A transformer fine-tuning application routes to art unit 2122 at 78 percent, and Examiner 1 is the most likely to receive it at 15 percent, with nine other candidates covering most of the remaining mass. Sample report: the record behind the figures is illustrative.
This report predicts routing from the prosecution record and from the claims as drafted. The office's own classification and docketing decide the assignment; read every figure as a probability, not a promise. It is not legal advice.
The inventive symbol, G06N3/0455 for transformer architectures, has sent 97 percent of its recent first-listed filings to 2122, and the two additional symbols for training and machine learning in general spread across 2121, 2122 and the business method units without disturbing that lead. The forty nearest recent filings by title and abstract all sit in 2122. No parent application was claimed, so no family weight applies, and the unit is decided on the classification picture alone.
Inside 2122 the intake is flat: ten examiners took new cases in the activity window, the busiest at 17 percent and the lightest at 6 percent, and all ten carry every symbol in the picture. The ranking therefore follows intake share and recent docket fit rather than any single dominant examiner, and the one adjustment made, a 1.25 factor for Examiner 1 whose newest cases are transformer fine tuning, moves that examiner from second to first. The top-10 coverage of 78 percent equals the unit probability, because the ten names are the unit's whole active roster; the remaining 22 percent is the chance the application routes elsewhere.
Claim 1 recites computing attention over token embeddings with a transformer and generating text with a pretrained language model, so the inventive concept is the transformer architecture itself, classified at G06N3/0455. Training is claimed in the same independent claim and takes G06N3/08 as a second concept, and the machine learning context recited in claim 5 takes G06N20/00 at the lowest weight. The nearest-filing check carried all three symbols and listed G06N3/0455 and G06N20/00 as the most common first-listed symbols among the neighbours, so the draft stands. A claim 1 that recited the generated text's commercial use instead of the architecture would replace the inventive symbol with a G06Q symbol and move the application to 3689.
| Symbol | Role | Weight | Claim basis | Where it routes | Support |
|---|---|---|---|---|---|
| G06N3/0455 | inventive | 1.00 | claim 1: transformer with attention | 2122 97%, 2121 1%, 2171 1% | exact level; 309 first-listed, 609 carrying |
| G06N3/08 | additional | 0.50 | claim 1: training | 2122 30%, 2121 24%, 3689 24% | exact level; 861 first-listed, 1952 carrying |
| G06N20/00 | additional | 0.30 | claim 5: machine learning generally | 2122 37%, 2121 31%, 3623 30% | exact level; 755 first-listed, 1539 carrying |
The unit is decisive at 78 percent. The second candidate, 2121, holds 9 percent on the strength of the training symbol and allows at 59 percent against 68 percent in 2122, so a routing miss toward 2121 costs about nine points of allowance. The business method unit 3689 appears at 5 percent only because G06N3/08 is carried by some of its filings; the picture gives it no first-listed support. Every candidate unit has an active roster.
| Art unit | Probability | Workgroup | Filings | Allowance rate | Active examiners | Evidence |
|---|---|---|---|---|---|---|
| 2122 | 2120 | 3,299 | 68% | 10 | picture 68%, nearest filings 100%, family none | |
| 2121 | 2120 | 2,739 | 59% | 14 | picture 12%, nearest filings 0%, family none | |
| 3689 | 3680 | 2,171 | 61% | 8 | picture 7%, nearest filings 0%, family none | |
| 2129 | 2120 | 1,608 | 60% | 13 | picture 6%, nearest filings 0%, family none | |
| 3623 | 3620 | 2,131 | 63% | 13 | picture 6%, nearest filings 0%, family none |
| Filed | Art unit | Examiner | Title | Similarity |
|---|---|---|---|---|
| 2024-01-23 | 2122 | Examiner 4 | transformer attention token for improved processing | 0.67 |
| 2024-02-13 | 2122 | Examiner 1 | transformer attention token for improved processing | 0.67 |
| 2024-03-07 | 2122 | Examiner 7 | token transformer attention for improved processing | 0.67 |
| 2024-05-29 | 2122 | Examiner 10 | transformer attention token for improved processing | 0.67 |
| 2025-01-20 | 2122 | Examiner 6 | attention token transformer for improved processing | 0.67 |
The long bar is the probability that the examiner receives the application; the short bar is the probability within the examiner's own unit, which here differs from the long bar by the 78 percent unit probability alone since all ten sit in 2122. One adjustment was made after reading the rosters: Examiner 1 carries a 1.25 factor because seven of the ten newest cases on that docket are transformer fine tuning, the closest dockets in the unit. No other docket read as closer or farther than the symbol statistics already said.
