Two judges sit in the Orlando hearing office. Last fiscal year Maria Mandry allowed 189 of the 286 claims she decided and Douglas Walker allowed 187 of 283. Both come to 66.1 percent, so any scorecard built on approval rates treats them as the same draw.
They aren't. Mandry issued a partially favorable decision on 22 of her 189 awards, which is 11.6 percent. Walker issued one on 65 of his 187, which is 34.8 percent. Per hundred decisions that's 58 clean wins from Mandry and 43 from Walker, off allowance rates identical to the first decimal place, and none of the difference shows up in the one judge number the whole industry tracks.
That's the finding, and it holds across the system. Everything below comes from the ALJ disposition file SSA publishes itself, 1,023 judges and 317,462 decisions.
What a partial takes
Briefly, since everyone reading this knows the mechanics. The judge finds disability and then finds it started later than you alleged. Onset moves forward, the retroactive period shortens, past-due benefits shrink by however many months came off the front, and because the fee is a slice of past-due up to the fee agreement cap, the fee falls at exactly the same rate on any case sitting under that cap. On a long-wait DIB file that already clears the cap on past-due alone, the judge can move onset by a year and the firm collects the same money. So the fee exposure is real and it is bounded, and it concentrates in the smaller cases.
The rest of it isn't bounded. A closed period switches off ongoing entitlement, so the client who was going to send you referrals for a decade becomes a client who has to reapply. And partials are the awards that come back as appeals, or as the phone call where somebody explains why the check is smaller than the number everyone had in their head at intake. None of that gets logged as a lost case anywhere, which is roughly why nobody prices it.
The spread
Nationally, 27,294 of 184,676 awards were partially favorable. That's 14.78 percent, about one award in seven, and the median judge sits at 13.8 percent, close enough to the aggregate that you'd assume the metric was stable, and it is not: the fifth percentile judge partials 4.3 percent of awards and the ninety-fifth partials 32.6 percent. Seventy-nine of them sit at or below 5 percent while another eighty-three sit at or above 30.
At the edges, with a floor of 30 awards so small denominators can't run the show, Brian Lemoine in Hartford decided 138 cases, allowed 109 of them, and issued no partially favorable decisions at all. Not a low number. Zero. Regina Sobrino in Flint decided 113 and allowed 57, and 33 of those awards were partial, which is 57.9 percent.
Take a firm that gets 200 awards a year. Draw the fifth-percentile judge across the whole docket and about nine come back partial. Draw the ninety-fifth and it's about sixty-five.
It isn't geography
The first place anyone looks to explain a gap like this is the map, and the map is a dead end. Only 23.8 percent of the variance sits between hearing offices. The other 76.2 percent is inside them, between judges working down the same hall off the same rotating queue.
Which makes your own office the sharpest version of the problem rather than shelter from it. In the median office with four or more qualifying judges, the distance from the lowest partial share to the highest is 20.2 percentage points, and 26 of 148 offices are past 30.
The pairs are where it stops being a statistic. In Montgomery, Ricky South allowed 68.7 percent of what he decided and partialed 2.1 percent of those awards, while Reni Barnett-Jefferson allowed 67.3 percent and partialed 33.8. Their allowance rates are 1.4 points apart. Their clean awards per hundred decisions are 67.3 and 44.6, a 22.7 point gap hiding underneath two numbers that look the same. Colorado Springs runs from Debra Boudreau at 6.8 percent to Matthew Kawalek at 54.4.
The two boring explanations
Chance first, because with a couple of hundred awards each judges will scatter on their own. We built the null world where every judge in an office draws awards at that office's pooled partial rate at that judge's real volume, and ran it two thousand times. Median simulated within-office gap, 7.4 points. Observed, 20.2. In variance terms chance buys about an eighth of it and no more.
Then the idea that a judge simply had an odd year of cases. If that were it, the number would wander back toward the middle the year after, and it doesn't: across the 533 judge-office cells with at least 100 decisions in both FY2025 and FY2026 to date the two years correlate at 0.85, a little less sticky than the same judges' allowance rates at 0.93 and still behaving like a trait rather than an accident.
The honest counterweight is that the two numbers aren't independent. Allowance rate and partial share correlate at minus 0.48, so a generous judge really is somewhat less likely to partial, and that's about a quarter of the variation accounted for. Three quarters of it the allowance rate cannot see. Pin allowance down to the narrow band between 55 and 60 percent, which is 147 of these judges, and the partial share inside that band still runs from 1.3 percent to 45.3.
We ran the same tests on allowance rates a few weeks ago and wrote up what they showed. The same structure turns up in both, which is what you'd expect if both numbers are measuring the judge and not the docket.
What the file can't tell you
SSA publishes the count of partially favorable decisions and nothing at all about what was inside them. There's no way to see how many months a judge takes off an onset date, or whether a partial was a closed period or a later onset, or what any of it did to a real past-due calculation. The partial share is a proxy for the money and it is not the money. It tells you which judges hand you an award you then have to argue about. What that argument costs, I genuinely do not know.
Reproduce it
It's one file, SSA's public ALJ disposition data for FY2025, and it already carries a Fully Favorable column and a Partially Favorable column, so the metric is just the second one divided by Awards. Drop the rows where Office is SPECIAL REVIEW CADRE, keep rows with 100 or more decisions, and you should land on 1,065 judge-office cells and 317,462 decisions. One warning that cost us an afternoon: match office names by exact equality and never by substring, because ORLAND PARK and ORLANDO are both real offices and a match on "orland" silently merges them into one that doesn't exist. Data and code are at DOI 10.5281/zenodo.21392341 and on GitHub, and the per-office tables are here.
We build the operational side of this for disability firms, which is why we had the disposition file open.
None of this changes which judge draws your file. The rotation does that and it isn't up for negotiation. What it changes is what you think you know when you see the name on the notice, because an allowance rate answers whether you win and says nothing whatsoever about what you win, and the two have been treated as one number for as long as anyone has been keeping score. Mandry and Walker work in the same building and allow at the same rate. One of them turns 58 of every hundred decisions into a clean award and the other turns 43. That was sitting in a public file the whole time.