The Facts vs. Inferences Taxonomy is a knowledge governance framework for AI systems that categorizes information into four distinct epistemic states: FACT (inspected evidence), INFERENCE (logical deduction), HYPOTHESIS (testable growth assumption), and UNKNOWN (declared missing data) to prevent compounding AI hallucinations in strategic decision-making.
- Eliminates silent conversion of AI guesses into canonical company strategy.
- Maintains verifiable citation provenance (exact source URLs and timestamps) for every data point.
- Requires human founder confirmation before an inference is treated as canonical truth.
The Compounding Hallucination Problem
In standard LLM workflows, a model generates answers based on statistical next-token prediction. While this produces articulate prose, it creates a fatal vulnerability for go-to-market strategy:
If an AI tool incorrectly assumes your product is an "Enterprise HR platform" when you actually build a "Developer CLI tool," every subsequent prompt, including pricing recommendations, target customer personas, cold emails, and ad copy, compounds that initial hallucination into strategic catastrophe.
The 4 Epistemic Knowledge Tiers
| Tier | Definition | Verification Standard | Downstream Usage |
|---|---|---|---|
| FACT | Directly observed data from your website, pricing page, or confirmed founder edits. | Inspected URL / direct user input | Safe for immediate reuse by all agents |
| INFERENCE | Logical deductions drawn from multiple facts (e.g. inferring target buyer based on feature tiering). | Requires corroborating external evidence | Flagged as draft; requires founder verification |
| HYPOTHESIS | A proposed strategic angle or growth experiment that has not yet produced live market data. | 7-day sprint validation test | Tested in sandbox; never treated as established truth |
| UNKNOWN | Information that cannot be determined reliably from existing public or internal data. | Explicitly declared; never fabricated | Prompts founder for input or schedules targeted research |
How XGrowth Implements Epistemic Grounding
Within the Product Brain agent, all retrieved company data is parsed into structured, owner-scoped memory cards. When other agents (such as Competitor Intelligence or The Pitch & Copy) request context:
- Only FACT cards are injected into the primary system prompt without disclaimers.
- INFERENCES are visually highlighted with blue tags so the founder can inspect the rationale.
- When competitor data is incomplete, the system marks the metric as UNKNOWN rather than inventing pricing or feature claims.
Explore Related Architecture Guides
Frequently Asked Questions
Because Large Language Models generate plausible-sounding guesses. If an AI system treats an inference as a verified fact, that error compounds down the pipeline, leading to fabricated value propositions, imaginary competitor features, and wasted marketing spend.
1) FACT: Inspected directly from first-party URLs or confirmed by the founder. 2) INFERENCE: Logical conclusion derived from verified facts. 3) HYPOTHESIS: Testable growth assumption requiring market validation. 4) UNKNOWN: Missing data explicitly reported rather than guessed.
XGrowth tags all stored knowledge with provenance URLs and certainty ratings. Inferences cannot be reused by downstream agents (like copywriting or campaign creation) until explicitly verified by the founder.