GLIDE Foundation

What do they do? GLIDE provides food, health access, family services, crisis support, housing assistance and advocacy in San Francisco. More

Why this approach interests us

• A trusted, low-barrier organization can connect people with care and support they would otherwise miss.
• Its combination of practical assistance and health services can address several obstacles facing the same person.
• Treatment access, housing support and overdose-response networks provide concrete pathways to better health.

Our main reservations

• Service counts do not show how much additional completed care or housing stability GLIDE causes.
• The model is sensitive to survival and service-duration assumptions, and current additional funding capacity is not verified.

What do you get for your dollar? Our central health estimate is about $4.89 million per 10 Bay Area QALYs for support to the Foundation. It covers several health pathways but not all of GLIDE's community, youth and policy benefits; the earlier rental-assistance-only estimate is not the return on a general donation. More

Decision, recipient and the limits of a whole-gift health estimate

This update evaluates an ordinary unrestricted donation to GLIDE Foundation, EIN 94-1156481. That choice matters before any arithmetic. GLIDE's website presents both the Foundation's service work and GLIDE Memorial Church. Its giving page separates their donation routes. A person choosing the Church's route is not purchasing the modeled Foundation portfolio. Conversely, consolidated accounts that include Church and property activity are useful context but do not identify the exact destination of a Foundation donor's next dollar.

The phrase “whole gift” describes the cost boundary, not completeness of benefits. Every dollar of the hypothetical gift remains in the numerator, including amounts assigned to management, fundraising, policy and affiliate-related activity. Health is quantified only for named mechanisms with an explicit bridge from spending to finite additional outcomes. The missing mechanisms have unknown value. Calling them zero would be unjustified; quietly awarding an unspecified residual would be equally unjustified. This leaves a partial-health diagnostic, not a comprehensive whole-organization expected value.

The preserved central model asks what a $100,000 tranche might accomplish if the specified allocation, delivery costs, additionality and clinical assumptions held together. Its Bay health total is 0.2044715224666835 QALYs and its donor price is $4,890,656.595775775 per 10 Bay QALYs. These decimals reproduce the executable model; they are not a claim to measurement precision. A sensible reader should retain the order of magnitude and the dependencies, not attach confidence to the last digit.

There is no assigned probability distribution over the scenario set. We therefore report no weighted expectation, no probability of beating a threshold and no favorable-tail contribution percentage. Other reports in this project sometimes contain explicitly subjective signed mixtures. This one does not become more comparable by inventing weights after seeing its scenarios. Its central value and adverse or favorable diagnostics must retain their names.

The scope also explains why an older rental-assistance calculation cannot be substituted for this result. A conditional rental model asks about a specific assistance pathway, often tied to a historical transfer-oriented cohort. An unrestricted Foundation gift may support meals, clinical access, youth services, family activities, administration or other work. Even if a particular housing tranche looked inexpensive, the donor would need a credible route to that tranche and its fully loaded incremental capacity. Ranking a general donation using a favorable restricted-program calculation would hide a material choice.

This update preserves the accepted broader model rather than rerunning that favorable selection. It expands financial and operational evidence and adds diagnostics. It does not tune clinical coefficients to recover the old ranking. The updated Foundation financial series is not silently substituted for the consolidated figure inside every plausibility note: historical accounting costs, hypothetical marginal allocation and donor gift size play different roles. Where the new finance evidence supports a distinct allocation stress, that stress appears separately.

Our recommendation status is HOLD for someone whose overriding goal is buying verified low-cost incremental Bay health. This is not a recommendation against supporting GLIDE for broader reasons. It means that the public record does not yet establish either an attractive complete health expectation or an actionable marginal offer. A donor concerned with dignity, homelessness, community institutions or policy may accept uncertainty that a narrowly health-maximizing ranking cannot resolve.

There are several possible ways further evidence could change this assessment without altering its ethical frame. A current budget could show that unrestricted gifts predominantly sustain a high-value clinical access team rather than the assumed mix. A treatment registry could demonstrate that people referred through GLIDE start and remain in care substantially sooner than they otherwise would. A documented expansion might create new capacity with public clinical resources already secured. Alternatively, evidence of substitution, low retention or fixed service capacity could weaken the case. The model makes these routes legible; it does not decide them from the organization's mission.

For publication, the title, research index and calculator should consistently describe “whole Foundation gift; named-pathway health diagnostic.” The central price should never be labeled an all-benefit expected return. The null weighted fields are intentional, not missing values to fill with the central number. Likewise, Bay is the primary geography for this comparison; the San Francisco-specific number is a separate narrower estimate. These presentation distinctions are part of substantive correctness because they determine what donation choice a reader thinks the model supports.

What do they do?

GLIDE's publicly described portfolio is substantially broader than a single clinic or assistance fund. Its services combine basic needs, practical navigation, direct support and partnerships with clinical providers. This combination could matter because a technically effective treatment is not useful to someone who cannot reach it, complete an intake, maintain contact or feel safe returning. But the existence of an integrated service setting does not demonstrate the incremental clinical outcome of every activity within it.

The FY2025 impact report records medication-assisted treatment enrollment alongside testing, meal provision and housing-related support. The important clinical anchor is 56 people enrolled in MAT, not the far larger supply-distribution count. Even enrollment is an intermediate outcome: the report does not supply a public deduplicated longitudinal file showing medication type, treatment duration, retention, alternative access and subsequent health for those individuals. We use it as an operational plausibility anchor, not as 56 additional lives saved.

The HEAT page describes low-threshold health access and partner-delivered care. It identifies public Street Medicine involvement in medication starts and university clinicians in wound care and related services. These are valuable complements, not free inputs that may be ignored in an economic assessment. Publicly financed prescribing, medication, laboratory work and clinician time belong in the counterfactual and in a broader resource account where attributable. The Foundation may contribute trusted access and coordination without financing the entire treatment.

Current web pages contain mixed vintages. Some service descriptions are written as ongoing schedules while nearby text still describes a future launch in fall 2025. A listed timetable is evidence of an advertised service, not proof that every session has run or that a new donor can buy an additional place. We do not turn these inconsistencies into a claim that services are closed. We also do not smooth them away by treating all page content as a synchronized current operational record.

