# Research Navigator

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Anthony Baptiste](https://cerebralvalley.ai/u/abaptiste)
- **GitHub:** https://github.com/AnthonyASBaptiste/research-navigator-adaptive-loop
- **Demo video:** https://www.loom.com/share/9307073f5e31443080ecdbfb0c84c084
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3955

90% of early-stage startups fail because they build products nobody wants based on false validation. When founders use modern LLMs to research market size, regulatory feasibility, and willingness-to-pay, the models act as chronic sycophants: they summarize high-level web articles, paper over missing local facts, and rubber-stamp optimistic 85%+ viability scores.
In emerging economies like the Caribbean and Latin America, this false AI certainty is lethal. Public digital data is sparse, statutory filings are offline, and local commercial realities are unindexed.
Research Navigator is an Active Validation Coach that replaces sycophantic AI summaries with an adversarial, evidence-calibrated validation engine. Built during the Meta Global AI Developer Hackathon, we developed the Adaptive Epistemic Research Loop—a multi-agent architecture powered by Meta open-weights models that turns false AI certainty into honest, grounded field discovery.
How It Works:
1. Grounded Market Planning: When a founder submits an idea, our Planner Agent (Llama 3.3 70B) extracts core value propositions and binds queries to regional infrastructure (CARICOM, local central banks, inland revenue authorities).
2. Evidence Ledger & Grounding Index: The system gathers market signals, deduplicates domains, and computes a Source Grounding Index (SGI) to ensure claims are rooted in verified local data.
3. Adversarial Epistemic Auditor: Instead of accepting the findings, Meta Muse Spark Open performs a rigorous audit across 4 failure modes: Sycophancy, Contradictory Evidence, Data Deficit, and Unverifiable Habits. It explicitly flags 'Evidentiary Voids' where public proof does not exist.
4. Autonomous Secondary Retrieval: If voids are detected, the agent dispatches targeted secondary search rounds to challenge or confirm the gap.
5. Calibrated Epistemic Refusal: When data remains inconclusive, the engine refuses to fabricate a high score. Instead of a false green light, it awards an empowering 'Field Discovery • Pioneer Opportunity' verdict and prescribes actionable, real-world customer experiments.
6. Vali WhatsApp Co-Pilot: Powered by Fireworks Llama 3.3 70B, Vali transcribes and analyzes founder voice notes from the field, extracting Mom Test facts (past penalties, current manual workflows, willingness to pay) and logging them directly into the venture roadmap.
Empirical Benchmark Results (5-Archetype Test Suite):
- Caught 36 critical evidentiary voids completely missed by single-pass AI.
- Slashed unearned confidence from 89.2% down to an honest, calibrated 47.0% (-42.2% sycophancy reduction).
- Down-calibrated inflated viability scores from 78.0 to a risk-adjusted 36.4.
- Prescribed 12 structured physical founder experiments to prove willingness to pay before writing code.
Judges can explore our live interactive web showcase or verify raw execution provenance in under 10 seconds via our zero-key CLI runner.

## More from Meta Global AI Developer Hackathon

- [Amarizoft](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3952)
- [No Body](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3953)
- [ProtoMind](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3954)
- [Road to 5K](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3956)
- [Research Microscope](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3957)
- [Second Opinion](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3958)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/3955. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt