Finding out what a patient can actually get still means making phone calls.

Nearly every community has a resource directory. The difficulty is not building one — it is that keeping it accurate is continuous work, and that work is rarely funded as a permanent role.

• Sector standards for community-navigation resource databases are built around an annual verification cycle, with procedures for handling changes in between. The between-cycle work is the part that costs staff time. (Inform USA, Standards & Quality Indicators for Professional Community Navigation, v10.1)
• Open Referral's research in Washington, D.C. estimated that roughly 500 organizations providing about 1,500 services would need approximately two full-time equivalents to be reliably updated every six months. That is the real cost of accuracy, and it is a salary line rather than a software line.
• Directory inaccuracy has been measured carefully in adjacent settings. When U.S. Senate Finance Committee staff called listed in-network mental health providers, reaching ten listings took one to three hours, and a third of the numbers were wrong, disconnected, or never returned. That study looked at health-plan directories rather than community resources, so we treat it as the nearest available measurement rather than a description of your week. (Senate Finance majority staff, 2023)
• In a survey of critical access hospitals conducted during the pandemic, 92% (n=127) selected finding a hospital to take a transfer, or general issues with bed availability, as their single largest barrier to outbound transfers. The same respondents named hospital relationships and care coordination as what helped most. (Flex Monitoring Team, February 2023; 155 of 404 sampled, 38% response)

Part of that last finding is about supply rather than information, and no software creates supply. We think the part that is genuinely an information problem is worth working on, and we try to be careful about the difference.

This is a long-standing problem, and most of what is known about it was learned by 211 centres, information-and-referral professionals, statewide referral networks and state bed registries. Their work is the reason the difficulties are documented at all, and we have relied on it heavily.

• Referral and receipt are different measurements. A published evaluation of one statewide referral platform reported a closed-loop rate near 99% while the share of cases producing an actual service was between 38% and 65%. The same program's resolution rate moved with the availability of funding to reimburse community organizations, not with any change to the software.
• Connecting people to services is harder than it looks. The largest randomized evaluation in this area, run by CMS, found that navigation did not significantly increase connections to community services or the resolution of needs compared with a randomized control group. Its evaluators attributed the effects they did observe to relationship and trust-building work.
• Availability information decays quickly. At least seventeen states have operated behavioral health bed registries, and no published evaluation has shown one improved access. The recurring difficulty is that manually maintained availability goes out of date within hours, and that an open bed is not the same as an accepted patient.
• Volunteer maintenance has not proven durable. Open Referral's assessment of crowdsourced directories found few examples of sustainability — not because contributors were inaccurate, but because the work concentrated on one or two committed people and stopped when they moved on.

We take these as constraints rather than as failures. They are the reason we are not building a referral network or an availability feed, and the reason we would rather test our assumptions with a few organizations than assume they hold.

A resource finder is built and running in the PHapp® research preview. It is not generally available. The resources it holds today are those a public health emergency calls for — cooling and warming centres, naloxone and vaccination sites, shelter, food, hotlines.

• Records carry a verification date where one exists. When an organization confirms a record, we store who confirmed it and when. When a county's published feed still lists it, we store that the feed still listed it, which is a weaker statement and is kept distinct. Records whose confirmation has lapsed are marked rather than presented as current.
• Whether a place is open is determined on the server, from published hours, dated closures, season and the site's own time zone rather than the reader's. When hours have not been published, the answer is recorded as unknown rather than as closed. Because that determination is made once, in one place, any surface reading it receives the same answer.
• Corrections persist. When a person corrects a record that also arrives from an automated feed, that correction is kept and continues to apply to that field, while the feed goes on maintaining the fields it maintains well.
• Changes are recorded. Each change writes an append-only entry with the previous and replacement values, so a bad bulk update can be traced and undone.
• The data can be exported in Open Referral HSDS under a partner credential, so other systems a community runs can read it. The export carries names, coordinates and service descriptions — not street addresses, phone numbers or weekly hours, which a consuming system would need to supply or look up itself.

