An educational screening tool that helps Washington families identify which long-term care funding programs may be worth exploring with a professional.
CareNav does not determine eligibility, provide legal or financial advice, or estimate benefit amounts. Every report carries that disclaimer, and the system is built so it structurally cannot overstate: eligibility logic is deterministic and auditable, every factual claim cites a curated source passage, and missing answers surface as explicit unknowns rather than guesses.
CareNav helps a family figure out which Washington long-term-care funding programs are worth exploring, before they spend hours researching ones that don't apply.
Who it's for. An adult child helping an aging parent, usually mid-crisis: a hospital is discharging a parent in days, or a fall just ended "managing at home." They aren't benefits experts. They discover these programs one at a time, from pages written for caseworkers, while doing everything else a hard week demands. The job they're hiring CareNav for is orientation: which programs are even worth my time, in plain language.
A deliberate scope note. From what I saw in the ER, the deepest pain for these families is often further downstream, in the applications themselves and in getting a straight answer on whether a parent qualifies. CareNav does not try to solve those, on purpose. Determination depends on assessments no questionnaire can run (the state CARE assessment, a VA medical exam), and applications are a heavier build. CareNav targets orientation because it's the slice an automated, cited, honest tool can do well and safely, and because orienting a family faster is what unblocks them to reach the professional who handles the rest. The handoff isn't a limitation bolted on; it's the point.
The benefits professional who makes the actual determination is a downstream recipient of the report, not a user I've designed for yet. Whether the report is genuinely useful to them is a hypothesis I want to test, not a claim I can currently stand behind.
What I deliberately did not build, and why. These decisions define the product more than the features do.
- It does not determine eligibility. It screens for "worth exploring," never "you qualify." Eligibility depends on assessments no questionnaire can run. Claiming determination would be wrong, a liability, and would set families up for a false yes. The architecture is built so the system structurally cannot overstate: deterministic rules, every claim cited, missing answers surfaced as explicit unknowns.
- It does not guess missing answers. Every question is skippable, and a skip is recorded as an unknown, not a default. A mid-crisis family won't have every number at hand, and a tool that quietly assumes income or assets to produce a cleaner result is lying to them. The "what we couldn't determine" section is a first-class output.
- It is Washington-only, twelve programs. Funding rules are state-specific and change often. A shallow fifty-state version would be confidently wrong; a deep single-state version can cite primary sources. Depth over coverage was the call.
- It does not persist reports server-side. Reports live only in the browser session, trading shareable links and refresh-recovery for never holding a family's health and financial answers on a server. For this data, no-PHI-at-rest was worth the UX cost.
- It accepts a known recall cost in retrieval. Passages are scoped to the program they belong to, so a citation can never point at another program's document, even though this occasionally misses a cross-program passage a human would find relevant. A tool whose whole promise is "every claim is sourced" cannot afford a citation pointing at the wrong source.
How I'd measure success. North star: time-to-oriented-shortlist, from "I don't know what exists" to "here are three or four programs worth a professional's time, with sources," in one sitting instead of hours over weeks. Supporting metrics: intake completion rate, whether a benefits professional would call the shortlist a reasonable starting point, unknowns surfaced honestly rather than guessed (currently 100%), and family-reported trust (not yet instrumented). Guardrails that must never regress, and are treated as release gates: hallucinated-program rate (0%), citation validity (100%), refusal correctness.
Honest status: this is pre-user-testing. I have not yet put CareNav in front of real families or benefits professionals. Everything above is a designed hypothesis. The evaluation to date measures whether the system faithfully does what it intends (fidelity) and, through a blind labeling pass, whether some underlying rules were actually correct (they weren't, in nine cases, which I fixed). It does not yet measure whether the product helps a real person.
The assumptions I most need to test, roughly by risk:
- That solving orientation is valuable even though it isn't the deepest pain. The deeper pain is applications and determination; the bet is that orienting a family quickly and honestly still saves real time and unblocks the handoff. If not, the product should move downstream.
- That a hedged, cited shortlist reduces a family's time and anxiety rather than just adding a step before they call a professional anyway.
- That families will complete a ten-minute intake mid-crisis instead of abandoning it.
- That "educational, not a determination" is a position families accept, when what they emotionally want is a yes or no.
- That the report is useful to a benefits professional rather than something they'd re-verify from scratch. (Untested, which is why the professional is a hypothesis, not a designed-for user.)
What would make me pivot or kill:
- If families tell me orientation was never the hard part and CareNav just added a step before the paperwork, it moves downstream toward guided applications, where I saw more of the pain.
- If families won't finish the ten-minute intake, the flow is too long and gets cut down or restructured.
- If benefits professionals won't treat the output as a trustworthy starting point, its ceiling as a family-facing tool is capped and it either pivots professional-facing or refocuses on the handoff.