One specific limit is clearer: the HEAT page says the hepatitis C peer community-navigation program is paused. Other testing and general linkage descriptions remain present. The appropriate conclusion is narrow. We should not model the paused peer program as operating, but neither should we infer that all HCV-related access has ceased. The model's small linkage pathway is explicitly not the full effect of an intensive colocated treatment program or a presumed peer-navigation restart.

The Welcome Center advertises rental assistance and case-management access. The program page indicates process requirements rather than an unrestricted entitlement to assistance. Published hours and workshop requirements establish how a prospective participant may approach the service. They do not specify the number of currently unfunded qualifying households, the amount needed to support them or whether another agency would provide the same assistance if GLIDE did not.

A materially newer operation is the Transitional Age Youth Health and Wellness Center, advertised as a 24-hour service for San Francisco young adults ages 18–27 facing homelessness or housing instability. A GLIDE account of its ribbon cutting places the opening in January 2026. The service page identifies City homelessness services, BMO, the Human Rights Commission and partners as supporters. Its public description includes clinical and practical support, but no completed-care denominator suitable for an incremental health estimate.

This new center changes the interpretation of historical spending. FY2025 ended before the opening, so its functional mix cannot be assumed to describe today's full portfolio. It also changes what the unquantified residual contains. Youth services may have substantial benefits, but we lack a current cost allocation, participant overlap map, additional treatment course count and no-gift alternative. We keep the center visible in the report rather than pretending the older modeled branches exhaust current GLIDE.

Annual service counts need their own definitions. The original FY2025 return identifies 571 people receiving a combined 2,041 HIV, STI and hepatitis C tests. A person may have more than one test, and a test may be negative, repeated or part of care available elsewhere. Neither number is a count of new diagnoses, treatment starts or cured infections. The same caution applies to naloxone: 5,168 doses are not necessarily 5,168 kits, people, rescue events or people whose annual fatality risk changes.

Meals expose a concrete reconciliation problem. The impact report describes 620,513 total meals, including 122,854 to-go meals; the original return narrative appears to present the two categories additively. We do not add them. This discrepancy is not used to allege mismanagement or to manufacture a more favorable cost per meal. It is a reason to request a stable denominator before inferring a productivity trend. The current clinical model does not multiply either reported total directly by a health coefficient.

The same discipline applies to family and behavioral support. Families receiving a mixture of childcare, classes, groceries or vouchers are not completed psychotherapy courses. Group participants are not necessarily unique across sessions or enrolled in a manualized treatment protocol. Referrals are not completed care. An organization can be useful through all these activities while still lacking the specific public denominators needed for a causal cost-effectiveness estimate.

Our operational reading is therefore cautiously positive about access mechanisms and cautious about measured outcomes. GLIDE appears to offer real points of contact with potentially effective care. Public materials do not yet quantify how often those points of contact change the timing, persistence or quality of treatment relative to existing options. That distinction, rather than the sheer size of the reach numbers, drives the model and the next evidence priorities.

Spending breakdown

The latest three Foundation returns cover fiscal years ending June 30, 2023, 2024 and 2025. They should not be confused with calendar-year activity, the separate Church returns or consolidated audited accounts. The official financial disclosures page provides both types of documents. This review uses the Foundation returns for the recipient-specific series and the final consolidated audit for broader balance-sheet, resource and liquidity context.

Sources: original FY2023 return, FY2024 return, and FY2025 return. “Whole” here adds separately netted costs to the functional total; it is not a statement of pure cash disbursements.

The extra columns prevent a common understatement. The return's functional expense total does not include every cost netted against revenue elsewhere on the form. For FY2025 the functional total is $28,937,817, while adding $505,300 of direct event costs produces $29,443,117. Using only the smaller number would omit real organizational activity. Using the larger number and then adding the same event costs again would double count them. The portable finance table reconciles these components in tests.

Approximate shares of the whole expense boundary are useful for interpreting scale:

These are accounting classifications, not rankings of useful versus wasteful spending. Administration can sustain payroll, compliance, technology, safe facilities and coordination. Fundraising may maintain future services. The model nevertheless keeps these costs because a donor cannot assume that the whole gift becomes direct clinical spending. Equally, an accounting program ratio does not reveal the marginal allocation of an unrestricted gift.

Within program services, the three principal named return categories and the remainder are:

The decline in the reported meals category is not sufficient evidence of improved efficiency. It could reflect changes in cost allocation, delivery mix, recognition or operations. The service denominator also needs reconciliation. We therefore do not divide each year's reported category by superficially comparable meal counts and claim a trend in marginal price. A current program-cost schedule that reconciles to the return would be more informative.

Revenue totals are $18,761,101, $26,742,300 and $28,318,144 for the three years respectively. A further classification issue is important: the FY2025 Schedule O explains that government contracts were moved from program-service revenue to government grants to align with the IRS definition. A large movement between those lines is therefore not evidence of a comparably large new public funding stream. The return reports $10,963,932 in government grants for FY2025, but comparisons with the earlier program-revenue line require the reclassification explanation.

The same accounting caution applies to noncash contributions. The return reports $699,849 of noncash contributions. That number should not simply be subtracted from whole expenses to manufacture a “cash cost” denominator. Recognition, timing and tax-versus-GAAP scope may differ. This report calls the series whole accounting expense. It does not claim that every expense was paid in cash during the year, nor that unrecorded volunteer services were economically costless.

The final FY2025 consolidated audit presents a different organizational boundary. Its total expense is $30,766,602, including $18,342,460 of program expense, $1,922,106 of Church expense, $7,766,659 of administration and $2,735,377 of fund development. These figures are not interchangeable with the Foundation-only return. The impact report contains figures labeled pending final audit; we use the final audited values for consolidated context instead of those preliminary figures.

Natural expense categories answer a different question from functional categories. In the consolidated audit, personnel accounts for approximately $21.51 million, with other costs including food, occupancy, depreciation, technology, outside services, client assistance and events. Personnel cuts across program and supporting functions. Adding the natural-expense table to the functional table would double count the same spending. For this assessment, the natural categories indicate the importance of staffing and infrastructure, while the functional series supports the historical allocation diagnostic.

The audit's $5.91 million cash balance is historical, not a statement that $5.91 million is available for an incremental donor-selected service. Assets include property, receivables and investments; restrictions and board designations also matter. The consolidated liquidity note identifies about $22.25 million available or expected within one year under its definition as of June 2025. That is a financial resilience indicator, not a current clinical expansion budget or a promise that an additional gift is unnecessary.

Operating cash flow was negative by approximately $2.60 million in FY2025. This differs from an accounting expense/revenue comparison and does not on its own establish a fundraising gap. Investment transactions, capital spending, restricted resources, credit and timing all affect liquidity. A new donor should not be told that a negative operating cash-flow year automatically purchases additional services. Nor should the presence of substantial assets be used to infer that no donation can matter.

The audit describes a credit facility extended through March 2026. That date precedes this September 2026 review; renewal or current availability has not been established here. The FY2026 operating budget plans approximately $31.04 million of revenue and $31.02 million of expenses, but covers the year ending June 2026. It is a past plan, not the current FY2027 budget. No current marginal clinical offer can be inferred from its small planned surplus.

A useful next financial document would combine the strengths of these sources: a current Foundation-only budget, reconciled program definitions, committed restricted revenue, capacity already financed by public contracts and an explicit description of what unrestricted gifts enable. Until then, the three-year series improves organizational understanding but cannot bear the causal weight of a marginal spending estimate.

Clinical evidence and the distance from a study to GLIDE

The relevant question is not whether food, treatment, housing or supportive relationships can improve health in some setting. It is whether this particular additional Foundation gift changes a clinically meaningful outcome for people who would otherwise receive something different. Each pathway therefore needs an intervention, a recipient population, a comparator, a delivered dose and a duration. The source evidence supplies only some of those pieces.

For medication for opioid use disorder, the accepted evidence includes the Massachusetts post-overdose cohort. It associates buprenorphine exposure with lower all-cause mortality; the model retains an adjusted hazard ratio of 0.63 as an observational anchor. This is not a randomized estimate for GLIDE. People who receive treatment can differ from those who do not, treatment exposure changes over time, and the local population may have different risks. The model moves the ratio halfway toward no effect before applying it.

That causal transfer should not be confused with retention or donor additionality. Causal transfer addresses whether the source association represents an applicable treatment effect. Retention concerns how long effective treatment is delivered. No-gift treatment share concerns medication that would already be accessible without the Foundation-supported pathway. Funding additionality concerns whether the new gift changes the amount of the service rather than replacing other financing. These are distinct concepts, although they could be empirically correlated and should not be multiplied redundantly if future evidence estimates a combined effect.

The emergency-department buprenorphine trial supplies evidence that an active treatment-start pathway can improve engagement relative to referral-oriented alternatives. It does not establish the number of GLIDE clients who are newly retained in medication, nor does a short engagement outcome justify assuming permanent treatment. The accepted model therefore uses a finite active period and keeps treatment completion and alternative access explicit.

Rescue-network support has a different evidence problem. Naloxone can reverse opioid toxicity when present and used in time, but supply distribution alone does not identify the people whose annual mortality hazard changes. A kit may be stored, expire, replace another kit, reach a low-risk setting or be used by a peer on someone outside the assumed cohort. Conversely, a trusted peer network may create substantial protection with relatively few units. The model's risk-network person-year is a synthetic planning unit, not an observed conversion from GLIDE doses.

For meals, the accepted sources include a small medically tailored-meal trial and a larger trial with a null glycemic result. Both are more specific than ordinary communal meal provision. Tailored diets for selected patients cannot be treated as a mortality coefficient for every GLIDE meal. The quantified pathway is narrower: some consumed meals replace inadequate intake and may relieve short-lived physical or psychological symptoms. Its utility change is an analyst judgment, not a measured GLIDE QALY.

The distinction between selection and response matters here. A person who would otherwise obtain adequate food may receive a valued meal without the modeled health increment. A person with inadequate intake may still not experience the specific symptom improvement assumed. These are reasons to model eligible share and response separately. They are not reasons to deny the non-health value of avoiding hunger. The report keeps that value outside the quantified health total rather than converting it using an unsupported exchange rate.

Behavioral support is compared with BRIGHT, a nonrandomized site-period study involving persistently depressed people in residential substance-use care. Its structured sixteen-session, two-hour CBT course is not equivalent to attending a mixed outpatient group. GLIDE's published group count does not establish diagnosis, baseline severity, attendance or protocol fidelity. The model selects a symptomatic subgroup and applies a modest delivery/response prior before a finite quality-of-life benefit.

A depression score is not itself a QALY. Mapping it requires a health-utility relationship, duration, recovery in the comparison group and handling of overlapping symptoms. The model uses a separate utility prior rather than silently translating score changes into health-adjusted years. It also excludes additional substance-use mortality from this behavioral row because medication and rescue survival are modeled elsewhere. This reduces one obvious stacking risk, though it does not prove all residual symptom overlap has been removed.

Women's advocacy and safety support have mixed evidence. An earlier randomized advocacy study provides a positive mechanism lead, while a randomized counterweight reports a depression difference below its stated clinical threshold and no significant health-related quality-of-life improvement. The appropriate transfer is neither universal effectiveness nor universal nullity. The central local utility benefit is small, selected and explicitly weak; signed scenarios permit adverse or null clinical value.

Family services also require restraint. The selected parenting trial and follow-up evidence concern more specific interventions and populations than ordinary childcare or a broad family-resource program. We do not award future reductions in crime, improved earnings or lifetime child health from a family receiving groceries or a class. The model contains a short caregiver-symptom pathway and a separate selected child-symptom pathway, with no generic parent-plus-child multiplier.

Hepatitis C illustrates how the comparator can reverse a superficially attractive transfer. The low-threshold treatment trial compares enhanced colocated care with facilitated referral among RNA-positive people who inject drugs. GLIDE's publicly described linkage may resemble the comparator more than the enhanced treatment arm. It would be incorrect to apply the full trial's cure difference to every test, or even to every referral. The model uses a small earlier-effective-cure difference for a selected untreated subgroup and excludes the paused peer program.

