Evidence
0 of 55 gifts that would lower the check were let through, over every case whose right answer is known: Social Security’s own examples, real court cases and labeled bills.
The five-minute path
- The problem in one number 20 s
- Give, and pay in the sandbox 90 s
- Her report to Social Security 40 s
- Her own copy, in Easy Read 20 s
- An AI assistant asks first 40 s
- Her benefit calendar 40 s
What to look for at each step, the sandbox buyer login, and the guided tour
Eight stops in this tab, about five minutes. Escape ends it.
In the Gift Checker, $1,000 sent as cash cuts her check two months later. Press Pay the shop: she keeps all of it.
Give, and pay in the sandbox90 s
Type $1,500 in Give any amount. The rules split it across her needs and her ABLE account, and one PayPal checkout pays each shop and her ABLE account directly. Pay with the sandbox buyer below, or as a guest with PayPal's test card, then open the receipt.
Her report to Social Security40 s
SSI asks for changes in income by the 10th. October's report says nothing counts, with the rule behind every gift. Press As cash: the same gifts would have counted $538 and cut her check to $456.
Her own copy, in Easy Read20 s
The same report written to Maya: the answer first, then one idea per line with a picture beside each. Press En español, and Remind me by the 10th for a calendar reminder.
An AI assistant asks first40 s
A Claude assistant with no SSI knowledge calls the Giftwise MCP server, talks June out of sending cash, starts a checkout only she can approve, and reads the order back from PayPal.
Press As cash and drag Aunt June's $150 from October 28 to November 2. A different check drops: SSI counts income by calendar month.
Sandbox buyer login, for step 2
It works only in PayPal’s sandbox checkout window, and no real money moves.
- Sandbox email
sb-tvnmx53636325@personal.example.com- Password
••••••••
This is PayPal's sandbox test buyer: a practice account that works only in the sandbox checkout window, where no real money moves.
The three headline results
$331 to $663 a month kept in real court cases; 40 of 40 of SSA's own examples; 0 of 55 let through that would lower the check
Three numbers first
- $331 to $663 a monthFamily help counted against the person in all 14 real court cases; routed through Giftwise, the same help keeps this in 4 of them at 2026 rates.$38,103 kept by two families over the 41 and 37 months their cases cover. The rules engine's counterfactual on each case's own facts: what it keeps each month ($331.33 and $662.67), repriced at 2026 rates, times the months the opinion states.
- 40 of 40of Social Security’s own worked examples match SSA’s answer.Claude with no rulebook: 32 of 40. New examples the engine never saw, each set run once on it frozen: 4 of 5, 1 of 5, 1 of 5 and 6 of 15 on every value, 19 of 30 on whether the gift counts.
- 0 of 55gifts that would lower the check were let through as given.Counted over SSA’s examples, the court cases and the labeled bills, wherever the right answer is known.
Does it get the rules right?
40 of 40 SSA examples; court cases, 14 of 14 scored, of 36 cases; 59 of 59 bills; 100,000 random gifts
- Social Security's own examples40 of 40 match SSA's answer; Claude with no rulebook, 32 of 40.
- Examples the engine never sawFour sets of SSA's worked examples, each run once on a frozen engine: 4 of 5, 1 of 5, 1 of 5 and 6 of 15 on every value, and 19 of 30 on whether the gift counts.
- Real cases from federal courts14 of 14 scored, of 36 cases, reach the outcome SSA and the court did.
- The two reviewers, measured59 of 59 labeled bills keep her check whole, with each reviewer alone shown beside the referee.
- The referee, under pressureA wrong quote, a quote from the wrong rulebook, a right answer with a made-up citation, and two reviewers who disagree: what the code does with each, and the check whole every time.
- The two errors that cost money0 of 55 let through that would lower the check; 6 of 51 fine gifts sent another way, each with the check whole.
- 100,000 random giftsEvery one keeps all 13 promises: every cent placed, her check exactly as with no gift, ABLE within its yearly limit.
- The rules watch the Federal RegisterAsked live for every change to SSI's income rules and amounts since 2022; the 2027 amounts are in use.
- What the AI never decidesHow much each payee gets, which rule applies, the check, when money moves, who is paid and the report: code, every time.
Does the money go only where the rules say?
A receipt to read without paying, a report checked against PayPal, every PayPal API it calls, and 8 failure cases run live
PayPal is where the rules are enforced: money reaches only a payee the rules picked, in one order, captured once, and anyone can check a month's report against PayPal.
- A receipt you can read without payingA recorded sandbox gift, seeded with the sample page: every payee and capture id, the rule behind each line, and Check with PayPal now asks PayPal for the order live.
- A report anyone can checkNovember's report lists the recorded sandbox gift and carries a QR code: Giftwise rebuilds it, asks PayPal for each capture live and checks the audit ledger.
- Claude reads a bill, livePress Read this bill live: Claude reads the sample bill now, within the daily AI budget, and the rules engine routes it.
- Every PayPal API it callsOne order paying several payees, capture once with PayPal-Request-Id, approval links for agents, Payouts, Transaction Search and signed webhooks, each with its file.
- The PayPal suites, step by stepThe checkout and API suites write a redacted transcript: each step with its PayPal order and capture ids, payee count, statuses and timings.
- 8 failure cases, run liveA declined card, a double click, a browser changing the amount, refunds and more, against the PayPal sandbox.
- Give every monthAfter a gift, press Give this every month: the same amount is routed again each month, with what changed and why.
- A fund pays many familiesOne PayPal Payouts batch, each item paid straight to a shop or ABLE account, confirmed in PayPal's Transaction Search.