Long-term-care funding in Washington is fragmented across at least a dozen programs — Medicaid waivers (COPES, CFC), institutional Apple Health, the WA Cares Fund, two VA pension enhancements, two Medicare benefits, PACE, TSOA, MAC, and respite programs — each with its own agency, eligibility rules, income and asset standards, and application path. Caregiver accounts suggest this research can consume many hours spread over weeks (figures like "15+ hours" circulate; treat any such number as illustrative — I haven't sourced a rigorous estimate). A screening takes roughly 30 minutes (illustrative), from questionnaire to a cited, prioritized shortlist a family can take to a benefits professional.
Intake
↓
Rules engine
↓
Candidate programs
↓
Retrieval (top-k = 5)
↓
LLM explanation
↓
Cross-check + citations
↓
Report UI
v1, measured against the latest committed evaluation baseline
(eval/results/results.md):
- Explanation reasoning effort: low
- Evaluation size: 100 synthetic cases
- Mean latency: 44.2 s · p95 latency: 57.5 s
- Mean cost per report: $0.107 · total evaluation cost: $10.62
- Hallucinated-program rate: 0%
- Citation validity: 100%
- Refusal correctness: 99.0% (the single flag is a strict-metric edge case — a borderline bare-"if" conditional, not an eligibility assertion)
- Pipeline failures: 1/100 — the cross-check refused to serve a report with an uncited claim, the designed safety behavior firing under model variance
family answers intake (~10 min, every question offers "I don't know")
│
▼
deterministic rules narrow 12 programs to candidates (src/lib/rules/)
│ excluded programs never reappear
▼
program-scoped retrieval pulls passages from the (src/lib/rag/)
curated corpus for each candidate
│
▼
LLM explains and prioritizes ONLY those candidates, (src/lib/llm/)
citing retrieved passages; output is schema-validated
and mechanically cross-checked
│
▼
family reviews the report — then takes it to a benefits
professional, who makes any actual determination
Intake asks one plain-language question at a time (about 10 minutes). Answers can be a numeric age, a single choice, or a multi-select — including multiple diagnoses. Every question can be skipped: the "I don't know" and "Prefer not to say" actions are shown as secondary "Skip this question" controls, and a skipped answer is recorded as an explicit unknown, never a guess. A review step lists every answer before submitting.
The report opens with the programs most worth exploring first — a ranked index that links to each program — followed by a single "Your next step" drawn from the highest-ranked program's own next actions. Each program is a numbered card (matching the index) with a one-line reason it may apply, its next steps shown inline, and collapsible sections for what it generally covers, information still needed, and sources. Every factual claim links to a cited source passage. A closing "What we couldn't determine" section groups the answers left unknown, so a family can see what would sharpen the result.
The layout is responsive and readable on a phone:
- Built for adult children helping aging parents, older adults, and families navigating stressful care and funding decisions.
- Calm, government-adjacent, operational design — deliberately not startup-marketing aesthetics.
- Large typography and strong contrast for legibility.
- Accessible: keyboard support with visible focus, native disclosure controls, no dark mode.
- Educational only — it never determines eligibility.
- Never guesses missing information; unanswered questions surface as explicit unknowns.
- Washington State programs only.
- Educational screening only — not legal, financial, or eligibility advice.
- Program administrators (DSHS, the Health Care Authority, Medicare, the VA, and the local Area Agency on Aging) make the final eligibility decisions.
- Evaluation uses synthetic cases whose ground truth derives from the deterministic rules engine.
- Latency is about 40 seconds per report (mean 39.4 s, p95 51.7 s over the
committed 100-case baseline). Sub-10-second latency is not achievable
under a citation-validated architecture: a single fully-cited, multi-program
generation dominates wall-clock, while retrieval, embedding, and assembly add
only ~3 s. The intended mitigation is a UX one — setting expectations up front
and showing a progress indicator while the report is produced (see
docs/evaluation.md) — not a smaller number. POST /api/screenis rate limited per IP (default 5 requests/minute) because each report spends real money on paid APIs; exceeding it returns429with aRetry-Afterheader. The limiter is in-memory, so under multi-instance or serverless deployment the effective limit is per instance, not global.- Reports live only in the browser session (
sessionStorage), never persisted server-side. This is a deliberate privacy tradeoff — no PHI leaves the browser — at the cost of no shareable report link and no recovery after a refresh or tab close.
npm install
cp .env.example .env.local # fill in ANTHROPIC_API_KEY and VOYAGE_API_KEY
npm run dev # http://localhost:3000 — start at /intakeHeads-up: generating a report calls paid APIs and takes 40 s ($0.11 per
report at list prices).
Other commands:
npm test # unit + boundary tests (colocated *.test.ts)
npm run ingest # rebuild the RAG index from src/data/corpus/
npm run eval # 100-case evaluation harness — paid APIs, ~$11 and
# ~13 min per full run (use -- --limit N to sample)
npm run eval:generate # regenerate the synthetic test set
npm run build # production build/baseline hosts a deliberately primitive keyword-search comparator used by
the evaluation harness.
- Architecture — the three-layer design and why deterministic logic is separated from generated language.
- Governance — what CareNav does and does not do, trust boundaries, failure modes and mitigations.
- Evaluation — methodology, current results, the independent labeling pass, and known limitations (including latency).
- Accessibility — the automated axe-core audit, contrast ratios for the palette, and keyboard/focus confirmation.