demo_video_script.md

FALSIFY — 30-second demo script

A tight, judge-ready narration. Total runtime ~30s of talking over one python main.py --demo run. Capture the terminal + the output/graph.html for the GIF.

---

Setup (before recording)

• Terminal with a dark theme, font large enough to read on playback.
• .env has a working LLM_API_KEY (so embeddings + recall are real).
• Run once beforehand to warm caches; record the second run.

---

Beat sheet

[0:00–0:05] The hook

"Every AI memory demo fixes forgetting. FALSIFY fixes something worse — an AI that confidently remembers a fact that's been proven false."

Run: python main.py --demo

[0:05–0:12] Session 1 — the belief graph

"We're investigating: did Company X know about the defect before the recall? Two hypotheses — a March QA report, and a January supplier email. A conclusion rests on that March report."

On screen: the BEFORE panel — every node tagged ALIVE in green.

[0:12–0:20] Session 2 — one contradicting fact

"Now one line arrives: a forensic audit found the March report was back-dated. Watch."

On screen: the revision log —

✗ refuted: 1 evidence node
✗ invalidated: 1 conclusion
hypothesis A → ↓ superseded
hypothesis B → ↑ promoted (new frontier)
🗑 forgotten: Company X knew about the defect by March 2021

[0:20–0:27] The AFTER + scoreboard

"The March evidence is red. The conclusion built on it collapsed and was surgically deleted — from the graph and the vector store. Hypothesis B ignites as the new answer."

On screen: the SCOREBOARD —

FALSIFY : X knew by Jan 2021 (supplier email)      ← revised
RAG     : X knew by Mar 2021 (QA report)  [STALE]  ← still cites the refuted fact

[0:27–0:30] The close

"Same store, same query. FALSIFY revised its belief and the disbelief persists across sessions. Plain RAG can't. AI revised, not forgot."

On screen: open output/graph.html — the red refuted node, the missing orphan, the green frontier.

---

The single most important line

"It's not that the AI forgot where the context was — the context was wrong, and FALSIFY revised it."

That reframes the hackathon's "Where's My Context?" theme into FALSIFY's exact contribution.

---

If asked "why can't RAG do this?"

"Refutation is graph traversal over typed edges — this fact grounds that conclusion three hops away. A vector index has no edges to walk and no truth-state to filter on. This needs a knowledge graph — which is exactly what Cognee gives us."