Wound care is clinically plausible but comparatively weakly quantified in this packet. Earlier assessment and treatment could reduce pain and accelerate healing. Public materials do not establish the relevant case mix, wound severity, alternative-care delay or measured healing time. A short symptom window and explicit partner resource cost are therefore used as judgments. The model does not add unobserved hospitalizations avoided, sepsis deaths prevented or disability benefits.

Across these pathways, the strongest common counterfactual is not “nothing happens.” San Francisco has publicly financed clinical services, other nonprofits, emergency care and informal support. GLIDE may improve their accessibility, coordination or continuity. The incremental effect is the difference in timing and completion, not the total effect of medicine relative to no medicine. This is why credible service activity can coexist with a modest or highly uncertain donor-attributable health estimate.

The evidence hierarchy is deliberately uneven because the mechanisms are uneven. Medication mortality has a substantive empirical anchor but local access parameters remain uncertain. Some symptom pathways have clinical rationale and trial analogues but no validated local utility conversion. Policy and newer youth services lack a sufficiently specified health chain here. We describe each at its actual level rather than applying a single “evidence-based” label to the whole organization.

Monitoring and information sharing

Public reporting is useful when it allows a reader to connect resources, activities and outcomes without changing definitions midway. GLIDE provides multiple source types: returns, audited statements, budgets, an impact report and program pages. Their coexistence is a strength, but they answer different questions. The audit addresses financial presentation; the return describes tax-reporting categories and activities; the impact report presents selected organizational outputs; web pages describe access and services. None is a substitute for the others.

The first monitoring priority is a stable unit dictionary. “Dose,” “kit,” “test,” “person,” “enrollment,” “encounter,” “referral” and “course” should have distinct definitions. A year's large distribution total can be accurate while still being unusable as a health denominator. Conversely, a smaller completed-care count can be much more decision-relevant. The report should preserve this distinction rather than reward whichever metric is largest.

The meal discrepancy is an immediate, bounded example. The impact report identifies to-go meals as part of the total, whereas the return narrative appears additive. The remedy is a reconciliation of the two series, including whether each is meals served, meals prepared or distributions through another channel. No elaborate outcome study is needed to answer this. Until it is resolved, a trend in apparent cost per meal would be premature.

Naloxone reporting would benefit from separating units acquired, units distributed, recipients, replacement supplies, high-risk network reach and reported use. These categories are not interchangeable. A reversal report is also not necessarily a unique person, and successful rescue reports are selected observations rather than a prospective fatality rate. For the present model, the missing variable is meaningful additional coverage of a risk network over time, not a larger cumulative supply count.

Medication access requires an explicit cascade: eligible people encountered, people offered treatment, starts, medication type, retained months, interruptions and alternative access. A person already in treatment may still benefit from continuity support, but that is different from a new start. A start followed by immediate loss to follow-up should not receive the same finite exposure as sustained care. Missing follow-up should be reported, not treated automatically as either success or failure.

For testing and linkage, the corresponding cascade is unique people tested, actionable diagnoses, existing treatment status, completed referral, treatment initiation and relevant clinical outcome. Repeated negative screening may have value in risk management, but it should not be modeled as a cure. The paused HCV peer program makes it especially important to distinguish the current general linkage pathway from past or planned services.

Housing monitoring should identify people versus households, assistance type, baseline threat, other available funds and follow-up duration. Three-month housing status is not one-year causal stability. A recipient may remain housed for reasons unrelated to the payment, or the payment may avert a short but harmful displacement. The model needs the difference relative to alternatives, not simply the proportion housed after receiving help.

Symptom-focused programs need basic severity and dose information before ambitious utility claims. For behavioral groups, this could mean the number meeting a relevant clinical threshold and the number completing a defined course. For women's services, crisis encounters and sustained support should be separated. For family programs, a mixed-service family count cannot identify how many caregivers or children received a clinically relevant intervention. Privacy and participant burden should shape collection; precision is not an excuse for intrusive reporting.

The financial monitoring priority is a bridge from current budget to marginal offer. It should reconcile program totals to the Foundation accounts and distinguish direct service, common costs, donated inputs and public complements. Historical accounting ratios are useful but cannot tell a donor which constraint an additional gift relaxes. A current funding map could be more valuable than a more polished annual report.

The final audit reports no identified material weaknesses within the scope of its work, but that is not an opinion on all internal controls or on program effectiveness. Similarly, self-reported governance procedures in the return do not prove every procedure was followed perfectly. We neither ignore these disclosures nor overstate them. Financial assurance and clinical outcome assurance are separate forms of evidence.

Good monitoring should include unfavorable information. Waiting lists that do not move, unsuccessful referrals, early treatment dropout, adverse experiences and failed program launches can improve an estimate. A system that records only activity and success stories will tend to overstate impact even without intentional misrepresentation. The signed model is designed to accommodate null and adverse outcomes if such evidence becomes available.

For a practical next review, three items would be especially high value: the current FY2027 Foundation budget and committed revenue map; a deduplicated medication and rescue-access cascade with alternative-care context; and a reconciliation of annual service units. These would directly affect allocation, additionality and the survival-dominant mechanism. Broad satisfaction surveys, publicity reach or new coalition spending announcements would be less decisive for the present health estimate.

This prioritization is not a demand that GLIDE rebuild its reporting around an external calculator. It is a statement of what would change our confidence. A donor may choose to fund service provision rather than measurement. If so, uncertainty should remain visible rather than being erased by optimistic assumptions. The appropriate standard is proportional evidence for the decision being made.

Sources, reproducibility and research effort

This V2 review uses original Foundation returns for the latest three fiscal years, the final consolidated audit, the published operating budget, the impact report and current official program pages. Clinical evidence is retained from the accepted coverage model with explicit study-to-local transfer limitations. Dates are important: a document published in 2026 may describe FY2025, and a current web page can contain older operational text.

The saved source ledger distinguishes freshly inspected financial and operational evidence from retained clinical sources. The report does not claim that every retained clinical full text was newly reread in this interval. It also does not treat an organization's own service description as an independent outcome evaluation. Primary organizational documents establish identity, finances and advertised activity; external studies inform only the clinical mechanisms they actually examined.