| Rank | Examiner | Art unit | Probability | Why |
|---|---|---|---|---|
| 1 | Examiner 1 | 2122 | 14.6% |
|
| 2 | Examiner 2 | 2122 | 12.9% |
|
| 3 | Examiner 3 | 2122 | 10.7% |
|
| 4 | Examiner 4 | 2122 | 8.8% |
|
| 5 | Examiner 5 | 2122 | 6.3% |
|
| 6 | Examiner 6 | 2122 | 5.4% |
|
| 7 | Examiner 7 | 2122 | 5.3% |
|
| 8 | Examiner 8 | 2122 | 4.8% |
|
| 9 | Examiner 9 | 2122 | 4.6% |
|
| 10 | Examiner 10 | 2122 | 4.4% |
|
As a group the ten allow between 51 and 80 percent of disposed cases, issue about two actions per case, and reach a first action in a little under fourteen months. The 101 share is the figure that separates them: it runs from 16 percent to 58 percent of cases with at least one eligibility rejection, and the leading examiner sits at the high end. The 103 share is uniform at 75 to 87 percent. Interview and 112 cells read n/a because the record carries no interview or 112 fields for this unit, and Examiner 2 has too few disposed cases for an allowance rate.
| Examiner | Allowance rate | Actions per case | 101 share | 103 share | Interviews | Months to first action | Read |
|---|---|---|---|---|---|---|---|
| Examiner 1 | 66% | 2.2 | 58% | 80% | n/a | 14.3 | Allows 66 percent but raises 101 in 58 percent of cases; write the architectural improvement into claim 1 as a limitation, not an argument. |
| Examiner 2 | n/a | 2.0 | 41% | 80% | n/a | 14.0 | No allowance rate yet, with 395 of 397 cases pending; a new docket, and 101 in 41 percent of first actions so far. |
| Examiner 3 | 62% | 2.2 | 36% | 82% | n/a | 13.8 | Allows 62 percent with 101 in 36 percent; the obviousness rejection at 82 percent is the fight to prepare for. |
| Examiner 4 | 77% | 2.1 | 55% | 76% | n/a | 14.1 | The strongest allowance in the top four at 77 percent, but 101 in 55 percent of cases; eligibility is decided on claim text here. |
| Examiner 5 | 51% | 2.2 | 40% | 79% | n/a | 13.6 | The lowest allowance of the ten at 51 percent with 101 in 40 percent; expect a second action. |
| Examiner 6 | 77% | 2.0 | 18% | 87% | n/a | 14.0 | Allows 77 percent with 101 in only 18 percent; obviousness at 87 percent is the ground that matters. |
| Examiner 7 | 58% | 2.2 | 51% | 75% | n/a | 13.9 | Allows 58 percent with 101 in half of cases and the most actions per case of the ten. |
| Examiner 8 | 78% | 2.2 | 27% | 82% | n/a | 13.5 | Allows 78 percent with 101 in 27 percent; a favourable draw if the art record is deep. |
| Examiner 9 | 80% | 2.2 | 16% | 76% | n/a | 13.9 | The best allowance of the ten at 80 percent with 101 in 16 percent; the fastest to first action. |
| Examiner 10 | 66% | 2.2 | 25% | 86% | n/a | 14.2 | Allows 66 percent with 101 in a quarter of cases; obviousness at 86 percent. |
No parent application was claimed, so the routing rests on the classification picture and the nearest recent filings alone.
The lever on this application is whether claim 1 recites the architecture or the use. As drafted it recites the architecture and routes to 2122 at 78 percent with a 68 percent allowance rate. Recast around the commercial use of the generated text, with a G06Q30 inventive symbol, it routes to 3689 at 80 percent, where allowance runs lower and eligibility rejections are the norm. Recast around the network layer structure with G06N3/04 leading, it routes to 2121 at 80 percent, a unit that allows at 59 percent; that framing buys nothing. The draft is already in the best of the three positions, and the steering table shows the cost of drifting out of it.
