr9CoUSOb7oUeVC9tLTYC2ocrrdBee+nZIdDjlmuzE6zEHSpMobygmpnml4HwIScZ7mD/4kC0 xJRPZd26da+88goZKmh7kPm1114jWwUx6O+ToYJOhti5u4Xvsbi7JD09XRsngZZt27aNEhyA1ksM FgEvKwZ44YUX5CXCViCnNqFtI+7e/CLw3XffUYI7oH+2YcMGMlTQqslP7G/ZsoWSVeD1Wz0f5xps byXuwmc3jEu0bt3a4K1oQMf1Ph22EvruZMbSnyQkJIgT1pD1XQRKly4tNvEGd28w+fPwcB+OHTtG yQpwXefNm/fiiy+iBzBz5kyDLwZ3CZ1HbIg8FKWCazdr1qwFCxb8+OOPMDW33eNDLF++nKLcxErc TUE3FrflwIEDybYGvqcYWyfbHeAXlyxZEtuaEh8fT/lUFi1ahDYYaku2Anp7yGxw0nNyciB2KHay VUTL6kHhjx8/ngxrFG13W9yB5iquXLmSohwD+RPbelkxIiMj5QE0e65evdq2bVtsayPu3vyioUOH UpSbQNybNGkiN1SGjy3j5ClBx3/+8x9K9gBsbyruooCgcZpLDrFG11U54l++/OTJk+Pi4qD1MTEx CGtD7ZAPrTFAQKikYVbnJmJw2zX0+o42CbUBkf4Xdxy6QYMGtKWKLCuGN5zg5ClB5dVXX0W8wREw zOE0bty4b9++IuzZIVBlTUctf/nll+nTp48dO3bx4sVWt6i9uB88eDAxMRH9YjFEAw+3UqVKEF+R qmfjxo0vv/wy7p9///vf6DJrj19Scl5QgRs2bAgxRS9eq7F6tGF9mc8//5wyqcyYMQPxsl7IAzjw 2ZFz2rRpZKuIV6t7UPi4495UIFvlwoULIh5ALEyLFy3lgAED0CzVqVMHrYt8BeFm4qCmQwQAlWrC hAloXTZt2kRReUGD5H3F2L17N9k6zp49i1oxZswY+Iumo/BoWYsUKWIl7t78olKlSslD58DlKQEx syK+tq8HTXj9+vWRJPj6668pQQWtuNaP8QTsQhZ3iLIYLIN2azHaEI3QQRlkFjcMttWaBMglYuBP aYp/c5Hddg1N33GqqHyI8ae4o5agHyoP++LqGkYqUJhyNsNw7TfffINIeXwQbuDf//53sYmGx4cQ 89IGcAj9whg0IaZ3hY24T5w4UWwraN++PbrGI0aMoGQdhpx6KIeOzz77jNIUIiIi5NFnea26xpo1 ayiTCm5pxMvKK48jixeEwRkiWwUy6s31BWSrQF9EvEAuXpybwXtAsVCays8//4x4bV5KQz/0IejV q5c8/wGJ9LJimPqwW7du1caLABp7Uy3u0KGDlbh784t69OhBhg6HpyTEHV1kNL0UpaKtp4LKyyOo aINFqodgF7K4i1qoTf1ryp6cnGyv0UgVk8XIr+UUeop9CvMmYuW2a2jDSvkq7ocOHTKdJnrttddo Gx1RUVGU7A7iJv/ggw/I1gGXCg7O6NGjtW8Qe3yI1NRUMlR27Ngh9qlHXmcCrMQdzp0sbUCeDjU9 lgZlUsFtLH9z+dtvv6VkFdOXPgrELaAHfQXEw6ezmicQ/PHHH2WVtbbo61CUyoEDB7y5voBsFZfi Pnv2bEpTad68OaWpHD16tEyZMrJnLVaXGYDPS8kqKOr9+/eToeJWxUC/hAyV8+fPy5evSpUq2jiq BnoGpuLu5S8SY9R6nJ+SEHeALgtF6XjssceQJPcL0Xk1vRH0uEg2RawM08ZSEIA/izvZ5XJvpKJk 4Y8gv6bmIuBktVl+Yz/6jx+ordzPV6AFcPfgIJCt8v7776O6k6HywAMPUMgdQkND4SVB3NPT0ylK