Methodology · Anti-Hallucination

How to read source labels in your analysis

AI-generated analyses combine deck evidence with investor-style interpretation. We flag every claim by source and grounding level so you know what to cite verbatim, what to verify, and what to treat as useful interpretive reasoning.

Why this matters

A pitch deck analysis is only as useful as the source context behind each claim. Without source attribution, “deck shows $30bn TAM” and “the team is a strong ML/healthcare combination” look the same — but one is a fact you can cite to an investor and the other is investor-style interpretation that should be checked before you quote it as fact.

Our pipeline runs an explicit source-grounding pass after the main analysis. Every claim gets a source tag, a grounding label, and a slide reference when applicable. The widget at the top of your report summarises this for the entire analysis; per-section badges drill into specific parts.

The four sources

From Deck

Claim is directly supported by text, numbers, or visuals on a specific slide.

Highest. Cite verbatim to investors.

Founder-Provided

Claim comes from the BSA briefing form (your own answers) or other context you supplied.

High, but it's a self-declaration. An investor will still ask for evidence.

Public Knowledge

Widely-known fact about the sector or market (e.g. SaaS gross margin benchmarks).

Medium. May be outdated; verify timing-sensitive numbers.

AI Interpretation

Investor-style reasoning generated from the surrounding context. Not directly stated in the deck, and not necessarily public knowledge.

Useful as a hypothesis and gap detector. Verify any specific names, numbers, or quotes before using them as facts.

Source-grounding labels

Mostly Deck-Grounded Confidence

≥80% of claims in the section are deck-grounded. You can lift this section's text directly into investor conversations.

~

Deck + Interpretation Confidence

50-80% deck-grounded. Most points are solid, but verify key numbers and named entities before quoting.

i

More Interpretive Confidence

Under 50% deck-grounded. Treat the section as investor-style reasoning: useful for surfacing gaps in your deck, but verify specific claims before quoting them.

Submitted-deck grounding

The percentage of claims in a section that we could trace back to a specific slide. A lower percentage doesn't mean the analysis is wrong — it usually means your deck didn't have an answer to that question, so the model added interpretation.

Founder tip: A consistently low deck-grounded percentage across multiple sections is the strongest signal that your deck needs more concrete details — specific numbers, named customers, source citations. Re-run the analysis after adding them and you should see grounded % climb sharply.

How to use this in practice

  1. Mostly Deck-Grounded / From Deck sections— quote them verbatim. They're your own deck content reflected back, and you can use them in investor follow-ups, partner emails, or your data room.
  2. Deck + Interpretation / Public Knowledge sections — read, cross-check timing-sensitive claims against current market data, then use.
  3. More Interpretive / AI Interpretation sections— read them as hypotheses, not facts. Ask: “is the model right that this is a gap in my deck?” If yes, fill the gap with your own real data.
  4. Investor Q&A questions — questions tagged AI Interpretation are still valuable. They represent what a less-informed reader might ask, which is realistic since not every investor knows your sector deeply.

Worked example

Flagged claim from a real analysis:

“The deck completely ignores NASA's TDRSS, ESA's EDRS, the two most established systems in existence.”
AI Interpretation·Slide 8

What this means: The AI thinks your deck skipped two important existing competitor systems. Those systems are not mentioned anywhere on the slide. The AI brought them in from its own knowledge.

What to do:

  • Don't treat “TDRSS” and “EDRS” as established facts you should reference — verify they're real systems first (in this case, they are).
  • Ask: “is the AI right that I should address competing satellite-relay systems?” If yes, add a competitive comparison slide.
  • If you do add it, use your own research, not the AI's names.

Why we treat interpretive claims differently from deck facts

A reasonable question: if an interpretive claim contains a made-up competitor name — say “TDRSS” or “EDRS” — why don't we run it through Perplexity or a web search and replace the AI's guess with a verified answer?

Because the value of an AI interpretation isn't in being factually correct. It's in showing you what an outside reader assumes when your deck doesn't answer the question. If a casual reader fills in a gap with a made-up competitor, that gap is real even if the specific competitor name is wrong. The signal is: “your deck has space for the reader to imagine things — fill it with your own facts before they fill it with theirs.”

An external fact-check would replace the reader's imagination with a textbook answer and lose the diagnostic value. Worse, it would falsely suggest the verified facts are what your deck should say, when in reality you might want to say something completely different — a different competitive frame, a different positioning, your own data.

We'd rather show you an interpretation flagged honestly as the AI's own input than a sanitised report that hides what an outside reader actually fills in. Treat AI Interpretation as a mirror of how a stranger reads your deck, not as a fact about the world — and use it to add your own facts where the stranger guessed.

What we do to minimise hallucinations

  • Fact-check pass. After the main analysis, we run a separate Claude pass that compares every claim against the original slide text and flags fabricated metrics, invented people, misquotes, and unsupported facts. Critical errors get rewritten in-place.
  • Source-grounding labeling pass. A separate pass classifies every remaining claim by source so the badges you see are independently assessed, not self-reported by the analysis itself.
  • Conservative truncation handling. If the fact-check or labeling response gets truncated by output limits, we retry with stricter parameters. Better to spend an extra few seconds than ship an unreviewed report.
  • No silent failures.If a labeling pass outright fails, the badge says so. We'd rather show “could not classify” than fake source grounding.