The Civil-Industrial Complex: AI and the Future of Public-Sector Expertise

Between federal fiscal years 2025 and 2027, the Gallatin Valley Metropolitan Planning Organization in southwest Montana programmed roughly $682,000 toward its first long-range transportation plan: about three-quarters of it consultant services, the rest some 1,600 hours of its two employees’ time and the overhead on it. The plan is federally required, the budget is public, the process is the one every MPO in the country follows, and the firm doing the work will almost certainly do it well.

Nothing about this is unusual, which is exactly why it is worth looking at. Institutional arrangements hide in the ordinary case.

Across the United States a large share of the technical work of government — comprehensive plans, transportation studies, zoning codes, environmental assessments, infrastructure inventories, capital programs — is done by private engineering and planning firms. Federal programs supply the funding, the requirements, and the procedures that generate the demand; the firms supply the engineers, planners, geographers, GIS analysts, and developers. Much of the work is valuable and some of it indispensable: a town of eight thousand cannot keep a hydraulic modeler on the payroll, and there are decisions where an outside analyst’s independence is the whole point. But step back from any single contract and a larger relationship appears. A substantial part of government has organized itself around the recurring purchase of technical services, the firms depend on the stable market that public funding and regulatory obligation create, and the arrangement sustains itself.

In his 1961 farewell address Eisenhower warned of the military-industrial complex.[1] He was not objecting to government buying equipment from private manufacturers, and he did not say the defense establishment was unnecessary. His warning was about the enduring relationship — the way a permanent arrangement between the military, industry, and public expenditure could accumulate influence beyond its proper place in a democracy — and about the alert citizenry needed to keep it in that place.

I think the civilian version deserves the same scrutiny, and I will call it the civil-industrial complex: the institutional and economic interdependence of public agencies, private professional-service firms, and the regulatory and funding structures that sustain recurring government technical work. The term is not an accusation of conspiracy or corruption. It names a self-reinforcing arrangement in which the ordinary operation of government continuously generates demand for external services, while reliance on those services discourages the growth of equivalent capability inside public institutions.

Artificial intelligence makes the question urgent. As AI lowers the labor, time, and specialist expertise needed to produce analyses, maps, software, and reports, the economic conditions that justified outsourcing begin to dissolve. The question is whether public institutions will notice, or will keep purchasing the same products through arrangements built for an earlier technology.

What a plan costs

The long-range transportation plan (LRTP) is a good place to look because every MPO has to produce one: federal regulation requires a plan with at least a twenty-year horizon, reviewed and updated on a fixed cycle.[2] Producing it means traffic counts, crash records, demographic and land-use analysis, demand forecasts, condition and revenue assessments, and a document that can run to hundreds of pages — and the consultant’s contract is only part of the cost. MPO staff gather and transmit data, coordinate with member agencies, manage the contract, review deliverables, run committee meetings, handle public comment, and carry the plan through adoption. The public pays for the whole institutional process, and the Gallatin Valley MPO’s Unified Planning Work Programs (UPWPs) make the scale visible:[3]

Federal fiscal year	LRTP work element	Of which consultant services	MPO staff hours
2025, as amended January 2025	≈ $273,000	$225,000	620
2026, as adopted → as amended	$351,450 → ≈ $372,250	$275,000 at adoption	770
2027, as adopted	≈ $37,090	$20,000	203
Total programmed	≈ $682,340	≥ $520,000	1,593

Two amendments sit inside those numbers, and both can be traced in the documents. The original FFY 2025 program carried $348,000 for the element, including $300,000 in consultant services; a January 2025 amendment moved $75,000 of that into the following year because the consultant’s work was running later than planned, without changing the contract price, scope, or schedule. The adopted FFY 2026 program’s $275,000 in consultant services includes that $75,000, and its funding comparison still lists FFY 2025 at the pre-amendment $348,000, so the shift shows; the total above uses the amended FFY 2025 figure and counts the money once. Then, during FFY 2026, the FFY 2027 program’s comparison column shows the element at $372,250 and the reserve at $59, where the adopted FFY 2026 program had $351,450 and $20,859 — a $20,800 reallocation from reserve into the plan. The MPO’s published link to the January 2025 amendment has already stopped resolving; not malice, just a website migration, which is how institutional memory usually goes.

