Two Locations at GBP 15,000 Got Different Cost Grades on the Same Index

Somebody hands you a one-page index. Eight candidate locations down the side, three criteria across the top, and twenty-four small circles filled in three shades, green for good and white for bad. You are being asked to narrow a field with it, and the only thing you can verify is that whoever built it had a view.

I keep one of these because it fails in a way you can prove on the page rather than argue about. It sits in the market study appendix of a service desk review at a large European public-sector organisation, produced ahead of a retender. The exhibit profiles eight European near-shore delivery locations and then rates every one of them on cost competitiveness, resources and skills, and IT infrastructure. Two locations carry the same printed salary figure and receive different cost grades. That single contradiction is enough to retire the instrument, and the reason it matters in 2026 is that this exact artefact is back, wearing new labels.

What the exhibit actually contains

Two instruments are stacked on one page, and they are of very different quality.

The first is a set of eight fact cards. Poland at 65,336 IT and engineering graduates, 33 per cent English, 19 per cent German, GBP 15,000. Spain at 63,830 graduates, 22 per cent English, GBP 27,000. Romania at 39,315 and GBP 15,000. The Czech Republic at 13,738 graduates and the highest German share of the eight at 35 per cent, GBP 16,000. Bulgaria at 9,272 and GBP 13,000. Lithuania at 7,149, the highest English share outside Ireland at 38 per cent, and the lowest wage of the eight at GBP 12,000. Ireland at 7,079 graduates, 98 per cent English, and GBP 35,000. Hungary at 6,681 and GBP 19,000.

Every one of those is a number with a unit and a subject. A reader who disagrees can fetch a different figure for the same field and argue about sources. That is what a benchmark is for.

The second instrument is the index itself: eight rows, three columns, twenty-four grades on a three-level scale. It carries no weighting, no threshold, no rank number and no definition of what strong, medium or weak mean on any of the three axes. It is the part of the page a decision gets taken from, and it is the part that cannot be interrogated.

The same salary, two different cost grades

Set the fact cards against the grades and the seam opens immediately.

Poland: GBP 15,000, graded weak on cost competitiveness. Romania: GBP 15,000, graded medium on cost competitiveness. Identical printed input, different published output, on the same page, in the same instrument.

There are only three explanations, and the exhibit lets you choose none of them. The first is that cost competitiveness is a composite, so something beyond wage is in it, perhaps facilities cost, employer taxes, wage inflation, attrition or currency exposure. Plausible, and the page names no such variable and shows no such number. The second is that a threshold sits between the two locations and the printed salaries are rounded to a granularity coarser than the threshold, in which case the rounding has destroyed the audit trail. The third is that it is simply an error, which a reader with no thresholds cannot rule out.

None of these can be settled by anyone who was not in the room when the grades were assigned. That is the actual finding. A reviewer cannot reproduce the grade from the data the exhibit publishes, so the grade is not evidence. It is testimony, presented in the visual register of evidence.

A colour is an assertion. It becomes a measurement the moment the threshold that produced it is printed beside it, and not one moment earlier.

The second criterion behaves the same way

Cost is not an isolated slip. Resources and skills shows the same pattern with a different variable.

If graduate supply drove that column, the eight would sort roughly by pool size. They do not. Ireland has the smallest pool on the page and is graded medium, above three countries with larger pools that are all graded weak. Read across Ireland's card and one field stands out at 98 per cent English against 20 to 38 per cent everywhere else, which is a reasonable thing for a criterion called resources and skills to reward. It is also a reading I am supplying, not one the exhibit states. The moment a reader has to reconstruct the rule, the rule was not published.

There is a genuine analytical result buried under all this, and it is worth saying because it argues for the study rather than against it. No location scores strong on all three criteria. The two locations graded strong on cost are both weak on resources and skills. Both locations graded strong on resources and skills are weak on cost. Cheap locations on this index had no talent depth and deep ones were not cheap, and that is a real conclusion, arrived at honestly, which the page never prints in words. A better exhibit would have led with it.

24

Grades on the index, none of them stated as text

0

Thresholds or weightings printed anywhere on the exhibit

3

Criteria, and no location strong on all three

The grade was a fill colour

One more property of this exhibit deserves attention from anyone whose job includes getting evidence out of old artefacts.

The twenty-four ratings are not characters. They are drawing objects, and the level is carried in the fill: one green for strong, a muted variant for medium, white for weak. A text extractor sees an empty region, a search for the word weak returns nothing, and a screen reader announces nothing. Any pipeline reading decks into a knowledge base files this page as a table of salaries and graduate counts and silently drops the ranking that was the point of building it.

Recovering the matrix meant reading fill attributes and resolving positions through the shape groups, and a first pass over this page recorded it as not recoverable. The lesson generalises past PowerPoint. If your conclusion is encoded in presentation rather than in content, then it does not exist for anything downstream of the human eye, and it has no audit trail even for the human eye, because a colour carries no provenance.

What a defensible version looks like

The correction is not complicated, and the same review contains a worked example of it. The supplier market study run alongside this location index distributed ten suppliers into quartiles per measure instead of collapsing them into a ranking, used the six measures the contract already carried, and printed its own limitations on the face of the analysis. That study was built to be defended in front of the suppliers who lost, and the reasoning is set out in You Will Have to Defend the Shortlist to the Suppliers Who Lost. The location index, in the same review, was built to be looked at.

Three rules follow, and they cost almost nothing.

Publish the thresholds, or publish the raw variables and let the reader grade. Either is defensible. A grade with neither is an opinion in a table.

Put the grade in text next to the colour. The colour is for scanning, the word is for the record, and if a criterion is a composite, name the inputs and their weights even when the weights are round numbers somebody chose in a meeting.

Reconcile the grades against the printed data before the page ships. Two identical inputs with different outputs is a check a spreadsheet runs in a second, and nobody ran it here.

The same instrument, relabelled

This was a 2014 snapshot of a near-shore market, and it would be a museum piece if the form had died with the decade. The three-colour grid over an unpublished rubric is the standard shape of the AI vendor comparison, the model evaluation summary and the readiness heatmap now circulating inside enterprises, with columns marked capability, cost, governance, data residency and support maturity.

Every argument above transfers. If two models with the same published token price come out at different cost grades, say what else is in the number, whether that is context window, cache behaviour, retry rate under load or the price of the evaluation harness around it. If a governance column is graded green, state the control that was tested and the evidence that was accepted. And put the grades in text, because these grids are now read by other software as often as by people, and a fill colour is not data.

Source: the market study appendix of an IT service desk review at a large European public-sector organisation, ahead of a retender. The location index and its underlying benchmark data are a third-party market study carried inside that review, not our own research, and the figures above are reported as the exhibit prints them. The rating matrix is not printed as text anywhere in the file; it was recovered from the fill attributes of the rating shapes and mapped positionally. The review was produced to inform a retender decision; the source records no sourcing outcome and none is claimed here. The reading of it, including the two contradictions and the three rules, is ours.