| Framing | Inventive symbol | Top art unit | Probability | Allowance rate | What the claims would say |
|---|---|---|---|---|---|
| As drafted | G06N3/0455 | 2122 | 77.7% | 68% | Claim 1 recites computing attention over token embeddings with a transformer and generating text with the pretrained language model; the commercial use stays in the dependents. |
| Claimed on the commercial use | G06Q30/0201 | 3689 | 79.6% | 61% | Claim 1 would recite selecting a promotion for a customer audience from the generated text, which puts the transaction in the independent claim and the application in a business method unit. |
| Claimed on the network structure | G06N3/04 | 2121 | 80.0% | 59% | Claim 1 would recite the layer structure and the gradient update of the network, which leads with G06N3/04 and moves the application to 2121 without improving its position. |
The measures come from predicting 300 recent filings held out from the record, each with its own classification picture. The unit stage is accurate on this record and the examiner stage carries the error: the leading examiner is right about one time in four, the top five about three times in five, and the ten listed cover the receiving examiner in 94 percent of cases. What the backtest does not measure is the drafting of the picture from the claims; published predictors working from claim text alone reach about half of cases at top 1 and four in five at top 5 across all units, and the decisive band on this record says the picture is not contested.
| Measure | Value | Sample | Note |
|---|---|---|---|
| Art unit, top 1 | 100% | 300 recent filings after 2026-04-02 | How often the leading unit was the unit that received the case. |
| Art unit, top 3 | 100% | 300 recent filings after 2026-04-02 | How often the receiving unit was among the three leading units. |
| Art unit, top 5 | 100% | 300 recent filings after 2026-04-02 | How often the receiving unit was among the five listed. |
| Examiner, top 1 | 24% | 300 recent filings after 2026-04-02 | How often the leading examiner received the case. |
| Examiner, top 5 | 61% | 300 recent filings after 2026-04-02 | How often the receiving examiner was among the five leading names. |
| Examiner, top 10 | 94% | 300 recent filings after 2026-04-02 | How often the receiving examiner was among the ten listed. |
| Examiner, top 10, given the unit | 94% | 300 of those, within their own unit | The examiner stage alone: the ranking within the receiving unit, with the unit taken as correct. |
| Classification step | not measured | n/a | The figures above take the classification picture as given. Drafting the picture from the claims is the harder step; published predictors working from claim text alone land near half of cases at top 1 and four in five at top 5 across all units, and the confidence card already reflects how contested this picture is. |
The appendix states how the office routes an application and how the probabilities in this report are formed from the record.
A new utility application is classified first. Every claimed concept receives a CPC symbol, and the symbols that carry the claimed invention are marked as the application's classification picture. An automated routing procedure then compares that picture with the portfolio of every examiner, the tally of symbols on the cases each examiner has already acted on, and weighs the size of each examiner's new docket, the backlog waiting in the same symbols, and whether the application continues an earlier one. The examiner who best fits receives the case. The art unit is the consequence of that match, not a separate decision.
The classification picture is drafted from the claims the way the office would draft it. Each symbol carries an observed distribution over art units from recent filings, blended with the units of the nearest recent filings by text, and with the parent's unit when a continuing relationship is claimed. Inside each candidate unit every examiner with new cases in the activity window is scored from the signals in the table, the scores are normalized to a probability within the unit, and the final figure for an examiner is the unit probability times the within-unit probability. An analyst's reading of each examiner's recent docket may adjust a score by a bounded factor, and every adjustment is stated in the reasons.
| Signal | What it measures | Where it enters |
|---|---|---|
| Classification picture | The CPC symbols the claims would receive, with the inventive symbol weighted above the additional ones. | The art unit distribution and the examiner match. |
| Nearest recent filings | Recent applications whose titles and abstracts read closest to this one, and the units that received them. | The art unit distribution. |
| Family | A claimed parent, its unit and its examiner, weighted the way continuing applications are weighted in routing. | Both the unit and the examiner. |
| Intake share | The examiner's share of the unit's new cases in the activity window, which already reflects docket balancing. | The examiner score. |
| Symbol share | The examiner's share of the unit's recent cases carrying this picture's symbols. | The examiner score. |
| Portfolio similarity | The cosine similarity between the picture and the examiner's symbol tally, at symbol, main group and subclass level. | The examiner score. |
| Qualification | The weighted share of the picture's symbols that appear in the examiner's portfolio. | The examiner score. |
| Recency | How recently the examiner last received a new case; an examiner without new cases in the activity window is not a candidate. | The examiner score and the candidate set. |
| Applicant history | Earlier cases of the same applicant with the examiner, which tend to share a classification picture. | A bounded multiplier on the examiner score. |
| Analyst fit | A reading of the examiner's recent docket against this application, bounded to between half and twice the score. | A stated adjustment within the unit. |