These are programmed amounts, not audited expenditures, and they are not the consultant’s contract value. The element carries supporting activities that are not the document alone — though public involvement, data collection, and program administration are budgeted elsewhere and partly serve the plan, so the true cost is higher than this total, not lower. Whether the spending is excessive is a question I am not asking; answering it would take the scope, the deliverables, and the invoices. The point is how much institutional effort is organized around producing a periodic artifact.

One more fact about this case matters. It is the MPO’s first plan, from an organization with two employees — its work program budgets twenty-four staff-months a year — designated after the 2020 Census put the Bozeman urban area over the 50,000 threshold. It inherited no model, no data architecture, no analytical workflow from a previous cycle. It had, in principle, a clean choice between buying a document and building a system, and the procurement was organized, as nearly all of them are, around delivering a document: a draft due in September 2026, public review in October, adoption in November or December. Not because anyone erred, but because the template, the funding categories, the review expectations, and the procurement procedures all assume one. The work program gives a partial answer to the question this essay is about — whether the MPO comes out owning a model and a dataset it can run itself. Staff intend to fold the plan’s GIS layers into the MPO’s own system, and they note that hosting the regional travel model locally rather than with MDT is a question for the future. Those are the right instincts. Whether they survive the budget is the question.

Because the purchase is not a one-time event, and the work program says so. Its five-year funding projection carries the plan element at about $19,000 a year in 2028 and 2029, then $171,000 in 2030 and $179,000 in 2031: the next update, on the five-year schedule the regulations set. The same table projects the MPO’s carryover turning negative in 2030 and reaching about −$220,000 in 2031, and the program states that even after cutting memberships and out-of-state travel, operating costs are expected to exceed available funding by 2031 under the current structure. In the data element, staff propose returning the continuous count stations the MPO bought the year before, citing budget projections and partners’ installation of counters at signals, and a public-facing data hub or dashboard is listed for research and early development “as budget and staff capacity allow.” A performance-measure dashboard is being built with the consultant as part of the plan; who maintains it after the contract closes is not said. Nobody at the MPO is doing anything wrong. This is what the incentives produce: the next $350,000 plan purchase is already in the projection, the money to pay for it is not, and the instruments that would make the next plan cheaper to produce — the counters, the dashboard — are the items being returned or deferred.

The static document

There is something odd about the conventional process once you notice it. Transportation systems change continuously — population grows, subdivisions are approved, businesses open, traffic shifts, pavement deteriorates, crashes accumulate, and the investments the plan recommends alter the network — yet the information used to guide the next twenty years is assembled into a document that represents a single moment, and once that document is adopted its analytical content stays fixed until the next cycle. Some of the data are years old before the plan is finished.

So the traffic-volume map becomes a historical record, the demographic projections drift from the counts, proposed projects are built or abandoned, and the assumptions move away from the system they describe. The plan is not useless; it remains the authoritative statement of adopted policy. But the evidence under it ages from the day it is printed, and five years later much of the work is done again — new data, new maps, new projections, new technical reports, and even where the datasets and methods were kept, substantial coordination to produce the next version.

This made sense when geographic information was expensive to acquire, sophisticated analysis required specialists, and results travelled on paper. Those conditions have been eroding for thirty years; GIS, databases, cloud computing, and web applications have steadily reduced the effort needed to maintain, analyze, and distribute spatial information, and AI accelerates the trend sharply. The problem is not that long-range planning is unnecessary. It is that the static document may no longer be the right technological form for it.

A plan that does not stop

Imagine the MPO maintaining its plan as a living information system instead of reconstructing the region every five years. Network, volumes, crash records, demographics, transit, land use, and infrastructure condition live in one continuously maintained spatial database. New observations enter as they arrive and pass validation; history is preserved; workflows are documented; maps and charts are generated from versioned data rather than drawn once for a PDF. A resident picks an intersection and sees its crash history, traffic trend, and any proposed improvement. A planner compares demand under different growth scenarios. A commissioner asks how a project affects safety, access, condition, and the budget, and gets an answer from current data rather than from an appendix adopted three years ago.

AI does real work in a system like this. Agents acquire and clean data, build and maintain the processing pipelines, keep the software running, flag inconsistencies, draft the explanatory narrative, and answer questions against verified datasets. A planner asks for the corridors where volume is rising and crash rates are elevated, and the system assembles the data, runs the analysis, documents its method, and returns a map.