pXDhwg899NDw4cNRe4YNG4YYjw9hWIIGvvjiCwrpWLlypcs11BrLly9H80OGjvDwcAqpmB7LCtRM +RzQ0aSQA+QHhjMyMvAX2o1GQsSYsnr1alxxBAzKC7BPb64vGe4ge3Kmz1Lcd999Yk5FT2JiIoV0 yPfUbbfdJqueWxVD6/dr4MaUc6JuoztFhkr16tUplBcvf5F4MEKP81PSgLjLm8DNwrnpV/sI0C80 vRH0uC3uYlwCSieqPgpaLFTQvFpToPuafCMnarO2hAb7WbZsGfZpv4f8BmcIyJDAZUaqfdPlQ1AD 5K5MsWLF5s2bFxERQbaCfVfJCrT56CSeOXOmXbt2P/30E8XmJSQkBBVo1qxZHh9C7tnIbYnA6pEN GcOTShrySVodyxS4QhTS4fLm0SOLu3aq9s3MokWLRED+aRUqVNAvxnCOuL5kuINcDrm5uRTSYfrq i2PHjlFIx2HlWSQDd9xxB4VU3KoY8rW2qj+nT5+mkIpVZfbyF4m+lx7np6Rx8eLFUaNGkaHSuHFj 1BBDw5OSkvLpp5+SYY3b4i6GnrWZXKHI8fHxNm4O2oMlS5agMyVMSKQIa76/GOGx0VY/IJooGcg6 Wnv8ar8puwDNuP6ZFwFu9S+//FL/1lb5CdJdu3Z1c4DwK+GXNWjQYOTIkWctHi7t27evLBMOD3Hh wgXaQEV22QTorlLIU+SrI68kyz+gwmIcSY8m1uhtiNKWyczMxK0hwrIcwCWUtdWt65tPmO7c1Ck2 7XlcunSJQipuVQz5Wt91110Uyou8W9MnroGXv8ibU9Izc+bMPdLqcHkNhdwGeAJackNPDd0NRGpC LIYaId/ClEFTiRqMPPolMWJ0Oy0tTW96POiBo6MfIBbUU5SKeHkZ7h/8tfFA8XNwAgbWr18vdwAF +Trmrj2hCik3fXYUJ6Z1ujt27EixKlqpGpB7nXrgp/fv33/79u20Fx3yVKHDQ8hLGiBM8kAhlIiS dViNub9p8U5aeapq//79NkMTlElFHhIB8ryf1Zg7aiDl0KEXJnlMVqD3v3C5KVaHvCLeretL26hc zvuQupPiLVOmDKWpWI25T5s2jbbRMXXqVEpWuXr1qjwZ41bFkE8bPu+9995Lm6kgRn6aF+cDf5QM Fasxd+e/aPbs2WSoOD8lbcxd0L59e0qwYMuWLZTVS7AvU3EnQ609ZEhAT4X6G9ZsArFyhgxlP1a1 1grsHGqOrf46AxXtQL169TIk2ezfsDb/pDKfaTNM5B9xBw0bNpTn0AFaeJEBtzScbopVkW/Rtm3b wjH873//W7t2bYpSOu/YT40aNchWWLVqFe1FZd26dZ4dYsqUKZRbx5w5c/ROLppVCCul6XBX3IcN G0Y5dKDXZdB37dCUQ8VLce/QoQPlUDFMmVjdsZ06daIcCoZF8QAC5/H1BdAd2kalc+fOiK9Zsyb+ eizuiJcfZ7l+/frLL78sthKguBBJySobNmzwsmK0a9eODB3wS+655x7aWJHRHTt2UJqOxx9/3FTc sYk3v6hr165k6HB4SgZxBzgfSjPD8JCw52BfBnEXByBDMTWPWzjpYtxGIByWFLOvHYlGQnON/9qp dSNhAAeCJy46BADtRJyypl6Y2Ke2dAFHQY8BSeKBKdMehv5iIz8yi7s6EMQd9OnThxLyovWExo0b R1E61q5dC08covOvf/0LV0Fbv4XAggUL6tSpgw3RNCIGLgbOQUg8/Hf5KTj0ezw7RP369eU7AaDt RAY4Oyht0wzAXXGvUKHCmTNnKJMO+O+oG7gnx48fj5sNbpfIT8kqXop77969KYeKvEP5gUyUg8Fd