How we work — artificial intelligence, accuracy and limits
Part of the content on Grants.VarnaHub is generated with artificial intelligence. Here is what, with which models, how we check it, how often we are wrong and what we cannot guarantee.
What is generated with artificial intelligence
- Cost eligibility assessments (“Eligible costs” page) — from each call's guidelines.
- Call summaries (Word) — only points whose text was found in the guidelines.
- The draft in-depth analysis and sample budget — reviewed by a person at KAGAMI before you receive them.
- Translations of call titles and descriptions (Bulgarian ↔ English).
- Suggestions for empty profile fields — from public registers and your website; saved only after “Accept”.
Not generated with AI: the list of calls, deadlines and budgets — they come automatically from the official sources (the EC portal and ISUN). The preliminary check in “In-depth analysis” is rule-based, not AI.
Which models and where they run
- Gemma 4 (26B, Google, open weights) — reads the guidelines and produces assessments, summaries and translations. Temperature 0 (no “creativity”).
- bge-m3 (BAAI) — finds the guideline passages relevant to each cost type.
- Both run on our own server in Varna (KAGAMI). Call guidelines and organisations' data are not sent to external AI services.
How we check
- Every “eligible”, “conditional” or “not eligible” comes with a verbatim quote from the guidelines, and the software verifies the quote is really there.
- Without a verified quote the assessment becomes “unclear” — we do not claim anything that is not written.
- Since 16.10.2026 the quote must be about that specific cost type; otherwise it also becomes “unclear”.
- If a call uses a lump sum, cost types are not reported separately — we show that, not “not eligible”.
- Every night the software checks the whole site's data (fields, dates, references, totals) and raises an alert on errors.
How often we are wrong
We measure accuracy by manually reviewing a random sample of “eligible / conditional / not eligible” assessments (with a quote). “Unclear” is not an error — it means the guidelines do not say.
| Date | Version | Sample | Correct | Note |
|---|---|---|---|---|
| 16.10.2026 | 1 | 40 | 11 (28 %) | before the fix — mostly “lump sum” misread as “not eligible”, and unrelated quotes |
| 16.10.2026 | 2 | 40 | 34 (85 %) | after the fix (lump sum, related quote, scope sentences); with a sample of 40 the true share is likely between 70 and 94%. About 82% of all assessments are “unclear” — guidelines rarely list every cost type |
Therefore: assessments are guidance, not a legal opinion. Final eligibility is checked in the official guidelines and with the managing authority — or in the in-depth analysis reviewed by a person.
When all calls will be analysed
Loading…
Changes in the documents after the analysis
- The analysis is based on the documents available on its date (shown on each call: “analysed …”).
- Each call is checked at least weekly: in ISUN — whether documents were added or replaced (e.g. “Conditions — amendment …”, “Q&A”); on the EC portal — whether the topic description and conditions changed.
- On a change the call is marked “⚠ documents changed — being re-analysed” and is analysed again overnight.
- We do not detect a change in the content of an already published file uploaded under the same name and address — check the document dates in ISUN and on the EC portal.
Legal note
The content is informational and is not legal or financial advice. We label AI-generated content in line with the transparency requirements of Regulation (EU) 2024/1689 (the AI Act). Spotted an error? Write to us with the “✎ Note” button — every note is read and fixed.