Professional expertise does not disappear. Models still have to be evaluated, data validated, trends interpreted, and policy made through accountable public processes. What falls is the human labor spent building and maintaining the information infrastructure — and that is the institutional change that matters more than the automation of any single task. The MPO becomes the permanent custodian of a growing body of knowledge about its transportation system instead of the periodic purchaser of a snapshot.

I spent years on the agency side of these arrangements — running GIS for a county government, working as a software engineer, as well as regional and local planning — and the pattern described here is a familiar one. None of it should be taken on faith, though. Here is what the difference between a document and a living system looks like in a finished plan, with the one layer of that system I have actually built.

Take the Missoula MPO’s 2025 Missoula Connect update — a finished plan from an organization that does this well — and follow its crash data through the process.[10] The project team of MPO staff and consultants obtained five years of MDT crash records, 2019 through 2023, and built the safety analysis in the existing-conditions report in Appendix C: a trend chart back to 2008, a severity comparison showing that more than 70 percent of crashes involving people walking or biking caused an injury or death against less than a quarter of vehicle-only crashes, two full-page maps of fatal-and-serious-injury and pedestrian-and-bicyclist crash locations drawn as graduated dots, and the three leading human contributing factors. That analysis did real work: crash locations were a key input to project scoring, and the plan reports the five federal safety performance measures against the state’s targets. A summary chart made it into Chapter 2 of the plan itself. The rest sits in a 340-page appendix, and the 2023 records, which MDT had not yet verified when the maps were drawn, are now the plan’s last word on safety until the next update.

The public process ran on a separate track: three phases of engagement across 2024 — sixteen tabling events, a survey with 640 participants, an open house, neighborhood meetings, and an interactive online map where residents commented on and voted for proposed projects — then a 45-day comment period on the draft and a written disposition of every comment in Appendix H. On the IAP2 Spectrum of Public Participation — the framework the Gallatin Valley MPO’s own work program uses to evaluate its outreach — it is a thorough Inform-and-Consult process, with advisory groups that reach toward Involve.[11] But notice the shape. The crash data informs the public as finished graphics; the public consults on a list of projects; the evidence and the conversation never occupy the same place, and both stop when the plan is adopted. The plan’s performance chapter says as much, with real candor: if resources become available, the MPO may develop an online dashboard to make information available to the public and visualize progress between federally required updates. Gallatin Valley’s work program, two hundred miles east, attaches the same condition to the same idea.

That dashboard is not a large project. I built the Montana Crash Explorer from the same public MDT crash data the plan used: 219,480 crashes statewide, 2016 through 2025, on a map that filters by year, severity, county, hour of day, collision type, and contributing factor, with charts that recompute as the filters change and a link you can copy to share any view.[12] A Missoula resident can reproduce something close to the plan’s Figure 4-10 — fatal and serious crashes, 2019 to 2023 — and then do what the figure cannot: zoom to Mullan Road and keep only the night hours, look at ten years instead of five, or check whether 2024 and 2025 look like 2023. That is the Inform function of the appendix, running continuously, two years past the plan’s data cutoff, as a site rather than a page. Adding a comment layer, so that a resident can flag an intersection and say why, would bring the Consult function into the same place, and that is an ordinary feature to build rather than a research problem. What the explorer does not do is also the point. It does not decide which corridors matter, score projects, or set targets. It keeps verified evidence current and in front of the planner and the public, so that professional judgment is spent on judgment instead of on redrawing the map. A crash layer is one layer of what a living plan would hold; the fuller version is the subject of a later piece.

Why the arrangement persists

Why do agencies rely on consultants so heavily? The standard answer is the uneven distribution of expertise, and it is a good answer as far as it goes. A small government cannot keep every specialist that complex projects require, so firms concentrate specialized labor and spread its cost across many clients. A town that needs a hydraulic model once every twenty years should not employ a hydraulic modeler.

The logic weakens for recurring work. Data maintenance, map production, spatial analysis, performance monitoring, and technical reporting come around every year, and until recently even these demanded expensive software, long training, and a great deal of staff time. AI changes the calculation. An experienced planner or GIS professional with capable AI tools can do work that used to take several specialized people, the marginal cost of one more map or analysis falls, and the advantage of concentrating that labor inside a large firm shrinks for whole categories of work. The consultant’s role shifts toward specialized engineering, complex modeling, independent review, and the unusual problem — the work it is still uneconomic to keep in-house. The boundary between what must be bought and what can be kept moves, and that boundary is exactly where the civil-industrial complex lives.

