Boring Demo Generator

Make hype survive daylight.
Give me a falsifiable claim, a real context, and a timer. I’ll spit out a boring, reproducible demo plan that doesn’t need narration or vibes.

Short version:

• Claim + Context = what you’re proving and where it breaks first.
• Environment = how honest you’re being about reality.
• Risk tolerance = guardrails + refusal behaviors.
• Time-box = demo size (forces decisions).
• Metrics + Constants = what you watch vs what you freeze.
• Artifacts = receipts you’ll publish so a stranger can rerun it.

What each field actually does

• Claim – The falsifiable sentence the whole plan orbits. Weak claim ⇒ weak plan. Write it like a bug you can close.
• Context – Constraints that make the claim non-magical (lighting, device, language, aisle width). If it’s not here, it’s “variable.”
• Environment
  • Lab/Sim → permissive, fast iteration, more “controls,” less credibility.
  • Staging/Prod Clone → realistic constraints, fewer excuses.
  • Live Floor (Guardrails On) → harsh light; adds explicit safety lines.
• Risk tolerance
  • Low → inserts “fail-safe / show refusal” language and spotter/kill-switch lines.
  • Medium → allows narrow ranges, retries with notes.
  • High → you’re signing for paperwork; plan says that out loud.
• Time-box – Caps scope. The generator builds a smallest-falsifier path that fits your minutes.
• Metrics – What gets logged live (latency, FN/FP, near-misses, confidence drift). Omit this and you’ll narrate instead of prove.
• Hold these constant – The “no fiddling” list (model hash, firmware, camera height…). If it wiggles here, result’s trash.
• Artifacts (checkboxes) – Appends receipt steps to the plan:
  • Signed at capture → filename + timestamp + operator, up front.
  • Config snapshot → model/version/settings/hash, reproducibility anchor.
  • Raw media + logs → audit trail, not vibes.
  • Hashes → immutability receipts.
  • One-page receipt → the shareable proof (links everything).

Pass / Fail (how the tool judges you)

• Pass: A skeptical stranger can watch the clip and match counts to the log without narration.
• Fail: Hidden knob twists, one-off “it worked once,” or any unlogged intervention.
• Borderline: If a surprise happens and you didn’t stop to note it, that’s a fail—you moved the goalposts.

Outputs & buttons

• Generate Plan → builds Setup → Procedure → Pass/Fail → Safety → Receipt Chain.
• Export Markdown / JSON → paste in changelog/PRs.
• Share link → URL encodes inputs so teammates can open with fields prefilled.

Presets (use these when you don’t want to think)

• Smoke test (kills hype fast)
  • Env: Lab/Sim · Risk: Low · Time-box: 15
  • Metrics: latency; failure counts; refusal behavior
  • Artifacts: Signed at capture, Config snapshot, One-pager
• Stakeholder demo (credible, safe)
  • Env: Staging/Prod Clone · Risk: Low · Time-box: 20–30
  • Metrics: p95 latency; TP/FP/FN; interventions
  • Artifacts: + Raw media + logs
• Precision benchmark (nerd fight ready)
  • Env: Staging · Risk: Medium · Time-box: 30
  • Metrics: PR curve deltas; confidence drift; throttle events
  • Artifacts: + Hashes for inputs/outputs; dataset slice manifest

Mental model

Freeze fewer things than you want, log more than you think, and make the result boring enough to survive a hostile room. If you need to narrate, you didn’t prove it.