The original model and input files remain alongside the V2 wrapper. The results file contains the exact default output, the retained twenty scenarios and five new diagnostics. Tests compare the full saved object and the original central calculation, reconcile financial components, exercise null and signed cases, reject malformed outer inputs and verify that a larger gift retains full cost while output is capped. One independent scaling assertion uses a tight floating-point tolerance; saved-output parity remains exact.

The historical allocation diagnostic was specified before its output was interpreted. Its broad categories come from the FY2025 Foundation return, while within-category splits remain judgments. The five- and ten-year horizons and reduced rescue-capacity cases isolate uncertainties without changing the accepted default. No scenario weights were added after observing results.

The long-form assessment and its scope limitations are necessary to interpret the headline estimate. No organizational interviews informed this assessment.

This assessment was reviewed September 11, 2026. The source ledger distinguishes new financial and operational checks from clinical evidence retained from the previous assessment.

This method leaves some important questions unanswered. It does so visibly: no current priced marginal offer, no complete portfolio expectation, no measured local utility conversion and no verified current allocation. Those are substantive limitations to carry into the published assessment, not blanks that a renderer or integration script should silently fill.

Qualitative assessment

A QALY-focused comparison does not capture every reason to support GLIDE. Food can relieve hunger and uncertainty even when a measurable health-utility gain is small. A trusted place can offer dignity, belonging and practical help. People may value being treated respectfully without needing that value translated into a medical endpoint. These considerations are real reasons for giving, not defects to be eliminated from a donor's preferences.

The report's narrow quantified benefits should therefore not be read as a complete theory of welfare. A small modeled meal contribution does not mean that meals have little human value. It means that this calculation includes only a selected, finite symptom pathway rather than the full value of nourishment, convenience, security and community. The same distinction applies to childcare, social connection and administrative help.

Policy work is particularly difficult to compress into this framework. GLIDE may contribute to coalitions, public debate, service preservation and budget decisions. The impact report names large public appropriations, but a coalition's total budget win is not GLIDE's own spending or its individually caused impact. A defensible policy-health estimate would need a specific decision, a counterfactual probability, GLIDE's contribution, affected people, displacement and a finite outcome bridge. We do not have that chain here.

Nor is policy necessarily positive in every relevant dimension. A reform can have benefits, implementation failures, opportunity costs and unintended effects. A signed policy model would need to represent these rather than award a generic positive residual because advocacy is mission-aligned. Leaving policy unquantified avoids an unsupported calculation; it does not assert that the expected effect is zero.

The newer youth center similarly warrants attention without a premature coefficient. It may reduce barriers to care, provide respite and help young people navigate unstable circumstances. Its public and corporate financing, partner services and overlap with other programs are central to the counterfactual. A young population does not automatically imply a large lifetime health return. The model would need actual service pathways and finite incremental outcomes, not age alone.

An integrated organization can also have complementarities that are hard to allocate. A meal may make it easier to attend an appointment; a trusted case manager may make a referral effective; reliable facilities may support clinical partners. Allocating common costs to one branch does not prove the branch operates independently. The current model's separate pathways are an accounting and reasoning aid, not a claim that GLIDE is a set of isolated mini-charities.

These complementarities cut both ways for cost-effectiveness. They could make the combination more effective than isolated services. They could also mean that a small restricted gift cannot reproduce the apparent effect without the whole infrastructure. A donor comparing GLIDE with a narrow clinical implementer should account for both possibilities. It is inappropriate to credit the benefits of the integrated setting while excluding its common costs.

There are ethical considerations in choosing what to measure. People should not need to demonstrate a large expected health gain to deserve respectful treatment. A health-maximizing donor may nevertheless use a limited budget to prioritize interventions with more measurable benefit. Those are different decisions. This report supports the latter comparison while acknowledging that the former principle is not reducible to its result.

A donor who values GLIDE's community role may reasonably support the Foundation despite the HOLD on a verified health-maximizing recommendation. The honest explanation would name those values and the uncertainty. It should not borrow the authority of a favorable old rental estimate or an unweighted central scenario to imply a level of quantitative confidence that is absent.

Our bottom line is therefore deliberately two-part. GLIDE is a real, broad service organization with credible pathways to effective care and meaningful non-health aims. The public evidence does not yet establish that an ordinary gift buys low-cost incremental Bay health at the thresholds used in this project. Better marginal capacity and completed-care evidence could change the second statement without changing the first.

Executable model: full gift, unique cohorts and finite health

The portable model retains the accepted input specification and calculation engine as separate original files. The V2 wrapper adds the financial series, independent diagnostics and stricter outer validation. Its default central result is exactly equal to the original engine's result on the preserved inputs. This separation makes it possible to reproduce the old numerical position while seeing what the fresh evidence does and does not change.

The default gift is $100,000. The assumed absorbable tranche is also $100,000, but this is a planning cap, not observed room for more funding. When the gift rises above that cap, modeled output does not continue linearly: the full larger gift remains in the cost numerator while activity is capped. This guards against an attractive price being extrapolated indefinitely. It does not establish that the first $100,000 is actually absorbable.

The allocation covers the entire gift:

These are retained marginal-allocation priors. They are not quoted budget percentages. The historical-mix diagnostic replaces the broad program shares using the FY2025 Foundation return, while retaining explicit within-bucket judgments. Neither family is empirically established as the destination of a current ordinary gift. In particular, the new youth center makes an old annual mix less representative of today's operations.

Funding additionality is 0.4 in the central scenario. Interpreted narrowly, a nominal service unit financed through the tranche produces 0.4 additional units relative to the counterfactual funding and capacity path. This is not inferred from the fraction of revenue that is governmental. Public contracts, donors, volunteers and staff capacity may interact nonlinearly. The zero-additionality scenario is therefore substantive: it represents a gift that sustains or substitutes for an already financed baseline without generating the modeled incremental health.

For rescue support, $500 per nominal annual risk-network participant converts the $10,000 allocation into 20 nominal person-years, below the capacity prior of 100. Applying funding additionality gives eight incremental equivalents. These people are not derived by dividing the reported naloxone doses by an arbitrary kit ratio. The cost and coverage are assumptions about a meaningful package of protection in a high-risk network. Their lack of local calibration is one of the most important reservations.