- The ABLE share, for realPayPal settles the shops; the ABLE share goes through the plan's own gift page, by Ugift code or gifting link. The gift page and every receipt with an ABLE line show the hand-off step by step.
What it changes
$38,103 for two families in real court cases; the problem as Social Security measures it; who each route reaches; how it compares
- The problem, measured by Social Security454,843 checks lowered for shelter help in December 2024, by $247 a month on average; in-kind support 8.59 percent of SSI improper payments; no tools to find living arrangement errors; the $2,000 limit and ABLE's reach, each from its source with the date.
- What routing would have keptSocial Security counted help from family or friends against the person in all 14 real SSI court cases scored. Routed through Giftwise, the same help would have kept $331 to $663 a month at 2026 rates in 4 of the 14 as the opinions state the facts, and money in all 10 cases computed in dollars if the person could open an ABLE account. In the two real court cases whose opinions state how long the help lasted, the same help routed through Giftwise would have kept one family $13,584.53 over 41 months and another $24,518.79 over 37 months, $38,103.32 in all, at 2026 rates.
- Why not pay the shop yourself?In Maya's sample year, one shared page keeps $1,747.33 that 10 of her monthly checks would lose with each giver paying on their own.
- What SSA asked backIn two of the court cases, SSA asked for money back: $4,870.00 (a cousin's deposits push savings over the limit) and $11,442.20 (children deposit $1,500 to meet a bank's minimum balance).
- Who each route reachesItems, care and school for all 7.4 million people on SSI; rent and utilities through ABLE, where the disability began before 46.
- How it comparesTrue Link, Purple, Talli, TrustHelm, Giftly, ABLE gift pages, trusts, DB101, WIPA counselors and fundraising sites, each from its own pages with the one step Giftwise adds.
- Who runs it firstBenefits counselors, state ABLE programs and payee organizations, and a 120-day pilot.
Running it
Checked every night from outside, 13 test suites, the site under load, and what a real deployment changes
- Checked every night, from outsideEvery page on the path, the public API, a gift paid by test card across its payees, and each failure case, on Render Workflows.
- Test results13 suites, each with its command and date; WCAG 2.2 AA on 144 page views.
- Under loadOn the deployed site and on a laptop. Deployed, one visitor at a time: Maya's page in 178 ms.
- Status and OperationsEvery service, latency and errors by route, payments and webhooks, and the alerts, public on two pages.
- Running it for realWhat the demo runs on, and what a real deployment changes, file by file.
How it is built and designed
Every sponsor tool with its file, designed for the people who use it, WCAG 2.2 AA on 144 page views
- How it is builtEvery sponsor tool and PayPal API with its job and file, and Claude's structured outputs checked by the rules engine.
- Designed for the people who use itBuilt to W3C's guidance for cognitive and learning disabilities, Easy Read, GOV.UK confirmation pages and Baymard's checkout research. Light and dark, Large text, Spanish for givers, WCAG 2.2 AA on 144 page views.
- Ask the rulebookElasticsearch finds the passages, Claude answers only from them, and every quote is checked word for word.
- The organizer's sideAdd or change a need and watch Giftwise choose where it is paid, find a product with Channel3, and send each gift to a Zapier webhook.
- The calendar's plannerPress As cash, then Plan the cash dates: the planner moves gifts on the Bryntum calendar and compares the result with paying the shop.
- Maya's year in AG StudioEvery gift and her check each month, compared with cash; press a giver and the Giftwise widget reruns the rules.
- A sample you can resetMaya's page resets to the same starting point at any time, and every earlier receipt still opens.
What is new
A report anyone can check against PayPal, a referee in code over two rulebooks, and an assistant that asks before a gift
What is new
A report anyone can check
PayPal pays only the payees the rules picked, in one order captured once, so every gift has a PayPal record. Each month's report to Social Security carries a code and a QR code: opening it rebuilds the report, asks PayPal live for every payment in it and checks the audit ledger, with no login.
November's report, with real sandbox paymentsTwo rulebooks, and a referee in code
When a bill might be a housing cost, two Claude reviewers read different rulebooks and must quote them word for word. Code checks every quote, takes a reading only when both agree, and otherwise picks the route that keeps her check whole either way: 59 of 59 labeled bills.
One bill, two reviewersAn AI assistant that asks first
Over MCP, an assistant checks a gift before it is given and starts a PayPal checkout that only the giver can approve, then reads the order back from PayPal's own MCP server.
Watch it talk June out of cash
What is first here
Four things none of the compared products does, each tied to its comparison row
- A gift routed before it is the person's money
- A third party's gift is checked by a rules engine built from the statute and paid to the payee the rules picked, and one PayPal order pays up to 10 of them. True Link and Purple control or watch money the person already holds; Giftly delivers cash with a suggestion; ABLE gift pages take money into ABLE only; Talli, TrustHelm and trustees pay vendors from a trust by judgment. The comparison, row by row
- A report to Social Security anyone can verify
- Each month's report carries a code; opening it asks PayPal live for every payment and recomputes the ledger's hash chain, with no login. None of the compared products publishes a report of this kind. November's report
- Two rulebooks, quotes checked by code, a referee
- Two reviewers read two different rulebooks (20 CFR 416 and POMS); code checks every quote; when they disagree, the referee takes the route that keeps the check whole under either reading. Published designs abstain on one judge's confidence or vary the model; this one varies the text each reviewer reads. Where the design comes from
- Accuracy published against SSA's own material
- 40 of 40 of Social Security's worked examples, 4 of 5, 1 of 5, 1 of 5 and 6 of 15 in the four sets the engine never saw, and federal court cases, 14 of 14 scored, of 36 cases, each on its own page. None of the compared products publishes accuracy against SSA's material. How the rules were checked