The finder's own screens are still being localized, and the resource listings themselves have no translation path yet, so both are English today. There is also no way to reach resource information by text message; the finder is reached by app or web.

These are the hard parts. Some of what follows is designed and not yet built; where the evidence points a particular way, we say so. We would rather describe the problem honestly than the ambition.

• Matching a situation to services that fit. Stating who a person is and getting back the places that can take them. Designed, not built. The evidence here is genuinely mixed — the largest randomised evaluation found navigation alone did not increase connections — so the question we are working on is what has to be true for matching to help rather than whether to offer it.
• A surface built for professionals. The finder today is a resident surface. What a coordinator needs is a different screen, and it does not exist yet.
• Knowing whether a referral became a service. The hardest measurement in this field: published evaluations report closed-loop rates far above the share of cases that produced an actual service. We would rather solve knowing than produce a number that flatters everyone, so we have not built a referral rail — that is a sequencing judgement, not a permanent refusal.
• Availability that stays true. Capacity changes hourly and decays faster than it can be maintained by hand; at least seventeen states have operated behavioral health bed registries and no published evaluation has shown one improved access. Our working answer is standing conditions that change monthly, which a small team can realistically keep accurate. If someone finds a way to make live availability hold, we would want to learn it.
• Being clear about what a listing promises. Published conditions describe a stated rule. Admission remains the organization's decision, and no surface of ours should imply otherwise.
• Doing it without holding patient information. The fields under design describe what a resource requires — county served, age range, insurance accepted, language, walk-in or referral. No name, date of birth, diagnosis or free-text field, and a stated situation is a query rather than a stored record. That constraint is written into the specification before the code exists, and we are publishing it so it can be held to.

Three assumptions could turn out to be wrong, and we would rather name them.

• Whether maintaining a listing is genuinely a few minutes a month. If it is not, this approach carries the same staffing cost as every other approach, and the honest description of what we offer would narrow considerably. This is the first thing we want to test, and it needs no software to test.
• Whether contributors keep contributing. The recorded difficulty with volunteer maintenance is not inaccuracy but attrition. Weighting contributions decides whose information is used; it does not by itself give anyone a reason to contribute. Some combination of accountability, compensation, or a funder requirement is likely to be necessary.
• Whether structured eligibility can be collected at all. County feeds rarely carry it. If organizations do not publish it, the honest answer for most records will be that eligibility is unknown, and the design has to handle that as a normal case rather than a gap.

Funding notices name capabilities in a fairly standard vocabulary. The answer differs term by term, and is set out below so you can decide quickly whether we belong in your application.

• Resource directory, resource inventory. This is the part that exists, in a research preview rather than generally available. Records of local services, each with a verification date where one exists. Where a site has published its hours we determine open or closed on the server; on the current corpus most records have no published hours, and those are shown as unknown rather than as closed.
• Community health worker support. Partly. A CHW needs to know what is currently open and who each service can take. The first is the layer we work on. The second is unbuilt — there is no schema and no vocabulary yet for who a service can take, and no way to match a person to it.
• Real-time resource directory. We do not use this phrase and would not write it into a proposal. A directory is only as current as its last confirmation, and the useful thing to state is who confirmed a record, by what method, and how long ago. That is what we record.
• Community information exchange. Not itself an exchange. The HSDS export described above is what an exchange your community runs would read, and its contents are worth checking against what yours needs.
• System navigation, care navigation. This is a human role, and we do not replace it. The layer that would let a navigator state a situation and get back services that fit is designed but not built.
• Closed-loop referral, referral management. Not built. The section above explains the sequencing — it is a hard measurement problem we have not solved, not a door we have closed.
• Screening for health-related social needs, and referral from screening. No screening instrument, and no patient record by design — that constraint is deliberate and we intend to keep it.
• Bed or appointment availability, capacity dashboards. Not built. Our working answer is standing conditions that change monthly rather than capacity that changes hourly; see above.

Where the answer above is not yet, that is what it means — not yet. If one of these is the capability your funding names, tell us; what a notice actually asks for is the most useful thing we can learn right now.