The medication branch allocates $3,000 at a $16,000 annual Foundation access/support cost, yielding 0.1875 nominal units before further discounts. The model charges at least one annual budget unit for setup even though the central effective treatment period is half a year. A 0.5 no-gift treatment share and 0.5 effective-retention factor, alongside 0.4 funding additionality, produce 0.01875 incremental treatment equivalents. This small number prevents the 56 historical enrollments from being treated as automatically marginal to the new gift.

The survival engine forms three exclusive cohorts: rescue only, medication only and both. Overlap is 0.5 of the smaller incremental cohort centrally, giving 0.009375 people in the shared group. The combined group is removed from the single-intervention groups before health is summed. That prevents awarding two separate full survival gains to the same person. It does not assume that everyone receiving multiple services is independent across every quality-of-life branch; those branches have separate unique-health factors.

The baseline all-cause mortality hazard is 0.07 per year. Rescue support reduces the active-period hazard by 0.005, while medication applies a hazard ratio shrunk from 0.63 toward one by a 0.5 transfer factor, giving 0.815. In the overlap cohort the active hazard is the reduced rescue hazard multiplied by the medication factor. An all-cause endpoint is used once; opioid deaths are not added on top of all-cause deaths.

Rescue protection lasts one year and medication protection half a year in the central scenario. After each funded exposure ends, its effect is removed. Thereafter both counterfactual trajectories face the same baseline mortality hazard. A survival gap created during care can persist, but no new treatment effect is generated after support ends. The integral stops at fifteen years, uses health utility 0.7, discounts at 3% and allows a quarter-year delay.

This structure addresses repeated-rescue double counting. A person does not receive a fresh fifteen-year benefit after every potential rescue. The model compares survival probabilities for a unique cohort under two hazard paths, so the gain is bounded by the finite health of that cohort. It remains a simplified expected-cohort model: hazards are constant within segments, and it does not estimate individual risk histories, fentanyl exposure changes or long-term recovery transitions.

Housing uses a $10,000 fully loaded access/stabilization episode rather than the approximate $3,045 transfer per reported rental-assistance person. It applies a 0.056 incremental stability difference, retained from a prior evidence-transfer assumption, capped by the room above 0.8 no-gift stability. A unique-health factor of 0.8, utility change 0.1 and one-year window follow. This is not evidence that every assisted person would otherwise become homeless, or that every prevented displacement yields the same health effect.

Shorter quality-of-life branches follow a common sequence: funded nominal courses, additional courses, relevant symptomatic share, effective-response/transfer share, distinct health-time, utility change and finite duration. Mortality and alternative-care catch-up reduce the integral where specified. Adverse course effects can subtract health. The “response” factor is a local transfer judgment, not an extra adherence discount applied to an already identical local trial estimate.

Meals use a thirty-day course of ninety meals costing $1,080. The model selects 20% as replacing inadequate intake, 50% as having the relevant response and 65% as distinct symptom-time. Utility is 0.015 over thirty days. Parenting uses a $4,500 course, selected share 0.3, response 0.35, unique share 0.7, utility 0.03 and a half-year window with substantial catch-up. These small contributions are not claims that food or parenting support has little total value; they reflect the narrow health bridge being quantified.

Childcare uses $13,000 per selected course, 0.2 symptomatic share, 0.2 response and utility 0.03 over half a year. Women's support uses $4,000, 0.35 selection, 0.3 response, 0.6 distinct health and utility 0.04. Behavioral support uses $6,000, 0.4 selection, 0.3 response, 0.6 distinct health and utility 0.08. Their finite windows and catch-up assumptions exclude indefinite symptom improvement or lifetime social returns.

HCV linkage uses $300 per unique linkage episode, 0.06 actionable untreated infection share and a 0.05 earlier-effective-cure difference relative to existing care. Utility 0.03 is integrated over three years with alternative-care catch-up and mortality. Wound care uses $600 per course, 0.4 symptomatic share, 0.5 earlier response, 0.7 distinct symptoms and utility 0.06 over fourteen days. Neither branch adds secondary transmission, avoided admissions or survival without a separate explicit chain.

The resource account adds gross associated external inputs, including clinical partners and common complements. It totals approximately $132,985 of associated funding for the $100,000 gift, then excludes $3,045 of rental cash transfer from the resource measure, producing approximately $129,940. This exclusion does not reduce the donor's gift cost. A transfer is not itself a consumed real resource in this convention, but the donor still pays it.

External resources are charged on nominal associated activity rather than only on successful additional health. In particular, HCV treatment resources are included for the nominal actionable subgroup before donor additionality and cure-response discounts. This may include resources that would also be used in the no-gift world. The result is therefore a gross associated-resource diagnostic, not a causal net social-cost estimate. It should not be presented as an audited economic cost or a precise measure of public spending induced by the gift.

Bay geography is 0.99 and San Francisco is 0.95, nested priors rather than a measured residence file. The all-beneficiary quantity is larger than either geographic subset. A site in San Francisco does not prove that every beneficiary resides there, while Bay policy spillovers are not assigned automatically. The donor price for Bay outcomes retains the whole gift; it does not reduce cost by the Bay share and thereby cancel geographic dilution.

What do you get for your dollar?

The central result remains exactly unchanged: 0.2044715224666835 Bay QALYs from the hypothetical $100,000 gift, or $4,890,656.595775775 per 10 Bay QALYs. The all-beneficiary total is 0.20653689138048836, and the San Francisco total is 0.19621004681146392. Their corresponding donor prices are approximately $4.84 million and $5.10 million. These are different geographic denominators, not competing versions of the same Bay figure.

The named survival pathways produce 0.20332877354402284 QALYs before geography. Housing adds 0.0017198322434524124, and the seven shorter symptom pathways together add approximately 0.00149. Thus the central diagnostic is overwhelmingly driven by survival assumptions. Adding several small clinical rows has improved scope transparency, but it has not made the result empirically diversified. A reader should not mistake a long list of pathways for many independent sources of well-established expected value.

