docs/demo-script.md

90-Second Demo Script — Underdog Lab

Record with a warm Space. All scenario examples use real challenge or World Cup 2026 data.

Beat 1 — Hook (0–15s)

“Most AI football predictions are unsupported guesses. Underdog Lab separates language understanding from probability. A small local model extracts WHAT changed. Transparent math owns the numbers.”

Show the World Cup 2026 Live tab with the tournament snapshot card:

• 48 teams, 72 group fixtures
• Recorded results (as of June 12 cutoff)

Beat 2 — World Cup forecast (15–30s)

Select Group A. Show the table with Mexico and South Korea at 3 points. Scroll through fixture probabilities.

Point to the evidence card: “The fitted model beats the previous baseline across 8,265 no-lookahead test matches. The 360M language model never writes a probability.”

Beat 3 — Scenario extraction (30–50s)

Switch to Challenge tab. Select a hidden historical match.

Enter:

Canada's striker is confirmed out.

Click Translate scenario.

“The local model extracts: key_attacker_unavailable. Team: home. Severity 1.0, certainty 0.9. It never predicts a score — it only classifies the story.”

Point to the Before/After probability bars.

Beat 4 — Backend honesty (50–60s)

“When the model produces zero-weight output, the system flags it visibly as a deterministic fallback. No silent degradation.”

Show the backend label in the factors card.

Beat 5 — Stress test (60–72s)

Switch to Scenario Stress Test. Move severity from 25% to 100%.

“Every adjustment is bounded and versioned. You can inspect exactly why the probability changed.”

Beat 6 — Reveal and scoring (72–85s)

Return to Challenge. Commit a forecast. Reveal the result.

“You, the baseline, and the scenario-adjusted model are all scored with proper scoring rules. Lower log loss wins. A surprising outcome does not by itself invalidate a calibrated forecast.”

Beat 7 — Tiny Model Lab + close (82–90s)

Show Tiny Model Lab tab.

“The QLoRA adapter failed its ship gate, so I rejected it. The safer base model ships locally with visible fallback. Every probability remains auditable.”

Recording notes

• Use a warm Space (pre-loaded model, pre-cached first scenario).
• Record at 1080p. Zoom browser to 120% for readability.
• No voiceover over slider movement — let the visual speak.
• Show the backend label clearly when mentioning honesty.
• Keep the Space URL visible in the browser bar for the first and last 5 seconds.