Nothing guarantees the boundary moves on its own, because the arrangement reproduces itself without anyone behaving badly. A small agency’s staff carry a long list of obligations, budgets are tight, and a permanent position is politically difficult, administratively slow, and often barred by the funding source. A consulting contract solves the immediate problem cleanly: a defined product, no long-term commitment, the obligation met. If the agency comes out of it without the knowledge, the software, and the documented workflows to repeat the work, the next cycle starts where this one did, and over time the arrangement hardens — staff expertise concentrates in contract administration rather than analysis, firms accumulate experience with the recurring requirements, and officials learn to buy deliverables rather than maintain capacities. Everyone acts rationally and the whole perpetuates a division of labor that may no longer be efficient. That is path dependence, and the longer it runs, the harder it becomes for the agency even to know which capabilities it ought to have.

Who produces the knowledge that governs?

Cost is the smaller issue. Government decisions rest on specialized technical knowledge — forecasts, performance analyses, demographic projections, environmental studies, engineering assessments, financial models — and consultants produce a great deal of it. Elected officials keep formal authority, but whoever produces the evidence shapes which choices look reasonable. In a corridor study, the alternatives someone chooses to define, the forecast assumptions, the performance measures, and the financial constraints all steer the policy discussion before anyone votes. A model does not reveal an independent future; it encodes beliefs about how the world works and returns conclusions consistent with them, and no one has to manipulate anything for that to happen. Government analysts carry their own assumptions, so this is not a complaint about private firms in particular. It is a point about accountability: an agency that relies on outsiders to define its problems, build its models, and interpret the results becomes dependent not only on external labor but on external expertise to understand the consequences of its own decisions.

This is not a hypothetical concern. In 2007 the Government Accountability Office examined support-services contracting at the Department of Homeland Security and found that more than half of the 117 statements of work it reviewed involved services closely supporting inherently governmental functions, and that in nine case studies officials had not assessed the risk of decisions becoming dependent on contractor judgment.[4] Federal policy has circled that finding ever since; I have traced the record in Twenty Years of Buying Public Expertise.[5] The short version: government recognized the problem, wrote it into policy in 2011, and its auditors were still finding the same gaps in 2024 and 2026.[6][7]

Government should be able to interrogate independently the assumptions, methods, and conclusions behind the technical information it uses. AI could strengthen that capacity by giving agency staff the means to reproduce an analysis, test a model’s assumptions, and validate a consultant’s work. Or it could add one more layer of dependence. Which happens is a choice about deployment, not a property of the technology.

The broken hammer

There is a philosophical way to see what AI is doing to professional routine, and I find it clarifying. In Being and Time Heidegger distinguishes equipment that is ready-to-hand (zuhanden) from objects that are present-at-hand (vorhanden).[8] A carpenter absorbed in the work does not regard the hammer as an object with properties; the hammer disappears into the hammering. When it breaks it becomes conspicuous, and in becoming conspicuous it reveals the whole arrangement of tools and purposes that was invisible while everything worked.

A transportation planner inside an established system treats the technical appendix, the consulting contract, the periodic report, and the GIS as ordinary parts of the job — the equipment through which planning gets done, invisible in the doing. But the purpose of transportation planning is not a document. The purpose is to understand a transportation system, anticipate its future, evaluate alternatives, and guide public decisions; the document, the contract, the software, and the model are instruments. AI makes those instruments conspicuous. When software can generate the maps, run the geoprocessing, build the application, and draft the narrative with little direct human intervention, it becomes impossible not to ask which parts of the conventional process are necessary. Why spend months assembling a static existing-conditions appendix if the same information can be kept current? Why organize a continuing public responsibility around a periodic production event? Heidegger was not offering a theory of technological progress, and I am not claiming new tools automatically supersede old ones. The point is that disruption exposes the machinery to scrutiny, and scrutiny lets us separate the enduring purposes of government from the historically contingent means we have used to pursue them.