The gross-associated-resource price is approximately $6.35 million per 10 Bay QALYs, compared with the $4.89 million donor price. This difference reflects modeled partner inputs and the stated transfer convention. It does not imply that a donor must personally pay the resource total. It also does not establish that every associated public resource is marginal. Both views are useful only if their boundaries remain visible.

The historical-mix diagnostic is not a preferred revised estimate. It uses actual broad Foundation program categories but still needs judgments to divide health spending among rescue, medication, HCV and wounds, and to divide other program spending among housing, women's services, behavioral support, policy and affiliates. The actual categories are more informative than an unsupported single clinical share, but they do not eliminate uncertainty about the marginal gift. New FY2026 operations make the distinction especially important.

The shorter-horizon diagnostics isolate a clinically meaningful uncertainty. They do not assume that every overdose survivor lives only five or ten years. They ask how much of the modeled benefit survives if the integration window is shorter while other accepted assumptions stay fixed. The large change shows why an explicit finite horizon is important. A survival model can be finite and still highly sensitive to the choice of that horizon.

The rescue-capacity diagnostics similarly isolate coverage uncertainty. They reduce the maximum nominal risk-network coverage to a quarter or half of the central nominal count, leaving the clinical hazard assumptions unchanged. They are not estimates of actual GLIDE capacity. Their purpose is to show how strongly the apparent return depends on converting dollars into meaningful additional protection rather than merely distributing units.

The accepted twenty-scenario suite is retained separately. It includes zero additional funding impact, zero capacity, independent harm, clinical nulls, weak and favorable delivery, rescue and medication nulls, overlap extremes, adverse clinical effects, zero distinct symptom benefit, longer funded courses, zero discounting, geography changes and a larger gift with capacity held fixed. These scenarios are not equiprobable draws. Their existence demonstrates inspectability and signed outcomes, not a probability that the central result is correct.

A zero or negative net-health scenario does not receive a misleading positive “cost per QALY.” The model's null-price convention is retained and numerical outputs are checked for finiteness. This matters when benefits and harms nearly cancel: an extremely small positive denominator can otherwise create an infinite ratio. The wrapper validates its allowed gift range, and the preserved core supplies its own parameter checks. Tests also compare the entire saved output, not only a rounded headline.

At the central result, reaching $1 million per 10 Bay QALYs would require roughly 4.89 times as much quantified Bay health for the same gift. Reaching $100,000 would require roughly 48.9 times as much. Those are arithmetic thresholds, not proposed coefficient changes. Unquantified pathways might add value, but we have not established an amount sufficient to bridge either gap. A favorable story about policy or youth services is not an appropriate substitute for that missing chain.

The adverse central result should not be improved by dropping common costs, narrowing to a successful clinical subgroup or treating public clinical resources as free donor-created capacity. Conversely, it should not be interpreted as an upper bound on GLIDE's total health value. It is conditional on a selected set of pathways and priors. The most decision-relevant next step is to identify the real marginal offer and its completed-care outcomes, not to adjust assumptions until a ranking threshold is met.

Uncertainty, overlap and what would be an actual model correction

Some uncertainties are ordinary parameter ranges; others change the meaning of the model. The cost of a support course can be varied numerically. Whether the course exists as a purchasable marginal service is a different question. Similarly, a utility range can express uncertainty about symptom benefit, but it cannot repair a denominator that counts visits when the model requires unique completed courses.

The largest current structural uncertainty is the rescue-risk network. The central model represents eight additional equivalents after financing adjustment. It does not establish who they are, their baseline risk, the duration of their protection or their access to naloxone without GLIDE. A local study could reveal that the effective cohort is larger, smaller or differently selected. Until then, the same accepted rescue hypothesis that makes the survival branch plausible also makes its result fragile.

Mortality evidence should be used symmetrically across organizations. A high-risk post-overdose cohort is not automatically the right baseline for all kit recipients, students or passersby. GLIDE's service population may include people at substantial risk, but this report does not infer that risk from a neighborhood label alone. The 0.07 hazard is a planning prior. A comparison with another organization should examine population selection and alternative access before treating different hazard assumptions as evidence of different effectiveness.

The finite survival engine is a substantive improvement over event-linear lifetime accounting, but it is not a calibration study. It avoids restarting a lifetime after repeated rescue and combines overlapping medication and rescue exposure coherently. It still assumes a simplified common hazard after support ends. If treatment changed long-run risk beyond the financed period, the present model could omit benefit; if the assumed survivor population faces greater later mortality, it could overstate benefit. Neither possibility warrants a one-sided adjustment without evidence.

Several discount factors could overlap empirically. Funding additionality, treatment substitution, retention, selected symptoms, response and unique health have different intended meanings. But a local causal study might estimate an effect that already includes some of them. If such a study were adopted, the appropriate correction would be to remove redundant factors, not to keep every conservative-looking multiplier. In the present model they are priors filling distinct missing links, and no single local treatment-effect estimate is multiplied by all of them.

Likewise, a positive response factor should not conceal a signed clinical result. The scenario system permits adverse hazard changes, adverse utilities and independent harm. Potential harms include delay of more appropriate care, burdensome participation, distress or resource diversion. These are not observed GLIDE harm rates. The model includes small explicit harm terms and adverse scenarios to avoid assuming all activities can only help.

Overlap is only partly measured by the code. Rescue and medication overlap is explicitly partitioned into exclusive cohorts. Symptom branches use “unique health” factors to limit overlapping mood, safety or physical symptom-time. These factors are not a person-level linkage file. Someone may receive housing, women's support, meals and behavioral services. A future integrated outcome registry could justify more precise deduplication; current aggregate program counts cannot.

The model also distinguishes public-financing displacement from care alternatives. Existing public money can mean a private gift replaces another funding source, but it can also provide a valuable complement that makes private navigation unusually productive. Separately, a person may receive care elsewhere even if GLIDE's budget genuinely expands. Applying a generic “public funding penalty” to both mechanisms without specifying them would obscure the counterfactual. This report retains a broad funding prior and explicit clinical alternatives rather than deriving either from revenue shares.

