The Luck of the Draw

§ 1 · Holding the case constant

Is it the examiner, or the case they were handed?

Every profile on this site carries a per-examiner number. The fair objection is that examiners do not receive the same cases. A docket carrying more crowded classes will draw more refusals. The examiner working that docket will look harder for those refusals. So we tested it the way you would test any other claim about a docket. Take every application whose first likelihood-of-confusion refusal (Section 2(d)) was issued in 2021 or later. Keep the ones the applicant responded to, and give the record at least ten months to resolve. Score the ones that resolved either way, where we could read the next action: did the refusal come off, or did it stick? Then adjust each examiner's rate for what they were handed rather than what they did. That means the Law Office, the year, the filing basis, whether the applicant had counsel, and the classes on the application.

What the examiner chose to do with the case is left in on purpose. The registrations they cited, the other grounds they raised, the number of rounds they ran: that is the examiner. The applicant chose none of it. Nobody picks their examiner.

That leaves 154,649 responded refusals that resolved either way, across 988 examiners. The 872 with at least 25 resolved each carry the analysis, with a median of 157. They are 819 of the 940 examiners active on the day the record was read, plus 53 who had already left. The refusal came off in 55% of those responded refusals.

§ 2 · The spread

Nearly two thirds at one end, under half at the other.

After the adjustment, an examiner at the tenth percentile of the corps lets the refusal come off 46.4% of the time. One at the ninetieth lets it come off 65.8% of the time. That 19.4-point gap is the spread, not noise. If examiners were interchangeable within their Law Office and year, the same calculation would produce a gap of two to four points. Before a word of the response is written, the draw has already moved the odds by about 19 points.

§ 3 · The years that follow

Four in five stay at the same end.

A number that describes the past is only useful if it forecasts. So split the record at the end of 2023. Take the 630 examiners with at least 25 resolved refusals in each half, rank them on the first half alone, then score them on the second:

Read the table top to bottom. The examiners who were hardest in 2021 to 2023 let 46.6% of their refusals come off in 2024 to 2025. The easiest let 65.0% come off. The groups were ranked once, on the first half, and never re-ranked on the second. Two years on, four in five of the hardest tenth were still on the hard side of the middle. The same share of the easiest tenth were still on the easy side. An examiner's past is a forecast.

§ 4 · Five ways this could be nothing

One ruled out. Four bounded, with the bounds on the page.

Nothing here says which end is right. Examining attorneys apply the same statute to different marks, and a refusal that stuck may have been the correct call. What follows is what the difference is not.

1. Hypothesis 01

   “It's the Law Office, not the examiner.”

   Ruled out

   Compare each examiner only to the others in their own office and the spread is 19.6 points, the same as the corps. Drop the office from the adjustment entirely and the gap widens to 21. The variation lives inside the offices.

2. Hypothesis 02

   “It's the cases they were handed.”

   Bounded

   Before any adjustment the gap was 23 points. Holding constant what the examiner was handed took four of those points off and left 19. Two of those controls are read from the file as it stands today. Drop the class terms entirely and the gap is 19.6. Now also hold constant what the examiner then chose: the registrations cited, the grounds raised, the round. The gap falls to 13.5. Part of that is the examiner's own habits. Part may be the crowding of the field. The number of citations measures that too. We report both. No adjustment can see the mark itself or how crowded its field is. Neither of those picks the examiner. The applicant does not choose the examiner. One exception is on the USPTO's own books: an applicant's companion filings go to the same examiner, and a few desks take particular subject matter. Count one application per applicant per examiner and the gap is 19.5. Drop the one class with a known desk and it is 19.

3. Hypothesis 03

   “It's the unresolved cases.”

   Bounded, stated plainly

   7.6% of responded refusals had not resolved when the record was read. 85% of those were suspended, usually because the cited mark is itself a pending application. The rest were abandoned, still open, or unread. Score each examiner's unresolved cases at that examiner's own resolved rate, and the spread is 20 points. Now suppose every one of them broke against this finding, in the worst possible per-examiner pattern. The spread would be 11.6 points and the persistence 0.42. Nothing suggests that pattern. Across 872examiners the unresolved share barely tracks the rate (rank correlation −0.11). We report both.

4. Hypothesis 04

   “It's how you scored the refusals.”

   Checked where it can be checked

   A refusal came off when the next action no longer carries it, when a notice of allowance issued, or when approval, publication or registration followed the response. It stuck when the next action still carries it. Three quarters of the wins rest on a notice of allowance, which is the USPTO's own record. A sixth rest on our reading of the next action, and those are the ones we tested. The application's status record confirms 72% of them, against 29% of the refusals scored as stuck. Score every one of those 14,877 next-action wins as a loss instead and the spread is still 18 points. One scoring rule was settled after the counts and before any spread. On a use-based filing, a notice of allowance counts as a win only where approval, publication or registration followed the response. It covers 18% of the rows, and undoing it moves the spread by under a point.

5. Hypothesis 05

   “It's only the applicants who responded.”

   Bounded, measured the other way

   Applicants respond to 48% of the likelihood-of-confusion refusals in this record and walk away from 51%. How often an examiner's refusals draw a response barely tracks how often they come off (rank correlation −0.05). Count every refusal the applicant never answered as one that stuck. The same calculation across the whole set gives a spread of about 10 points on a base of 26%, instead of 19 on 55%. The 19 answers the question once a response is filed. The 10 answers it on the day the refusal arrives.

§ 5 · Reading the record

Three things to take to the docket.

1. 01

   Price the draw.

   About 19 points separate an examiner at the easy end of the corps from one at the hard end, on whether a likelihood-of-confusion refusal comes off. A single draw is usually quieter: two thirds of examiners sit within about 8 points of the middle. 19 is the conservative reading. A second method puts it at 22. That method models each examiner directly instead of standardising them.

2. 02

   Read the examiner's own number for what it counts.

   The profile figure is a different count from the one used here. It is how often that examiner's Section 2(d) refusals come off after a Final, once the applicant contests it. Where the sample is thin it is pulled toward the corps average, with the plain count in its tooltip. It is published only at 25 or more contested Finals that resolved. Read a thin sample as a direction, not a decimal.

3. 03

   Know what it cannot tell you.

   It cannot say why one examiner differs from another. It cannot say which argument worked. It says nothing about any examiner by name. It is corps-level evidence that the per-examiner number is worth reading.