ukacq — Full Demo Guide

Howard · 2026-07-21 (refreshed) · ~35 min full flow · ~12 min highlights (steps 1,2,4,5,7)

Demoing to Nick Ashford (Fordhouse)? His interest is aerospace & defence, he supplies the Airbus QSPL data privately, and the standing rule on his work is "we are evidencing, not opining — that is Nick's job." Read the Nick-specific notes at the foot of this page before using the generic flow — several steps change, and one thing he would most expect to see is currently down.

Login: one is enough. The session cookie is scoped to .ukacq.com with a 30-day expiry, so logging in once on any page authenticates you on every subdomain. Verified against auth.py on the live box. Three approved emails: adam@fantell.uk, cto@mathtec.co.uk, adam.rudd@edinburghinvest.com. Screen-share viewers need no login.

The single most important timing fact: a live engine run takes 20 min 35 s (measured median, n=15). Do not start one during the demo. Show the completed benchmark instead — step 7 tells you exactly which one.

Flow

Landing page — the pitch live
Log in — once, covers everything
The Fountainhead — app directory live
Company Cards — 3.2M companies live
Select Game — the 19-criterion matrix live
Define Game — three input modes building
LLM Game — engine vs one-shot building
Settings — the governance reveal

Step durations below are presenter time — my estimate of how long to spend talking. They are not measured. Where a system duration is measured, it is labelled and sourced.

The Flow

Step 1

Landing page — the pitch

https://ukacq.com · public, no login

"ukacq — acquisition intelligence built on evidence, not opinion. Everything behind this page is live, fed from primary-source data, and every behaviour is a setting somebody decided, not a default somebody buried."

Scroll slowly through the three numbered sections.
01 Platform — 12 app tiles. Three are live and clickable; the rest are dimmed and marked Coming soon, and are not clickable (fixed today — they used to hang).
02 Data — pause here. "3.2 million UK companies, 6.8 million employee records, 10.8 million PSC ownership records." verified against live counts today.
03 Approach — "Evidencing, not opining. Code before tokens. Primary sources. Traceable claims."

What "835 dimensions" means: I have not verified this figure or the 10.8M PSC count against the live box — only the 3.2M and 6.8M. If asked to substantiate, use the two I checked.

Step 2

Log in — once

https://company.ukacq.com

Enter your email → a 4-digit code arrives from noreply@aubrains.com (code expires in 15 min).
Enter the code. You are now authenticated across all ukacq subdomains.

"One login, no password, whole platform. The session lasts 30 days — I won't have to do this again during the demo."

Do it before the audience joins if you'd rather not wait on an email. The session persists, so you can log in an hour early and the demo starts already inside.

Step 3

The Fountainhead — the app directory

https://apps.ukacq.com

Point out the status dots: live building planned
Scroll to Digital twin cards — 10 card types, each with its own subdomain. One of the ten is built (company); say so.
Click a live tile to move on — the tiles are links.

"This directory isn't a static page — it renders from its own settings. Which apps show, how they group, whether planned ones are visible: all rows in a table, not lines in code."

Step 4

Company Cards — the data showcase

https://company.ukacq.com

Use these searches. I tested every one against the live database today.

Type this	What you get	Why it's safe
`BAE SYSTEMS MARINE`	12,428 employees · £2.77bn revenue · PSC	trusted history, sane figures
`GKN AEROSPACE`	3,456 employees · PSC chain	trusted history
`00229770`	same company, by CH number	shows number-search works

Two searches that will embarrass you — I tried them:

Rolls Royce → zero results. The registry stores it hyphenated. Use ROLLS-ROYCE if you want it.
BAE SYSTEMS (OPERATIONS) → 16,609 employees and revenue of £7,103. That is a units bug, not a real figure (see below).
Precision Castparts → zero results. Not in the UK registry under that name.

Walk the blocks in order: Identity → Headline → Ownership (PSC) → Employee history → Classification → Sources.
The point to land: every block names its source at the bottom. "Every fact on this card says where it came from. That's the whole product philosophy in one line of small print."

Honest provenance — fixed today. Employee history now reads "provenance unverified" where the underlying series isn't traceable to a filing, and "Filed accounts" only where it is. Only 575,487 of 6,848,125 rows (8.4%) qualify as trusted. Before today the card claimed "Filed accounts" for all of them. If someone asks how you know the data is honest, this is the answer — the system tells you when it doesn't know.

Known defect, be ready for it. 551 of 8,365 companies with more than 100 employees show a revenue under £100k — a units inconsistency in the source (some rows in pounds, some in millions). That is 6.6% of large companies. If you hit one, name it as a known data-quality defect being tracked, not a display bug. It is filed with Pixel, who owns that lane.

Step 5

Select Game — the criterion matrix

https://select-game.ukacq.com

This is the strongest surface you have. It needs no explanation — the data argues for itself. If you only have ten minutes, this and the company card are the demo.

"23 acquisition games, scored on 19 criteria of equal weight. No weighting tricks, no composite index, no opinion. The engine orders them; a human decides."