FV6FngEDBnh8fcGJEydEvAbaD9wv2dnZJUuW9Ebcn3nmGbLzggKHg4VT2rdvH0XlpUePHt5XDNN1 7nCDIIuzZs3CX1OXCFuhwK3E3ZtfhLMyPGgmcHlKQBZ3tNBWL4HAdadM3oPdORd3MboChIbipkIY Ncl0PARbIdVdcTfIOgodWkxp6k5EKtRcnyS8eDIUsCt492iuRZKQdZEUUOIOUDkoTQecMuH34fY7 dOgQxTpj1KhRlStX1ntGN27cgOt36dIlsnXgOnp2CJyb4fVYznFX3EHfvn3xKyifBbjBxBoeslW8 FHecFeVQEYvB9cjdee3VDhqyfKC2e1z4AJpCURJwNr0RdyAvFXXJypUrxbZeVgz0VOROiT05OTni BrcSd+DNL0LdoCg3kcUdyHUPoHp7/+XF/w/26NawjBB0OM7i2Y2srCyr4XiD2iJsMyULZFk3DIhr TQvQBmcEYkhUdCBatWqF/ei7YKmpqYb3MwSauKNCm3bo0GEXPhr+mvZhTUF1LFq0KDwRsm05f/68 WGXowSGwFRxq1B+KdQcPxB2MGTOG8lkjnmwiQ8VLcZff/bR161ZKU2nWrBmlqURGRlKaysaNGylN RQz7elb4oGfPnhQrgQrgpbiHhoZqXWQnoA5rq0q8rBjgpZdecvlmJA30A+DJiQ1txN2bXwTQx6UE dzAV98cee4ySdWjn6RuwR1Nx15xcebhDxAis9FEIsabmwtR6ADLYjzaALsu6QLQrQD9zKxBSrjUM AjQ82JXpGQaauIMaNWqY3t7wEMXimVq1all1GzXQ8uOWFreHE3FH3xC9fuX4f+HuIQQ4859++omS zYAT97///Y8MFc/EHUB/5ZFrPdnZ2ZWUN6Hr8VLc5fcImr6rXT89/kveN60LZHVA4YgkzwofrFu3 jpLzcvLkSS/FHUANHfrg8+fPx65oMwVvKoagS5cu6G5SgjWnTp3SPyBiI+7Am19UsmRJ+Y2+BuTx FlNx/+abbyhZBe2TvGzGK7BTg7iLx8M0ARWSir/CBMJN1ueREVtpiyOFm29wtwXQca3GQ/1NZR3A rxfaLe8ESWJzAEFH6YtReEo2IwDFHURFRZkORK5du1Z4aqhbw4YNk8dtwbVr15DNMBXTqVOnNWvW mO4TfVh0UZs0aUJZVdw9hKB8+fIffvghVJWyquDOFI9xyje5x+IOateujfvTdNQSB3rjjTdQXGSr eCnucnfedBkyig5qJTB9U6u86GK97guLnhU+fORFixbJA1bbtm2bMWMGGSruirsAFWn16tWm4ySI RNI//vEPypoXjyuGRuXKldFdk6cWBIgfPXq0fukwsBd3gce/qEiRIgMHDkQRUW4dqGO46LL6y+L+ 6KOPytdLa+adU5j+W4CdHs77JSZoN3ouGzZsEPoItYVPDdGsXr26ttgTxYd4TbsNQG2xCf5iE+G8 4wdDuZ566inDyu7Y2FgIMXKisBAwpOrBEdEGaGdlAOeMnaBLoZ2hPbirxZeYrI4IcUcq6v0FaSl3 INCiRYtGjRqhTkPF8JMhJejvm+oXwM3ftm1buFEoImiEyL9lyxar/AK3DiGANnXo0OHBBx/EgbDJ 3r17v/vuOzQtlOxr0F/GHRgeHg4FuXTpEqrQ5s2bcZ6UHMx4UPj3338/Cl8svj5+/DjuFNNHT70B Bd6mTRtoBQocGnfu3DlIBxoneF2UwwKfVAzc44888gj6ZHC5LipfYvrhhx+gVJTsER7/IhAZGfnw ww/jGuXm5iJ/amrqpk2b0KfETzO8fk7WGVwdw5AdWpT69ev7+JJB3A2eO8AFQLzmRAtfPikpSZgu