Research › Reading targets
When shots land on the wrong bull
On this page
GroupLab project. Researched and written with Claude. Testing and data collection by the developer. Written 2026-09-22; the data is from 2026-09-20 to 2026-10-06.
No figure: The article gives the offsets and the counts. The developer's standing consent of 2026-09-24 now allows those range sheets to be shown, and no figure has been drawn from them yet, including the three sheets it describes.
What we found
Nearest bull is the wrong rule whenever the rifle is not zeroed for the load. On one 20 shot sheet every single shot landed nearer a bull it was not aimed at, and the group that comes out of reading it that way is tight, confident and meaningless.
How sure we are
Three sheets from one afternoon, with the shooter's own written record of which bulls were aimed at. Enough to show the failure clearly. The fix is built and held by tests on these sheets, but five sheets from one shooter are not a measure of how often it is needed.
Why a bull matters at all#
GroupLab measures each shot from the bull it was aimed at, and pools those offsets into one group. That is what lets a 25 bull sheet with one shot per bull give you a 25 shot group: every bull is its own aiming point, and the group is about the rifle rather than about your aim.
It only works if the software knows which bull each shot was aimed at. The obvious rule is "the nearest one".
Where that falls apart#
The developer shot a 6 ARC load at 100 yards through a rifle zeroed for something else. He aimed at bulls 2, 3, 4 and 5 of every row, twenty shots.
Every shot landed high and left of its aim point, by more than the distance between two bulls.
So every one of the twenty is nearer a bull it was not aimed at. Nearest bull assigns all twenty to the wrong bulls, and then measures each shot's offset from that wrong bull. The offsets come out small, because each shot is close to the bull it got assigned to. The group comes out tight.
The number is precise and it is about nothing at all. It is not the rifle's dispersion; it is a measure of how neatly a constant offset happens to line up with a grid of bulls.
The second sheet: two different offsets#
The .22 LR sheet was shot in a strong, shifting crosswind, and the windage was adjusted between row 2 and row 3. So rows 1 and 2 share one point of impact, rows 3 to 5 share another, and inside each the wind moved the shots around by a real amount.
There is no single offset to correct here. A sheet-wide correction would be wrong for one half of the sheet whichever half it was fitted to.
The third sheet: a whole row low, and one shot from nowhere#
The primer test sheet had five shots on row 1 where they were aimed and five more aimed at row 2 that landed on row 3. One of those five landed far from everything else, low and left, next to bull 21. The developer fired it, saw where it went, and does not know why.
So this sheet has a group that is where it should be, a group that is a whole row low, and a single shot that is neither.
What GroupLab does today#
It says it is unsure, which is the important part, and it lets you move a shot onto the right bull by hand. It does not quietly hand you a tight group built on wrong assignments.
The review queue names the shots whose nearest bull is not the bull the matching gave them, and the figures carry their doubt. On the 6 ARC sheet it raises the problem rather than reporting a group.
What it will do#
This part was written in September 2026. All three are in GroupLab now, and the sections after it say how they turned out. Three things, in the order they matter.
You tell it which bulls you aimed at. Click them on the sheet, or use a preset: every bull, a row, bulls 2 to 5 of every row. That one piece of information solves the 6 ARC sheet outright, because with twenty aim points named and twenty shots to place, the assignment is a matching problem with one answer rather than a guess.
Then it can work out a common offset. Once it knows which bulls you meant, the shift is measurable, and a sheet where everything landed a bull low becomes a sheet it can read correctly.
And you can still move any shot by hand, because the shooter saw where the shot went and the software did not.
A full sheet, every shot about 0.8 inch high#
On 2026-09-26 the developer shot a full 25 bull sheet of 6.5 Creedmoor at 100 yards, one shot per bull. The rifle's zero that day put every shot about 0.8 inch high of its own bull, just outside the 1 inch ring, with the bulls 1.5 inches apart. So each shot is nearer the bull above it than its own, and the top row's shots sit up by the codes and the title.
Matching every shot to a bull one to one, as GroupLab did then, gave 5 of the 23 shots it found to the bull above their own. The group came out with a mean radius of 0.59 inch, from shots measured 0.6 inch low on one bull and 0.9 inch high on the next.
GroupLab now reads the whole sheet first. When a sheet has at most one shot to each bull and the shots all landed off by the same amount, which is what a rifle's zero does, it works out that amount and gives each shot to the bull it was fired at. It is used only when it is certain, and it says so on the analysis screen: all shots are about 0.78 inch high and 0.37 inch left of their bulls. On this sheet every one of the 23 went to its own bull and the mean radius is 0.21 inch. Give each shot to its nearest bull instead undoes it. The scan is kept as a test, so this can never quietly break again.