The geographer David Harvey called the technological shrinking of the world time-space compression: transportation and communication keep reducing the time it takes to overcome distance, and the reorganization of production that follows redistributes economic activity and power.[9] AI compresses a different distance — between an analytical question and its answer — and the effect is geographical in Harvey’s sense when it changes who can do the work and where. A two-person MPO in Montana can now do analysis that previously required a large metropolitan agency or a Denver consulting firm; a rural county can maintain a serious land-use and infrastructure information system without a technical staff; an independent geographer can build what once took a development team. The geography of expertise is what is changing, and that is the economic foundation of AI’s disruption of the civil-industrial complex.

Two objections

The strongest objection is independence. A consultant’s analysis carries weight precisely because the consultant is not the agency, and a plan produced in-house is easier to attack as the staff’s preferred answer dressed up in numbers. This is real, and the response is to separate the functions rather than deny the problem. Keeping the data, the models, and the workflows inside the agency does not preclude buying independent review of them. It makes independent review cheaper and more meaningful, because the reviewer can rerun the analysis instead of taking the appendix on trust. Independence is a property of review, not of who owns the database.

The second objection is turnover. Small agencies lose people constantly, and the failure mode of in-house technical systems is well known: the one person who understood the platform leaves, and nobody can run it. That is the consultant critique mirrored — dependence on one employee instead of one firm — and it is why documentation, interoperable formats, and reproducible workflows are not refinements but the whole design. A living system is an institutional asset only if a new hire can inherit it. If the platform cannot survive the departure of its builder, the agency has not built capacity; it has hired a consultant with a W-2.

What it would take

The shift begins with a different philosophy of investment: buy the capacity to produce the product, not the product. For an MPO that means a permanent transportation information platform with five properties.

1. Public ownership and control of the data, methods, models, and documentation, in interoperable formats, reproducible without any particular vendor.
2. Continuous maintenance, with new counts, crash records, and demographic estimates entering the system as they are validated — and no change to adopted policy implied by a change in the data.
3. AI-assisted analysis and reporting, with every significant product carrying its sources, assumptions, processing history, and validation.
4. Formal decision-making kept distinct from the evidence. Participation, financial planning, interagency coordination, and adoption procedures stay where the law puts them, and adopted policies remain identifiable as decisions.
5. Continuity across cycles, so new staff inherit documented data and workflows instead of reconstructing the agency’s technical foundation.

Consultants still contribute where specialized expertise is needed. The agency simply retains the ability to understand and maintain its core functions.

Federal rules allow this. 23 CFR § 450.324 requires MPOs to review and update their plans at fixed intervals, permits revisions between updates, and provides for electronic availability; § 450.316 explicitly contemplates visualization and electronic access to planning information.[2] Nothing requires the analysis to be frozen between adoptions. The distinction that matters is between the platform and the plan: observed conditions, models, and performance measures can be updated continuously, while adopted policies, the fiscally constrained program, and other legally significant elements still require formal review and approval. A small MPO could build a living existing-conditions platform now and evaluate it against the conventional process.

It should be evaluated, because AI is not a universal solvent. These systems make mistakes, produce unreliable output, and create dependencies of their own, and a badly designed internal platform can consume resources and return nothing. So the right experiment is a pilot: a conventionally produced existing-conditions appendix against an internally maintained, AI-assisted system, compared on full cost — consultant fees, staff hours, subscriptions, hosting, maintenance, quality assurance, administration — and on usefulness. How fast can new data be incorporated? Can the analysis be reproduced? Can staff revise the method without outside help? Can the public see the evidence behind a recommendation? How much institutional knowledge remains when the cycle ends? The static report will probably be cheaper in year one. The proper unit of comparison is not the cost of one plan but the lifetime cost and public value of maintaining the capacity to plan.

The other dependency

There is an irony here. The technology that could reduce an agency’s dependence on engineering and planning consultants is itself supplied by private firms, and by a small number of very large ones. An MPO that no longer needs an outside GIS analyst may instead depend on a proprietary model, a cloud platform, and a stack of subscriptions. Prices rise, interfaces change, services are discontinued, formats lock, and sensitive public data flows through systems the agency does not control. Dependence would not disappear; it would be reorganized.

The lesson is not to avoid AI but to distinguish technological capability from institutional autonomy. Government need not own every computational system it uses, but it must keep control over its information, its methods, its adopted decisions, and its ability to change providers without losing what it knows. Open standards, portable data, documented workflows, reproducible analyses, independent validation — the same list as before, because it is the same problem. Swapping an AI vendor for an engineering consultant would leave the civil-industrial complex intact under a new name.

