We don't trust a single source. Each AI model has different training data, different retrieval strategies, and different blind spots. By posing the same research question to three independent models, we get triangulated intelligence where each model's strengths cover another's gaps.
The prompt: "Map every Dutch government funding program available to early-stage AI/tech entrepreneurs in 2026-2027. Include exact amounts, rates, eligibility, deadlines, and strategic positioning."
Why multi-model matters for trust
If one model says WBSO budget is €1.817B and the others agree, confidence is high. If only one model reports a program exists, we flag it unverified. If they disagree on a number (Take-off: €40K vs €60K), we investigate: the €60K is WO track, €40K is HBO track. Disagreement reveals nuance that single-source research misses entirely.
Agreement = confidence. Disagreement = investigation. We compared every data point across all three outputs. The table below shows how conflicts resolved into higher-quality data than any single model produced alone.
| Data Point | Claude | GPT | Gemini | Resolution |
|---|---|---|---|---|
| WBSO 2026 budget | €1.817B | €1.817B | €1.817B | 3/3 confirmed |
| WBSO starter rate | 50% | 50% | 50% | 3/3 confirmed |
| Take-off Phase 1 amount | €60K WO / €40K HBO | €60K (unspecified) | Not covered | Claude most precise: two separate tracks exist |
| MIT AI 2025 success rate | 14/65 = 21.5% | Not mentioned | "Highly competitive" | Claude sourced InnoVein; exact figure verified |
| STUDI program (€28.3M) | Not found | Full details | Not found | Only GPT found it. Verified on RVO.nl |
| LIOF fund structure | Basic listing | Minimal | LVFF + LSCF + MKB full | Gemini had deepest regional detail |
| DEI+ 2026 budget | Mentioned | Mentioned | €134M, first-come mechanism | Gemini had exact budget + mechanism |
| OPZuid ERDF 2026 | Not found | Not found | €10.9M, 50%, defence/transition | Only Gemini covered southern ERDF |
| Eurostars NL budget | €22M/yr, €500K cap | Not detailed | €22M, 50%, Call 11 open | 2/3 confirmed with exact matching figures |
Governments organize by program name. Entrepreneurs think in problems. No founder wakes up saying "I need MIT Haalbaarheidsprojecten." They say "Where do I start?" or "Is this worth my time?" We identified the top questions by reading what the data structurally reveals when you ask: "what would a person with zero context need to decide?"
Signal: Information overload
48 programs across 9 tiers. The data has a natural ordering: non-competitive programs first (guaranteed), competitive last (risky). Low-TRL first, high-TRL later. This ordering IS the answer to "where do I start?" We just had to surface it.
Signal: Eligibility confusion
Every program defines "SME" differently. WBSO's definition differs from MIT's. The data contains these distinctions already. Arranging them as a matrix against founder profiles makes years of confusion scannable in 10 seconds.
Signal: Opportunity cost
We had two numbers: amount and competition level. Multiplying amount by success rate and dividing by estimated effort gives expected-value-per-hour. This metric exists nowhere on any government website. But the data to compute it was there all along.
Signal: Sequencing blindness
The data has deadline dates AND TRL ranges. Plot them together and a natural calendar emerges: WBSO first (rolling), MIT Haalbaar April 7, Take-off rolling, MIT AI May. The sequence writes itself from two fields we already had.
The key insight
We didn't survey 1,000 entrepreneurs. Every dataset contains implicit questions. The 48-program matrix, when sorted by certainty and effort, answers "where do I start?" without anyone asking. The eligibility criteria, when cross-referenced against profiles, answer "am I eligible?" automatically. The job of synthesis is surfacing answers the data already contains.
Every answer on the "Top 3" page is the output of a formula, not a feeling. Here's the exact computation behind each one:
Answer 1: "Start here" sequence
Ranking formula: (1 - rejection_probability) * amount / estimated_hours
WBSO scores highest: 0.95 * €60,000 / 4 hours = €14,250/hr. MIT R&D AI scores lowest: 0.21 * €350,000 / 120 hours = €613/hr. The sequence isn't advice. It's arithmetic applied to three known variables.
Answer 2: Eligibility matrix
For each of 48 programs, we extracted every eligibility criterion. We defined 6 common founder archetypes. For each cell: if ALL criteria met = YES. If one criterion requires a fixable action (register BV, get LOI, relocate) = CONDITIONAL. If a structural disqualifier exists (wrong TRL, wrong sector, no university link) = NO.
Answer 3: Effort/reward quadrant
X-axis derives from competition type: non-competitive = 4-10hrs, first-come = 20-50hrs, competitive = 80-200hrs. Y-axis uses known success rates or proxies: non-competitive = 90%+, first-come = 50-70%, competitive NL = 20-45%, EU = 5-15%. Bubble size = max amount. Every position is computed from data.
Why "show your math" builds trust
- Reproducibility. Run the same prompts through the same models. You'll get similar data. The methodology is open, not a black box.
- Traceability. Every number traces back to a program, a model output, and an official source URL. Click through and verify yourself.
- Visible disagreement. Where models disagreed, we show it. Where data is estimated (effort hours), we label it estimated. No false precision.
- Formulas over feelings. "WBSO first" isn't advice from an expert. It's the mathematical result of ranking by expected value per hour. You can disagree with the formula. But the formula is transparent.
- Named gaps. We explicitly say what we don't know: exact processing times, real success stories, application templates. Honesty about limits builds more trust than fake completeness.
This is v1: manual synthesis into structured answers. The next stage automates, personalizes, and expands. Same transparency principle at every stage.
The principle that never changes
Show your work. Every recommendation traces to data sources, computation logic, and confidence levels. The tool gets smarter. The transparency stays constant. That's how trust compounds over time.