The two shots nearest the top edge were not found at all: they landed in the printed codes and the title, where GroupLab does not look for holes. That is its own defect, and the reason to zero first rather than rely on this.
A row higher fits just as well#
On 4 October 2026 the developer fired one shot of .300 Norma Magnum at each of bulls 1 to 15 of a 25 bull sheet and photographed it on a phone. Every shot landed about an inch low and right, so each hole sat nearer the bull below its own, and the phone gave most of them to that bull.
The whole-sheet reading above was there for exactly this, and it stood down, rightly unsure. Read as fired at bulls 1 to 15 with one common offset, the holes fit with a scatter of 0.312 inch. Read one row higher, as fired at bulls 6 to 20 with a different offset, they fit with exactly the same scatter. On a grid of evenly spaced bulls the holes alone cannot tell those two apart: moving every aim point down a row and the offset up a row describes the same holes.
What does tell them apart is how a sheet is shot. A load sheet is fired from bull 1, in order. So GroupLab was changed to take, among the readings that fit as well as the best, the one whose shots sit on the first bulls, allowing one hole in ten outside that order. All three of the developer's photographs of that sheet, and the phone's own, then gave bulls 1 to 15, as the shooter's drawing of the sheet does.
When the tie is real, it asks#
Two days later the same sheet, photographed on a tablet, gave two different answers on two runs. Both runs found the same 14 holes. The difference was one hole added by hand at bull 10, where the shot had torn the paper's right edge and nothing found it. With 15 holes, the one in ten allowance let the reading a row higher pass as well, the two readings tied, and GroupLab fell back to the nearest bull. Tried at 399 places a thumb could have put that hole, the rule found no offset at any of them.
So the matching became one problem. Every reading is a one to one matching of holes to bulls around one common offset, re-centered on the middle of the holes and repeated. The order of preference is:
- Holes that fit the bulls as aimed, with no offset, as well as any other reading are read that way, so a zeroed sheet is never moved. A scan from 20 September shows why this comes first: with its holes partly merged, a shooting order 0.8 inch off fitted almost as well as the truth.
- Otherwise the shooting order, bulls 1 to k for the fewest k that fit, when it fits as well as any reading and its offset is smaller than the spacing between bulls.
- Then a single reading that clearly fits best.
- When none of those holds, GroupLab asks Which bulls did you fire at?, with its own guess first, up to two more readings, and Choose the bulls myself.
On the tablet photograph it now gives bulls 1 to 15 at all 399 places, without asking. Each hole is drawn with a line to the bull it was given, so a wrong reading shows at a glance, and a click on a bull lights its shots.
The lesson for anyone building this is that a sheet of one shot a bull is a matching problem with a prior, not a set of nearest neighbors. The order people shoot in is information, and a tie the holes cannot break is a question to ask, not a guess to hide.
What you can do now#
If your rifle is not zeroed for the load you are testing, either zero it first, or expect to move some shots by hand and check the review queue before you trust the figures.
And if a group looks better than you shot, look at the assignment before you believe it. A tight group from a sheet you know you scattered is not good news; it is the software measuring something else.
What this means#
Zero the rifle for the load before you shoot a multi-bull sheet. That is the whole practical finding, and the 6.5 sheet says it again: zero first, or a sheet of one shot per bull cannot tell whose shot is whose. On one twenty shot sheet every single shot landed nearer a bull it was not aimed at, and what came out was a tight, confident and completely meaningless group.
Be suspicious of a result that is too good. Reading shots to the nearest bull turns a consistent offset into twenty small groups, each one clustered around a bull. The numbers look better than the shooting was. That is the dangerous shape of this failure: it does not look like an error.
Say which bulls you fired at when you know: GroupLab now lets you, by rows, columns or a list, and what you say always wins over its own reading. Write the bulls down at the bench too. It takes a moment and it is the only thing that survives a wrong assumption about the zero.
Sources#
- The three sheets and the shooter's own record of them:
docs/NOTES-FROM-PLANNING.md, entry 120. - What GroupLab did with them:
docs/PHASE1-RESULTS.md, \"Entry 120\", sections 2 and 7. - The sheet shot a bull low, and the shooting order:
docs/notes/archive/results-351-375.md, \"Entry 374\", the C bull sheet. - The tie on the tablet, the 399 places and the question:
docs/notes/archive/results-376-400.md, \"Entry 376 Part A\". - The code:
src/GroupLab.Core/Detection/ImpactOffset.cs,ImpactOffsets.ReadWholeSheet, held bytests/GroupLab.Core.Tests/Detection/ImpactOffsetTests.cs.