Why we buy the report

The relationship between government and consulting grew up when expertise was expensive, software required long training, and small agencies could not do complex professional work themselves. Consulting was a sensible response to those conditions, and AI is dissolving them — compressing the time between recognizing a public problem and producing the analysis to address it, and making it practical to maintain continuously what could once only be produced periodically. In doing so it exposes how much of our institutional arrangement reflects old constraints rather than enduring necessity. The question was never only whether government spends too much on consultants. It is whether government has become structurally dependent on purchasing its own capacity to understand and act. The long-range transportation plan shows it clearly — the purpose is understanding, the document is a means, and we have built an elaborate institutional process around the means — and the same is true of land-use planning, environmental assessment, hazard mitigation, and asset management, every field where government repeatedly buys a representation of something that never stops changing.

None of this guarantees a better public administration; the benefits depend on how the technology is deployed, how its results are evaluated, and who controls the systems through which public decisions are made. Eisenhower’s warning was not that government and industry should never cooperate. It was that an enduring relationship between them acquires momentum of its own, and only an informed public keeps it serving democratic ends. The same holds here. The most consequential thing AI brings to government may not be cheaper versions of the services it already buys. It may be the realization that some of those services can become capabilities of government itself — and, once that is possible, the question of why we organized so much of government around buying them in the first place.

References

[1] Dwight D. Eisenhower. Farewell Address to the Nation, 17 January 1961. National Archives, Milestone Documents.

[2] 23 CFR Part 450, Subpart C — Metropolitan Transportation Planning and Programming, § 450.316 (interested parties, participation, and consultation) and § 450.324 (development and content of the metropolitan transportation plan). eCFR.

[3] Gallatin Valley Metropolitan Planning Organization. Unified Planning Work Program (landing page). FFY 2025 UPWP, Amendment No. 1, TPCC, January 2025 — the MPO’s published link to the amendment no longer resolves; the FFY 2025 figures are from the amendment memorandum as posted at the time. FFY 2026 UPWP, approved by TPCC 27 August 2025 and by MDT and FHWA/FTA 5 September 2025: work element 300 at Tables 11–13; funding comparison, including the pre-amendment FFY 2025 figure, at Table 26. FFY 2027 UPWP, approved by TPCC 22 July 2026 and by MDT and FHWA/FTA 1 September 2026: work element 300 at Tables 11–13; funding comparison at Table 24; five-year funding allocation projections at Table 27; data element 302 and public involvement element 102 for the count stations, the data hub, and the IAP2 reference.

[4] U.S. Government Accountability Office. Department of Homeland Security: Improved Assessment and Oversight Needed to Manage Risk of Contracting for Selected Services. GAO-07-990, September 2007.

[5] Twenty Years of Buying Public Expertise. Companion post; full source list there.

[6] U.S. Government Accountability Office. VA Acquisition Management: Oversight of Service Contracts Needing Heightened Management Attention Could Be Improved. GAO-24-106312, 25 January 2024.

[7] U.S. Government Accountability Office. Artificial Intelligence Acquisitions: Agencies Should Collect and Apply Lessons Learned to Improve Future Procurements. GAO-26-107859, 13 April 2026.

[8] Martin Heidegger. Being and Time (1927), trans. John Macquarrie and Edward Robinson. Harper & Row, 1962. §§ 15–16 on readiness-to-hand and the conspicuousness of broken equipment.

[9] David Harvey. The Condition of Postmodernity: An Enquiry into the Origins of Cultural Change. Blackwell, 1989. Part III, on time-space compression.

[10] Missoula Metropolitan Planning Organization. Missoula Connect 2050 Long-Range Transportation Plan Update, Final Plan and Appendices A–J, 2025. Crash analysis in Appendix C (Figures 4-8 through 4-11); safety performance measures in Appendix G; engagement in Chapter 3 and Appendix A; disposition of public comments in Appendix H; the dashboard remark in the Final Plan’s performance management chapter.

[11] International Association for Public Participation. IAP2 Spectrum of Public Participation, 2018.

[12] Episteme Geospatial. Montana Crash Explorer. Interactive map of MDT public crash records, 2016–2025.