There is a risk of optimistic omission as well as optimistic inclusion. Policy, infection prevention, vaccination, youth services and community effects may be material. A partial-health report that ranks itself as complete would misleadingly omit them. Yet adding an arbitrary residual would create a different error. The appropriate publication remedy is a prominent scope label and unresolved expected-value status, not an automatic positive adjustment.

Three findings would require immediate model or presentation correction. First, evidence that the assumed ordinary gift route actually goes to a different legal recipient would invalidate the recipient boundary. Second, evidence that modeled nominal units are visits or supplies rather than meaningful courses would require a revised conversion. Third, any headline that labels the central partial diagnostic as a complete weighted expectation would be substantively false even if the arithmetic remained correct.

By contrast, a more recent budget that differs from the assumed allocation does not mechanically prove the central prior wrong. It would inform a new marginal allocation judgment, which should be documented before seeing the resulting price. The present historical-mix sensitivity illustrates that discipline. We preserve the old default, expose a source-informed alternative and identify the missing current marginal budget rather than claiming that an accounting fraction is a causal parameter.

The next independently elicited scenario distribution should be specified before inspecting its ranking consequences. It would need correlated states of demand, capacity, displacement, clinical effect and geography, with explicit null and harm possibilities. It would also need a decision about unresolved portfolio pathways. Until that work is done, reporting a central scenario and a transparent range is more honest than attaching a numerical confidence level that the evidence has not earned.

Funding and previous grants

We found an official Foundation giving route, but not a dated offer specifying additional clinical capacity at a stated gift tier. These are different levels of readiness. A working donation button establishes that an organization accepts money. It does not establish that $10,000 or $100,000 will create a certain number of extra treatment starts, retained patient-months or safer-use person-years.

The most useful offer would describe an operational bottleneck and its current financing. For example, a navigation team might have qualified demand and public clinicians available but lack outreach staffing. In that case an additional private gift could plausibly create completed care without financing the entire medical service. Alternatively, staff and clinicians may already be at capacity, or public contracts may cover the same work. A general statement that demand is high does not distinguish these situations.

Public funding is visible in multiple parts of the portfolio. The return's government-grant category is material, but its recent accounting reclassification prevents simplistic year-to-year comparisons. Program descriptions and the impact report also name public and philanthropic support. The newer youth center explicitly identifies public agencies and BMO. These sources establish that the no-gift world includes substantial existing financing; they do not establish the amount of private displacement at the margin.

Clinical partnerships create a similar ambiguity. Public Street Medicine and university clinicians could make Foundation access work highly productive if they have unused capacity and GLIDE reaches otherwise untreated people. If their capacity is fixed, more referrals may rearrange who receives care rather than increase treatment. If clients would reach those providers anyway, the effect may mainly be reduced delay or burden. The model has broad priors for these issues, not a verified current operating map.

Rental assistance is a particularly important example because the earlier favorable estimate can sound like a concrete purchase. Historical transfer totals and recipients document assistance delivered, not an open tranche. A current donor would need to know whether funds can be designated, whether eligible cases are waiting, the fully loaded cost of processing and support, the role of other assistance funds, and how long the intervention changes housing relative to alternatives. A historical transfer per person is not the price of an incremental stable household.

The FY2025 financial statements show both substantial resources and financial pressures. Neither observation resolves marginality. Assets may be restricted, invested, tied up in property or intended to sustain future operations. Negative operating cash flow may reflect timing, investment decisions or a genuine operating challenge. A claim that GLIDE urgently needs a certain amount for health expansion would require a current forecast and program-specific evidence, not a subtraction of annual revenue from expenses.

The last located operating budget is for FY2026, now ended. Its contract revenue and planned expense categories are useful evidence of the scale and composition of planned operations, but they are not a current FY2027 funding gap. The new youth center opened during that budget year. A donor deciding now should seek an updated plan that separates continuation commitments, restricted funding, new activity and contingency reserves.

The model's $100,000 capacity cap is therefore an assumption rather than a fundraising recommendation. It prevents unlimited scaling in the calculator but cannot certify that even a smaller tranche is productive. The correct status of capacityOfferVerified is null. Replacing that null with a positive number merely because the organization has a large budget would confuse organizational scale with marginal absorbability.

For an unrestricted donor, the first practical question is allocation: how would the next tranche differ from the counterfactual budget without it? A response that names broad annual programs is less useful than one identifying staffing hours, service capacity, coverage duration and financing already committed. The second is substitution: which other funds, public contracts or partner resources would change if the private gift arrived? This should include whether the gift frees flexible money for another activity, not only the restricted label on the payment.

The third question is delivery: how many additional unique eligible people would complete the relevant course, and over what period? A range may be adequate if its basis is clear. For medication, retained patient-months are more informative than referrals. For housing, unique households and the assistance counterfactual matter. For rescue support, meaningful network coverage and high-risk reach matter more than doses. For symptoms, baseline severity and completed sessions are essential.

The fourth question is measurement: what existing records could verify the proposed mechanism without imposing a disproportionate new research burden? A small marginal offer need not promise a randomized trial. It could still report deduplicated eligibility, start dates, retention, alternative-care delays and adverse outcomes with honest missingness. Routine operational data are more useful when definitions remain stable and are linked to the modeled unit.

Finally, the donor needs a legal and financial route. A restricted offer should specify the Foundation as recipient, the purpose, treatment of excess funds and whether common costs are included. We have not verified such an agreement and have not contacted GLIDE. The official general giving page is a navigation route, not a negotiated contract. Any future restricted arrangement would require its own updated model rather than silently inheriting this ordinary-gift result.

There is no need to reject exploratory modeling solely because the offer is unknown. The model can still identify what evidence has high decision value. But there is a need to distinguish research promise from donation readiness. Here the best next evidence is a current marginal spending and capacity map, followed by completed-care data for the survival-dominant pathway. More general impact stories would add context but would be unlikely to resolve the central health-ranking uncertainty.

A reasonable donor conversation could also reveal that GLIDE's priority is not clinical expansion. It may be maintaining trusted community infrastructure, preserving flexible support or addressing a new program need. Those can be legitimate purposes. If so, the decision should be evaluated on those terms rather than forcing the organization into a medical cost-effectiveness offer it has not made.