01
Who publishes the numbers
A census of corporate research in Greece: 68 publication programmes, 64 publishers, 15 sectors. Who produces the figures business decisions rest on, how far the press carries them, and which ones answer engines cite.
INDEPENDENT RESEARCH
Numbers nobody
else published.
Seven papers on Greek business and on what AI systems actually cost to run. Each one publishes the data behind it, the code that produced every table, and a ledger of every claim considered, including the ones that failed checking and never made it into the text. Free, no form, no sign-up.
- 7 papers
- 68,304 words
- 308 claims, 0 unresolved
- 78 data files
01 - GREEK MARKET RESEARCH
What it actually costs to sell online in Greece, what a Greek order really carries, what 900 Greek shops send to a browser, and who publishes any of these numbers in the first place.
02
What a Greek online order actually costs
What the €36.1bn e-commerce figure counts and what it leaves out, plus a landed-cost model built only from published tariffs that you can run against your own basket.
06
The Greek e-shop technical audit
900 Greek domains measured on one day. 36.8% send no security header at all, 11.1% fail a validating TLS handshake, and five shops out of 505 pass five elementary checks. No site is named.
07
The Greek SME digital bill
What a year of selling online in Greece costs, priced only from published pages. Of 32 cost lines, six have no published price at all, and they are disproportionately the mandatory ones.
02 - MEASURING AI SYSTEMS
What an agent costs per run, what makes one work, and why the same year produces both a 5% and an 88% AI adoption figure. Measured on running systems, including the versions that got worse.
03
The adoption gap
Twenty widely-cited AI statistics graded on six criteria. Why the same year yields 5% and 88%, where Greek firms actually sit once size classes are separated, and why the skills-gap explanation does not survive the spending data.
04
What actually makes an agent harness work
Three agent harnesses on one model, 45 graded sessions. Includes three faults found in the measuring apparatus itself and two published findings withdrawn as a result.
05
What an agent actually costs
Seven versions of one production agent on one benchmark, including the two that got worse, and a negative control showing the suite scores 22.8% for doing nothing at all.
THE APPARATUS
Nothing here asks
to be taken on trust.
Every paper ships the machinery that produced it. That is unusual enough to be worth stating plainly, because it is the part that lets you disagree with me on evidence rather than on tone.
The data, in full
Every figure in every paper recomputes from files published beside it: the crawler, the benchmark, the cost calculator, the raw results.
The claims that failed
Each paper carries a ledger of every factual claim considered, including the ones cut, downgraded or failed in checking. The ledgers were written before the prose, which is the point of them.
The mistakes, in the body
Five measurement errors were made and caught during this programme, and each is written up where it happened rather than in an appendix. Two had already produced published findings, which are withdrawn.
The corrections, dated
Every change since publication is in a public log with its date and its reason. A corrections policy nobody can inspect is not a policy.
TERMS
Free, and yours
to build on.
Everything here is published under a Creative Commons Attribution 4.0 licence: copy it, quote it, chart it, build on it, including commercially, with credit. Analysis code is MIT. There is no email gate on any of it, because a paper that argues you should check rather than trust cannot ask for your address first.
INDEPENDENCE
Nobody paid
for any of this.
No sponsor, client or vendor funded, reviewed or approved this work. Where my own businesses or software are the subject, and in three of the seven they are, that is stated in the first paragraph of the paper as well as in its declaration of interests.
Found a mistake?
Corrections are welcome and are recorded with the date and the reason in the public change log rather than edited in silently. If a number here is wrong, that is worth more to me than the compliment.
Tell me what is wrong