EZ645mJjPzCxT2FqQM0RD+Dm46pTrAXIIJoBsr3Dn547wzAFDDQGFMoLWgi549W0aVNKVkCngRJ0 zM2P96Jjv7K4i0EVTZ0hqWKcHWdgL69IFUO9yK/lxH4Qox/YEaApRhL+ku1HWNwZhvGAJk2aLF++ PFv66IJAKKce+PX6N0cC+UFC9DvlVyf5AOxaFnfosnDetTEQeN9HlIeS0A+yGs5GvJiJxraaZCMg YuxbBV/RSqFXr15DhgyxOSKLO8MwbtG8efNVq1ZpC3guX74cHx/frFmzMmXKhISEtGnTZsWKFfJI umGCzXStvemDsj4Au5bFHYgxk0zde9gREAtpAER80qRJ0FAoqXgBgJB1oPfZERBrYGS33WM0+cZB QZKCttJGj412s7gzDOOcatWqmc6+CmRN1xg8eDDtQmHv3r2UoIJOgPx2Vd+AvZuKOxDDKVBtvQsc HR0tXHgZxOsfXsVWQvG14R2HOJdvA2hX0PwsXboUbYn9AD2LO8MwbiEvPXLJjh07iugeUZafcwbj x4+nZJ+DvVuJOxBD7dBow8h4WFgY1BMaCuEWSipnEMqud+RdImZubdDLN4A6t27d2oNRexZ3hmHc okSJEuulbznZcPDgQf0bIosWLfo/aVF/RkZGaGgo5fA5OICNuEOXoaTIk5mZKb+R0Qo43ciPrSDE zpUdoJHAJj6Rb3tY3BmGcRfo+5QpU2zGZwTXlS83GVbUmL69ABJHyfkBDmAj7gLNoU5LS4Nwa0sk DUDHkaqNn1itgg8EWNwZhvGMGjVqjB8/fufOnYY3NeXk5KSkpED9a9WqRVl1QHO+yQtcWMNrUXwM zsmluAP4znCoxW8Ae/bsQdMkxsQBwmIQRoCcVitqAgQWd4ZhvCc8PLx+/foNGzaEy2t4tvbmAwlz Iu4CSDb8cZsJ1YS8L3EMWFjcGYYp4EDCnIu7Bhx5iHh0dHRcXBz+5seweL7C4s4wTAHHM3EPdljc GYYJdtz+QDbDMAwT+LC4MwzDFEBcv/I3KysrPj6e7FuDTp06NWrUKHhf+cswDOMCiPstS7t27agU JL766itk4DF3hmECFhee+/PPP19D+S6+KeXLlx88ePCJEyc++eQTiioo5OTkvPfee2RIQNy7dOnC njvDMAWT6tWrw4HdvHkz2bcMvFqGYZgAhydUGYZhCiAs7gzDMAUQFneGYZgCiIsJVXuqV69++PDh AwcOvPLKKxR1a/Duu++2bNmSJ1QZhglYvBL3atWqHTlyhIxbj9KlS1+8eJEMhmGYQMIrcQexsbGG t87fIkDWR4wYQQbDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzD MAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMAzDMEwAUajQ /wNvCvocE80fWgAAAABJRU5ErkJggg== This value indicates the number of saves or revisions. The application is responsible for updating this value after each revision. DocumentLibraryFormDocumentLibraryFormDocumentLibraryFormRachel BransonAndrew Grogan22026-10-05T21:28:00Z2026-10-05T21:28:00Z2026-10-05T21:28:00Z034710840522Microsoft Office Word033795falseTitle1false47535falsefalse16.00000x010100ACE339D09936DB4EBD68BB5ACEDF8699