Research: what I measure, published with its data
I study what AI models do when they are given real work, and I publish the data and the method so anyone can check the result. One paper, one monthly dataset, one benchmark and one study so far.
- Paper
Instrument before you evaluate
Write-path invariants for deployed human-agent systems
A field report from a production human-agent system. Whether such a system can be evaluated is decided by its write path: what it records, with what attribution, and what it may overwrite. The paper names four write-path invariants and closes with a checklist a team can answer before it designs an evaluation.
- Authors
- Guglielmo Paesano, Romanus Berg
- Venue
- Human-AI Coevolution (HAIC) workshop, NeurIPS 2026
- Accepted
- 30 Sep 2026
- Presentation
- Atlanta, 13 December 2026
OpenReview(verifiable source, opens in a new tab)What the paper argues
- 80,059DOI10.5281/zenodo.23188442Open dataset, monthly
AI Visibility Index
October 2026 edition
Which businesses do AI assistants recommend when people ask them? Each month the same questions go to four assistants in every place and language on the list, and every answer is kept and published through AIGROW, the company I founded.
- Answers recorded
- 80,059
- Assistants asked
- ChatGPT, Claude, Gemini, Grok
- Countries
- 20
- Languages
- 11
- Places
- 108
- Business types
- 44
- Licence
- CC BY 4.0
DOI 10.5281/zenodo.23188442(verifiable source, opens in a new tab)aigrow.gg/index(verifiable source, opens in a new tab)
- 71.5%Benchmark, pilot edition
Which office work a model can be left alone with
36 models on 20 office tasks, 720 graded runs. 515 passed, 71.5% overall. The mean hides the finding: the same models that route a request almost perfectly fail most of the time at replying to a review.
| Kind of work | Passed | Rate |
|---|---|---|
| Request routing | 69 of 72 | 95.8% |
| Invoice extraction | 65 of 72 | 90.3% |
| Email triage | 30 of 36 | 83.3% |
| Meeting summary | 30 of 36 | 83.3% |
| Business calculation | 55 of 72 | 76.4% |
| Lead qualification | 26 of 36 | 72.2% |
| Bookkeeping | 51 of 72 | 70.8% |
| Appointment scheduling | 50 of 72 | 69.4% |
| Policy questions | 25 of 36 | 69.4% |
| Customer support | 48 of 72 | 66.7% |
| Product description | 21 of 36 | 58.3% |
| Quote writing | 37 of 72 | 51.4% |
| Review replies | 8 of 36 | 22.2% |
Every task and its grading(verifiable source, opens in a new tab)CC BY 4.0
Also published
Email spoofing: 6,937 local business domains tested, and 80.9% of them could be impersonated by a stranger sending mail in their name.
The study on aigrow.gg(verifiable source, opens in a new tab)
The AIGROW Score: a measure from 0 to 100 of how AI models see one business, always printed with its 95% interval.
aigrow.gg/aigrow-score(verifiable source, opens in a new tab)
An MCP connector that lets Claude and ChatGPT query the score directly, where they would otherwise guess.
A price index of 46 models, published beside the benchmark.
The same method, on your system
A system I build for a client is instrumented the way these studies are: it records what it did, attributed and append-only, so whether it works is answered from data. If you already run an AI system and cannot say how well it performs, that is a project of its own.
LLM evaluation, as a serviceCan AI do it? Work tasks, tested in public
Have a system that needs measuring?
Tell me what it does and what you need to know about it. I read every brief myself and answer in writing with a scope.
Or write to contact@guglielmo.io