A validation certificate exists so that a purchaser does not have to check the vendor’s arithmetic themselves. This page checks it. Your own browser downloads both published reports from the issuer’s content server, hashes them, hashes the individual chart streams inside them, redraws those charts from the PDF’s own bytes, and then recomputes the two percentages printed beside them. One reproduces exactly. Two do not.
The document is Validation Report — Review for: Hinge Health, issued by Validation Institute. Two editions are public: one valid through May 2022, one through May 2024. The certificate page of both reads:
“Claim: Hinge Health users have significantly lower costs for MSK care than similar, matched non-users. … Award Date: May 2021” Certificate of Validation, both editions, signed Benny DiCecca (CEO) and Linda Riddell
The study behind it is a matched-cohort claims analysis of program use in 2018:
“Users and non-users had 24 months of continuous enrollment in their health plan – 12 months before and 12 months after MSK care. A total of 374 Hinge Health users were matched to 374 non-users.” Method / Calculation, both editions
Matching was done on baseline total health care costs, baseline MSK care-related costs, gender, age, and employer — the report lists exactly those five. Not on procedure use. That will matter in section 4.
The report’s money headline is:
“Over 2 years, MSK medical claims spend was $2244 less per Hinge Health participant compared to the matched control group.” Findings & Validation, p. 5
Chart 1 on the facing page breaks the year-1-to-year-2 change into five components for each arm. Add them up and the $2,244 appears exactly: the control arm rises $1,796 per person, the Hinge arm falls $448, and 1796 − (−448) = 2244. That is a real reproduction, and it establishes that the chart and the sentence are describing the same computation. Everything that follows uses the same five numbers.
“The majority of the difference (88.88%) between Hinge Health participants and matched control group stems from lower use of surgery, injections, and emergency room visits.”same paragraph, p. 5
Invasive procedures — surgery, injections, ER — account for $1,656 − (−$309) = $1,965 of the $2,244. That is 87.57%, not 88.88%. Nor is 88.88% any other combination: 88.88% of $2,244 is $1,994.47, which is not a whole number of dollars, and no subset of the report’s five components lands there — the two nearest are 87.57% (invasive alone) and 89.57% (invasive + imaging). The button below enumerates all thirty-one subsets in front of you and shows the gap.
“In year 2, there were 68.7% fewer Hinge Health participants undergoing invasive procedures than the matched control group.” Findings & Validation, p. 6
Chart 2 prints the inputs to that sentence as percent changes from year 1 to year 2: −9% for the Hinge arm on invasive procedures, +175% for the control arm. Those two changes plus 68.7% over-determine the problem, so it can be closed by enumeration. Over every pair of whole-number year-1 counts from 1 to 374 in each arm consistent with the printed −9% and +175%, exactly 24 configurations produce 68.7% — and in none of them do the two arms start equal. Every solution needs the Hinge arm to begin with between 94.0% and 95.1% of the control arm’s invasive-procedure count.
Hold the baselines equal, as a matched design implies, and the reachable values are 66.8%, 66.9%, 67.0% and 67.1% — nothing else. The published 68.7% sits 1.6 to 1.9 points above the top of that range. This is not a rounding quibble: it means the sentence is describing a comparison whose starting point is not the one a reader assumes from “374 matched to 374”, and the report never prints the year-1 counts that would settle it.
Both PDFs are on the issuer’s own content server and both are still linked publicly. The files differ — different cover, different company blurb, different expiry date, and the phrase “Level 1 – Savings validation” quietly became “Savings”. The evidence did not change at all. The two chart images are byte-for-byte the same compressed streams in both documents, at different offsets in different files.
Press the buttons: your browser fetches both PDFs from
assets.ctfassets.net, hashes the whole files (they differ), then hashes the exact
byte ranges holding each chart (they match), then inflates those same bytes and paints them onto
a canvas — so what you see below is drawn from the file you just downloaded, not from an
image I uploaded.
The underlying data is program use in 2018, on 374 people, awarded in May 2021, and re-issued in 2023 with a 2024 expiry printed on it. A purchaser reading “2023 Validation Report” on the cover has no way to see, from the cover, that nothing behind it was re-run.
“Users voluntarily enrolled in Hinge Health. Though the matching took into account several factors, users may have some unmeasured factor that made their results favorable.”Limitations, p. 7
“While Hinge Health participants demonstrated a decrease in total MSK spend, additional research would be needed to determine whether total MSK spend would decrease for a whole population including both participants and non-participants of Hinge Health.”Limitations, p. 7
Both of those are correct and creditable. Neither is the finding on this page. The Limitations name the risk that the two groups differ in something unmeasured; section 4 shows that the report’s own printed numbers require them to differ in something measured — baseline invasive-procedure use — and that this quantity was never printed. That is not a caveat about generalisability. It is an arithmetic constraint sitting inside the certified claim.
The vendor is doing what vendors do. The certificate is the thing a purchaser is invited to rely on instead of checking, and it carries money behind it:
“Validation Institute is willing to provide up to a $25,000 guarantee as part of their Credibility Guarantee Program.” Validation and Credibility Guarantee, p. 8, 2023 edition
A guarantee is adjudicated against a number. If the number on the certified page cannot be reproduced from the chart on the facing page, the guarantee has nothing stable to be adjudicated against.
Every verdict above comes out of a function on this page. Each control below feeds those same functions a case where the answer is known, including cases where the honest answer is “yes” — a check that can only ever fail is not a check.
Which printed number on the facing page does this percentage come out of, and can you run it in front of me?
If the answer is a chart with whole numbers on it, the certificate is checkable and worth what it costs. If the percentage only exists in the sentence, then what has been validated is the sentence.