18 minutes. For someone who wants to know what's in the dataset, how good it is, and what they could ask of it.
The thesis for this audience: the dataset is large and the coverage is honest
about itself. Do not oversell breadth — sell the fact that you can tell verified from inferred, per row.
What's actually in it
Table
Rows
What it is
psc_stage
10,774,730
Persons of significant control — the ownership graph
company_employees
6,848,125
Employee counts + financial headline per company
ch_universe
3,228,658
The company universe — identity, SIC, status, filings
fact_candidates
657,585
Extracted financial facts awaiting corroboration
company_edges
340,170
Parent/subsidiary relationships
fin_* (16 tables)
0
The financial grid — built, empty. Be upfront.
All counts are COUNT(*) run today, not n_live_tup estimates.
That distinction matters here — the autovacuum estimate on this box has been wrong by two orders of magnitude
on at least one table.
1 · 5 min
Coverage honesty — lead with this, don't bury it
https://company.ukacq.com → BAE SYSTEMS MARINE
Show the card, then the source line under the employee history block.
575,487 of 6,848,125 rows (8.4%) have emp_history_trusted = true. Those read
"Filed accounts". The other 91.6% read "provenance unverified".
Search 14887623 to show an untrusted one side by side with the BAE row.
"Most of the employee time-series is inferred rather than filed. We don't hide that behind an
average — the label changes per row. If you're building a screen on employee growth, you now know which 8% you can
stand on."
Volunteer the units defect before they find it. 551 of 8,365 companies with
>100 employees show revenue under £100k — a units inconsistency in the source. That's 6.6% of large companies.
It's filed with the data owner. Searching BAE SYSTEMS (OPERATIONS) shows one: 16,609 employees,
"£7,103" revenue.
2 · 4 min
The ownership graph — the underused asset
10.77M PSC records and 340,170 parent/subsidiary edges.
On a card, show the Ownership (PSC) and Ownership chain blocks.
This is the part that answers questions a company screen cannot: who controls this, what else do they control, has control changed.
Current limit: chain traversal is shallow — the immediate corporate owner
renders, multi-hop traversal is not built. The data supports it; the query doesn't exist yet. Don't imply
you can walk a full ownership tree today.
3 · 5 min
Classification — where the dataset is weakest
https://define-game.ukacq.com → tab 2, "Pick from the data"
Tick a sector and watch the live counter move across 3.2M companies. The plumbing is real.
Then say the quiet part:sector is roughly two-thirds "services". Ticking
industrials + tech can still leave you 800,000 companies. SIC codes are a 2007 vocabulary applied to a 2026 economy.
"This is the honest weak point. We can count 3.2 million companies instantly and we can't yet
slice them the way an analyst actually thinks. The fix is a ratified 7-axis taxonomy — sector, business model,
offering, supply-chain tier, value-chain function, and three ownership facets — replacing one conflated code."
Why it matters to this audience specifically: the taxonomy work is the
difference between "3.2M rows" and "a queryable universe". If they care about screening, this is the roadmap
item to discuss.
4 · 4 min
What the scoring layer holds
https://select-game.ukacq.com
437 scored judgements — 23 games × 19 criteria, each with written rationale.
Scoping triage: 56 games, tiered 25 / 8 / 23.
Click a game to show per-criterion evidence.
The caveat this audience will care most about: there is no human-authored
reference set anywhere in this system. The 23 benchmark reports say "Screened by: Claude" in their own
headers, and the reference scoring labelled with a person's name was a model one-shot they had shown as an
example. We retracted that in writing on 2026-07-21. So no quality number from this layer is
defensible yet — the arithmetic is verified, the judgement is unbenchmarked.
Questions this audience asks
Question
Honest answer
"Can I export it?"
Not verified — I have not tested an export path. Don't promise it.
"Is there an API?"
Internal JSON endpoints exist behind the auth wall. No documented public API.
"How fresh is it?"
Two iXBRL bulk loads registered (2026-06, 2026-07). Check the Sources block for the exact date before quoting one.
"What about non-UK companies?"
Not yet. Everything is keyed on Companies House number. A scheme-agnostic identity layer is designed and unbuilt — that's what unlocks EU entities.
"Financial history depth?"
Headline figures per company. The 16-table year-by-year grid is built and empty. Say built-not-populated.
"Can I get the raw data?"
Commercial question, not mine to answer.
Demo 3 · data cut · Howard 2026-07-21 · all row counts are COUNT(*) on vpsm today