12 minutes. Four surfaces. Optimised for "is this real, and is it differentiated?"
The one sentence this demo has to land:"Every number on every screen names its source,
and where we don't know something the system says so instead of guessing." Everything below is in service of that.
Skip entirely: map (502) · define-game (facets are too coarse to impress) ·
any live engine run (20 min 35 s measured). This audience will not wait and does not need the plumbing.
1 · 2 min
Open on the landing page
https://ukacq.com
Scroll to 02 — Data and stop there.
"3.2 million UK companies. 6.8 million employee records. 10.8 million ownership records." Both of the first two are verified against the live database.
Then 03 — Approach: read the four principles aloud. They are the pitch.
"Nobody buys a database. What we've built is the discipline on top of it — every claim traceable to a filing, and no opinion dressed up as a finding."
2 · 4 min
A real company card
https://company.ukacq.com → search BAE SYSTEMS MARINE
Search, land on the card. Let them read it for a beat before you narrate.
Walk: Identity → Headline → Ownership → Employee history.
Then point at the small grey text under a block. That is the whole demo:
"See that line? Every block says where its data came from. And where the history isn't traceable to a filed account, it says 'provenance unverified' — it doesn't dress an estimate up as a filing. Eight percent of our employee series meet that bar. We show you which eight percent."
This is your strongest moment. Admitting the 8% is more persuasive than
claiming 100%, because it proves the label means something. If they push — "so most of your data is unverified?" —
the answer is: "most of our employee time-series is inferred, and we mark it. The company identity, filings,
ownership and classification are registry-sourced. The point is you can tell them apart."
Click into the top game (Fire, Air & Water, 124/190) and show the per-criterion scorecard with rationale.
Point out the excluded criteria: "Three criteria are shown but deliberately not in the arithmetic. Moat is one. Somebody made that call and you can see they made it."
"This is 437 individual scored judgements, each with its reasoning attached. The engine orders the field; it doesn't tell you what to buy. That's the line we hold — evidencing, not opining."
If asked "has this been validated against a human expert?" — say no.
"The arithmetic reproduces the reference sheet exactly. The judgement has no human benchmark yet, and we
retracted a claim that it did when we found the reference was machine-generated. That's on the roadmap and
we're not going to pretend otherwise." This answer wins more trust than it costs.
4 · 2 min
Close on the engine comparison
https://llm-game.ukacq.com — do not click Run
The page opens on a completed benchmark (demo-aerospace-t23). Two reports side by side, judge verdict on top.
Give them the measured numbers and stop:
Our engine
A single LLM call
Time
20 min 35 s
2 min 49 s
Measured over
15 runs
5 runs
"Ours is seven times slower, and that's the product. A one-shot gives you a confident essay in three minutes. Ours spends twenty minutes doing the work you'd otherwise pay an analyst a week for — and shows you every step it took."
Honest caveat if they probe the reports: the engine's citations are currently
thin — live search isn't wired yet, so it reasons from model knowledge. Say it's the next build. Don't let
them discover it by scrolling.
Questions you will get, and the honest answer
Question
Answer
"Where does the data come from?"
Companies House bulk filings, PSC register, iXBRL accounts. All primary, all free. No purchased datasets.
"How current is it?"
The iXBRL bulk load is dated in the source registry — check the Sources block on a card before you claim a freshness number.
"Is any of this AI-generated?"
The company data is not — it's registry extraction. The research reports are LLM-generated and labelled as such, with a gate that blocks reports carrying too little evidence.
"Who else uses it?"
Three approved accounts exist. Do not imply a user base.
"Can I try it?"
Yes — needs an email added to the approved list. Takes a minute.
"What's the moat?"
Not the data — that's public. It's the provenance discipline and the settings layer: every assumption is a row you can inspect and change, rather than a constant buried in code.
Demo 1 · investor cut · Howard 2026-07-21 · timings measured from game.race_job (n=20)