Clash detection found twelve thousand issues. Nobody fixed any of them.

Somebody runs the federated model on Tuesday night. By Wednesday morning the export is in the coordination folder: a PDF the thickness of a phone book, or a spreadsheet with a Status column that still reads "New" the whole way down. Twelve thousand issues. On the next job it will be eight thousand, or two thousand after grouping, and the digit is the least interesting thing about it. The artefact exists. The issues do not move.

Ask a coordinator what became of last month's report and you tend to get a shrug and a version number. It was run. It was circulated. Somewhere between those two verbs and a poured slab, the fixing was supposed to happen.

The number is not the finding

Four-figure clash reports are real, and most people reading this have produced one. The count gets presented as diligence — we ran clash, it is in the execution plan, here is the evidence — and it is simultaneously large enough that no individual can own an item inside it. Both those things are true at the same moment, which is exactly why the artefact survives from job to job.

Grouping helps a little. Twelve thousand collapses to nine hundred once you strip the duplicates, filter out insulation brushing a hanger, and set the tolerance somewhere sane. Worth doing, and most coordinators do it before anyone else sees the file. Nine hundred unowned issues still do not move.

What a report can and cannot do

Detection is a renderer with a count. Given these models, in this state, with these tolerances, which elements occupy the same space or break a rule. It answers that question well, and far faster than anyone flicking between two drawings on a light table ever managed.

What it has no way of carrying is a next action. A workflow needs an owner, a decision due, and a close condition that is not "we ran it again". A report has a timestamp and a list. You can attach it to a meeting invite. You can save it in the folder next to last month's. None of that assigns a discipline, withdraws a wrong dimension, or changes the drawings the steel fixer will build from.

It is tempting to think the industry bought coordination at the moment it bought detection. Folders have names, models are federated, someone is "on BIM", the licence renews on time. In that version of events the report is the work, opening it is the meeting, and the green cells — if they ever appear — are the close-out. Green is a statement about a register. Whether the duct fits is a statement about a building.

Found and owned are different verbs

"Found" flatters the software. An intersection was computed, an identifier written, a viewpoint saved. What has not happened is the part that costs money and takes nerve: deciding which element moves, which specification yields, which grid is right, and whether this is a real clash or a modelling tolerance dressed up as a risk.

Unowned issues do not stay neutral. They go local. The MEP lead knows the duct will drop below the ceiling zone in that corridor and has said so to his modeller. The structural package still shows the beam at the last issued depth. The coordination model may well show the drop. The issued two-dimensional set does not, and the set is what gets built, because that is what issued means to a foreman with a gang standing behind him.

Ownership is not a name in a column, although a name in a column is usually where it stops. It is the authority to change issued information — or to record, dated and attributable, that the issued information stands and the clash is a model error. If the register cannot name that person for a given row, you do not have assigned issues. You have a list of people who were in the room when the count was read out.

Close-out means something changed

This is the part that goes missing, so it is worth being blunt about what does not count.

Grouping is hygiene. "Reviewed" is a meeting verb: the clash was on the screen and somebody said they would look at it. A fresh run that comes back with eleven thousand is output again — if the published models have not changed, you have photographed the same collision twice and filed both photographs.

On a live job, close-out is one of two recorded acts:

The construction information changes — a revised drawing, a revised model state issued for the stated purpose, a decided dimension the installer can build to — and the old instruction is withdrawn from the people who still have a copy of it.
A dated decision says the issued information stands and the clash is rejected: tolerance, duplicate, accepted deviation, or a modelling error to be cleaned up in work in progress rather than built.

If neither has happened, the issue is open, whatever colour the cell has been painted.

ISO 19650 is unromantic about this, which is one of its better qualities. A model shared for coordination is not published construction information. A clash found in the shared state is a coordination fact, and it becomes a construction fact only when someone moves a container through the gate and issues it. Projects that skip the gate still pour concrete perfectly happily. What they cannot do afterwards is say which clash they had decided to live with.

The meeting that never closes

Wednesday, ten o'clock, a dozen people and a screen. Used properly, that hour does real work. It can pick the ten clashes that will stop first fix on level three. It can push a discipline into committing to a depth in front of witnesses. It can put an engineer and the person who has to install the thing in front of the same view, which is often the whole trick.

What it cannot do, by sitting, is close twelve thousand issues. Going through the report at four minutes an item is a viewing, not a workflow. What comes out the other end is minutes — MEP to resolve, structural to confirm, carry forward — with no revision number, no withdrawn plot and no named recipient at the workface. Those minutes are the report again, in prose.

The same failure travels downstream at the handoff to construction: the model changes hands and the assumptions quietly do not. If the meeting ends and the published set is the same set you walked in with, nothing has been coordinated.

Why nobody fixed them

Because fixing was never the job the report was hired to do. It was hired to prove clash had been run, and that proof has genuine commercial value in a bid, in an execution plan appendix, on a progress slide. It is also cheap compared with what follows: assigning, deciding, revising, issuing, withdrawing, and then checking that the print taped up in the trailer no longer shows the old depth.

The size of the count protects people, too. Nobody gets blamed for item 7,412. Forty owned clashes with two overdue is a management object with names against it, and that is uncomfortable in a way that twelve thousand never is.

The third reason is duller and more structural. Close-out crosses systems the report never enters. The collision lives in a viewer export. The decision lives in an email, or a comment thread, or a coordinator's notebook. The change lives in a native model on a design office server. The issue lives, if it lives, in a published PDF and a transmittal. Until those are one record, "fixed" has nowhere to be true at once — and when the official record cannot be trusted to finish anything, people keep their own copy of what they believe was agreed. That is where the unofficial set in the trailer comes from.

A count worth having

You do not need a better number than twelve thousand. You need to count something else.

Count the issues that have an owner and a close condition, rather than the issues the engine can enumerate. Count close-outs as changes to issued information, or as recorded rejections with a date on them. Then count how many of this week's collisions are last week's collisions wearing new identifiers — that exercise is quick, and it tends to be the one nobody volunteers for.

Stop filing the export as coordination. The appointing party — or the principal contractor, if they hold the coordination appointment — should treat an unowned clash count as incomplete work. When the next report lands, ask for the short list: who has the authority to change the issued information, what the close condition is, and which published container will carry the decision to the people holding the tools. That is the labour the report was hired to avoid, and it costs the assignment, the revision and the withdrawn plot, not another run of the engine.

Keep the clash-run statistic off the progress slide until those three can be named for the collisions that would actually get built. If they cannot, the twelve thousand issues are still open, and nobody has been paid to close them.