complai / docs /walkthrough_script.md
aliceberthale's picture
Strip inline DISPOSITION/CONFIDENCE tags from displayed answer; update walkthrough confidence note
d1b7bb4
|
Raw History Blame Contribute Delete
4.53 kB
# CompLAI — 3-minute walkthrough script
**Format:** voiceover screencast · **Audience:** Mistral application reviewers
**Persona:** a regulatory & sustainability compliance analyst at a European bank
**Spine:** pure product flow (ask → cite → quantify → human-review gate → sovereignty)
> Target ~3:00 at a calm 150–160 wpm. VO below ≈ 440 words. Pause where marked `[…]`.
---
### Pre-flight (before you hit record)
- Open the live demo, **Chat** tab. Clear any prior conversation.
- Have these two questions ready to paste (both verified to behave as the script needs):
- **Q1 (auto-answers, cited):** *"A corporate client imports 50 tonnes of steel into the EU in 2026 — what are their CBAM obligations?"*
- **Q2 (routes to human):** *"Can a client use voluntary carbon offsets instead of surrendering CBAM certificates?"*
- CBAM Calculator tab pre-set to: Steel — Basic Oxygen Furnace, 50 tonnes, ETS €65.
- Browser zoomed so citations + confidence bar are clearly legible.
---
## Shot list
| Time | On screen | Voiceover |
|------|-----------|-----------|
| **0:00–0:18** | Title / Chat tab idle | "This is CompLAI — a regulatory compliance copilot for financial institutions. Picture a compliance analyst at a European bank: they report under CSRD, assess EU Taxonomy alignment, and price the CBAM carbon exposure of the corporate clients they lend to. Today, one question means searching thousands of pages of regulation." |
| **0:18–0:48** | Paste **Q1**, answer streams in | "So let's ask one. *[paste Q1]* The answer streams back grounded only in regulation we've ingested — the full CBAM text from EUR-Lex — not the model's memory." |
| **0:48–1:15** | Expand Sources; hover the `Article 9` / `Article 21` chips | "And for a bank this is the part that matters: every answer cites down to the specific article. *[hover chips]* If an analyst can't trace a statement to a source clause, it won't survive an audit — so traceability is the default, not an afterthought. Next to it: a disposition — here, *required* — and a calibrated confidence score." |
| **1:15–1:25** | Point at the confidence bar (~87%) | "That confidence score isn't the model's self-report — it's calibrated against how well the regulation corpus actually supports the question. Honest, not the inflated ninety-nine percent a raw model emits." |
| **1:25–1:55** | CBAM Calculator tab → Calculate | "For the numbers, the model hands off to deterministic code — it never invents a figure that lands in a risk model. *[click Calculate]* Fifty tonnes of basic-oxygen steel, at today's ETS price: roughly six thousand seven hundred euros of certificate exposure, traced to the official emission factor. A client's carbon liability, quantified in seconds." |
| **1:55–2:30** | Back to Chat; paste **Q2**; review banner appears | "But the most important feature in a regulated setting is knowing when *not* to trust the model. *[paste Q2]* When confidence is low or the disposition is high-risk, it doesn't answer with false authority — it routes to a human reviewer, and says why. *[show banner]* It drafts and triages; a person signs off." |
| **2:30–3:00** | Slow zoom on answer + citations; end card | "And all of this runs on Mistral models — which is the whole point for a bank. The same system deploys inside your own perimeter: weights run in your datacenter, no client data leaves the building. Managed API in development, self-hosted in production — one config change. Grounded, cited, calibrated, and sovereign. A compliance copilot a regulated institution can actually put in front of its analysts." |
---
### Production notes / risks to narrate around
- **Confidence now reads ~87% on in-scope answers** (calibration was refit on retrieval grounding). Still say the word *calibrated* at 1:15 — it's the differentiator vs a raw model's inflated 99%.
- **Pick the verified questions.** Q1 auto-answers (disposition *required*, ~87%, grounded); Q2 hits *prohibited* and routes. Other questions may not land the auto-vs-review contrast as cleanly.
- **Don't show** multi-turn memory, the fine-tune, or the benchmark table — those are interview material, not the 3-minute cut. One storyline.
- **The sovereignty close is the Mistral-specific beat.** It's what separates this from a generic RAG demo. Land it last so it's what they remember.
- **If you have 20 spare seconds**, the RAG-vs-Base tab is a strong optional insert after 1:55 — but only if you stay under 3:10.