Lessons · Lesson 5 of 6
- 01 · A colour is a measurement, and a measurement carries its conditions
- 02 · The screen is an emitter and the garment is a reflector
- 03 · Why a render can never disagree with itself
- 04 · A digital material is two halves, and one of them is usually missing
- 05 · The library nobody curates
- 06 · The library is a lead-time asset, and the unit is days
The library nobody curates
Find the near-duplicates in a colour library, and set the rules that stop a correct search returning the wrong record.
Lesson 5 of 6 · 17 min
The situation
19 February. Tolbooth Navy on TW-4180 has come back from Chandori Mills three times, and each submission is further from the standard than the last. The mill is not incompetent and the merchandiser is not careless. On the third rejection the mill's colourist asks a question nobody at Thistlewaite had thought to ask: why is the standard we were sent measured on a twill, when the last one was a poplin?
The library holds two records called Tolbooth Navy. Both are real, both were properly measured, and both were approved. One is on a 118 gsm cotton poplin for a shirt programme. The other is on the 272 gsm cotton twill this overshirt is cut in. The merchandiser searched for the name, took the record the search returned first, and sent it.
A search that returns the right name and the wrong record is the failure mode of an uncurated library, and no error message appears at any point.
What is really in there
The migration brought across 214 entries. Run them against each other by measured colour instead of by name and you get this.
| Finding | Count |
|---|---|
| Entries in the library | 214 |
| Clusters where entries are the same colour under a different name, or the same name on a different substrate | 19 |
| Entries sitting inside one of those clusters | 51 |
| Distinct colours actually represented | 182 |
| Redundant records | 32 |
32 of 214 records — 15.0% of the library — are duplicates of something already in it. They built up the ordinary way. Somebody could not find a colour, so they created it. A supplier sent a standard under their own name for a colour the brand already held. A season's card was imported whole instead of matched against what existed.
Every one of those 32 is a chance for a search to return the wrong answer, and the chance is not spread evenly. Duplicates cluster on the colours used most, because those are the ones most people needed to find.
Six entries that are simply gone
The audit found a second thing. 6 entries point at measured files that lived on a colourist's laptop. She left in 2025, the laptop was returned and wiped, and the records now hold coordinates and a name and nothing behind them. The colours are still bought. Nobody noticed for two seasons, because coordinates are enough to look like a record, and enough to pass a colour that has already been approved once.
There is no recovery from that except measuring the kept physical standards again, if they still exist. On this library four of the six do.
Five rules, and what each one prevents
The fix is a set of rules about the record, not a piece of software. Every one of these can be applied in a spreadsheet on a Friday afternoon.
- An identifier that is not the name. A name is written for humans. It is not unique, it does not survive a rename, and it gets re-used across seasons. Give every record a stable identifier that nothing else uses, and point every reference at the identifier: a tech pack line, a purchase order, a material record. This is a settled principle in data management generally, not a colour idea. There is an international standard on naming and identifying data elements that says the same thing.
- One master record per colour, per substrate. Not one per colour. The substrate is part of the key, which is what the two navies were telling everybody. The same colour on a poplin and on a twill are two records, deliberately, and both are correct.
- A state on every record. Proposed, measured, approved, superseded. A record with no state is treated as approved by whoever picks it up, because there is nothing to say otherwise.
- Supersede, never delete. A live order points at the old record, and deleting it turns a working reference into a broken one. Mark it superseded, name the record that replaces it, and let it stay findable and unusable.
- *The record's fields include what it was approved for.* Substrate, dyehouse, date. Lesson 6 shows this is also what makes an approval re-usable. The same four fields do both jobs.
| Rule | The failure it prevents |
|---|---|
| Identifier, not name | The navy incident: right name, wrong record |
| One master per colour and substrate | Two people approving the same colour twice, on different cloth, without either knowing |
| A state on every record | A proposed colour being sent to a mill as a target |
| Supersede, never delete | A live purchase order pointing at nothing |
| Approved for what | A re-use offered where the substrate or dyehouse has changed |
What the wrong pick cost, and what the pass costs
| Line | Amount |
|---|---|
| Three lab dips matched to the wrong standard | USD 126 |
| Elapsed time before the mill's colourist asked the right question | 11 days |
| Float remaining in the season's colour calendar afterwards | 9 days |
The USD 126 is nothing. The 11 days is the whole of it, and where it landed makes it worse. Lesson 6 shows the season opened with 20 days of float on the colour calendar. The navy took 11 of them in February, leaving 9. Lesson 3's taupe tape then took the remaining 9 in May. Thistlewaite finished the season with nothing left, and on neither occasion had anybody done anything wrong.
Against that:
| Line | Amount |
|---|---|
| Distinct colours to review | 182 |
| Time per colour, checking against the near-duplicate list and setting a state | 6 minutes |
| Total | 18.2 hours |
| Cost at the studio assistant's rate | USD 254.80 |
USD 254.80 against 11 days. That comparison is the argument, and it is not really an argument about money. A curation pass is two and a bit days of one person's time, done once, with a rule that keeps it clean afterwards.
Check yourselfYour library holds one record for a colour, approved two seasons ago on a fabric you no longer buy. A new style needs that colour on a different cloth. What do you do with the record?Show the answer
Leave it alone and create a second record. The existing one is a true statement about a dyeing on a substrate you have retired. Editing it to point at the new cloth would falsify it and break any historical reference pointing at it. Create a new record for the new substrate, carrying its own measurement once the dyeing exists, and link the two so anybody who finds one sees the other. If the old fabric is genuinely gone for good, mark that record superseded and name its replacement. Superseded, not deleted, because somewhere a purchase order still points at it.
Check yourselfA supplier sends you a measured standard under their own name for a colour you are fairly sure you already hold. What is the cheapest way to find out?Show the answer
Measure the difference. Do not compare the names. Take the supplier's record and work out its colour difference against every entry in your library. If it is very close to something you already hold, you have found the pair, and you keep one master with both names recorded as aliases. This is the same operation as the duplicate pass and takes a minute per incoming standard. Doing it on arrival is what stops the library growing duplicates in the first place, and it is the only moment when both records are in front of one person.