The ranking table loads immediately. 23 games sorted by total out of 190. Top: Fire, Air & Water at 124.
Click any row → that game's full scorecard, every criterion with its evidenced rationale.
Tab: Scoping triage — 56 games with tier badges. Tier 1 = 25, Tier 2 = 8, Tier 3 = 23.
Tab: The 19 criteria — what is actually being measured.
Land this point: "Three criteria are deliberately excluded from the score — moat is one. They're shown, they're just not in the arithmetic. That was a decision, and it's visible."

If asked "is this ranking validated?" — the honest answer is no. The reference set it was checked against turned out to be Claude output, not human scoring. We retracted that claim in writing on 07-21. The arithmetic reproduces Adam's spreadsheet exactly (437 scores, top total 124 — verified). The judgement has no human benchmark yet. Say that plainly; it lands better than a hedge.

Step 6

Define Game — three ways in

https://define-game.ukacq.com

Tab 1 — Write it. Free text. Name + description.
Tab 2 — Pick from the data. Do this one live. Tick a sector, then a SIC division, and narrate the company counter falling. It is querying 3.2M companies as you click.
Tab 3 — Upload. Drag a pitch deck in (pdf/pptx/docx/xlsx/csv). It extracts text and infers parameters.
Tab: Games defined — 20 exist today.

"Define a thesis three ways. And when it can't infer a parameter from what you gave it, it says so rather than picking a default — that refusal is the feature."

Weakness to own before someone finds it: the facets are crude. sector is about two-thirds "services", so ticking two boxes can still leave you 800,000 companies. The right axes are Solomon's eight structural parameters, not yet wired. Frame the live counter as the proof the plumbing works — the taxonomy behind it is the next job.

Step 7

LLM Game — the engine versus a one-shot

https://llm-game.ukacq.com

Do not click Run. Measured engine median is 20 min 35 s (n=15 complete runs; min 13 m 52 s, max 35 m 32 s). The page will sit there. Show the finished benchmark instead.

The number that makes the case — all measured, from race_job, n=20, every job completed:

	Engine	One-shot
Median wall time	20 m 35 s	2 m 49 s
Range	13 m 52 s – 35 m 38 s	2 m 31 s – 3 m 25 s
Runs measured	15	5

The engine takes 7.3× longer. That is the honest framing: it is not faster, it is more thorough — and step 7's job is to show what you buy with the extra 18 minutes.

The page loads the most recent benchmark. That is bench 2, demo-aerospace-t23 — engine, one-shot and judge all completed, all three bodies populated (engine 14,470 chars · one-shot 24,083 · judge 5,039).
Walk the two reports side by side. Then read the judge verdict at the top.
Right panel — expected vs elapsed per step. The expectations are real medians computed from past runs, not seeded guesses (the view is percentile_cont(0.5) over observed rows — worth saying if the audience is technical).
Gear icon → settings drawer: model per pipeline step, track toggles, HITL points, judge config.

Check this before you present: the other bench row is __perm_test__ with every field NULL. If the page loads the latest by bench_id or started_at you get the good one — bench 2 is newest on both. If it loads unordered, it could render empty. Open the page once before the demo and confirm you see content.

The weakness, if pressed: the engine's reports currently carry very few real citations — search was never wired, so it runs from model memory. The evidence gate correctly flags this. The machinery is real; the evidence layer is the next build. Better to say this than to let someone notice the bibliography is thin.

Step 8

Settings — the governance reveal

https://apps.ukacq.com/define-game

"Every setting carries four separate questions: is it On? Is it Shown to the user? Can they Change it? Can they Re-scope its bounds? The admin controls not just the value — but who may see it and who may move it."

Show the settings grouped by category (A. Input, B. Facets, C. Pipeline …).
Three statuses: ratified decided · proposed suggested, unconfirmed · open genuinely unanswered.
The moment that lands: find a setting with status open and read its written question aloud. "This one is deliberately blank. The system is carrying the question instead of a guess. Nineteen settings are in that state right now, each with a question waiting for a decision."
Then show one where the user cannot change it. "Visible, locked. They get a 403 if they try. That's not a bug — that's governance."

Why this is the differentiator. Most platforms hardcode their assumptions and bury them. This one hoists them into rows you can see, audit and change — and when it doesn't know the answer it says open rather than inventing a default. There are 83 settings across three apps today.

Use demo.ukacq.com/settings to demonstrate the four-axis model itself — flip to Admin view and show a setting being hidden, locked, or bounded.

Readiness — what is true, per surface

No percentage scores here. There is no rubric that would make "78% ready" mean anything, so a number would be invented precision. Instead: what works, what doesn't, what to avoid.

Surface	Works	Doesn't	Demo?
ukacq.com live	Loads fast, responsive, self-hosted font. 3.2M/6.8M stats verified.	"835 dimensions" and 10.8M PSC unverified by me.	Yes — open here
apps.ukacq.com live	Settings-driven render, status dots, card directory.	Settings icon points at `app.ukacq.com`, which does not exist.	Yes
company.ukacq.com live	3.2M searchable. 10 blocks. Honest provenance labels (fixed today).	6.6% of large companies show wrong revenue units. Only 8.4% of employee series are trusted.	Yes — use the tested searches
select-game.ukacq.com live	437 scores, reproduces the spreadsheet exactly. Click-through scorecards.	No human-validated benchmark behind the judgement (retracted 07-21).	Yes — strongest piece
define-game.ukacq.com building	Three input modes. Live facet counter over 3.2M. Upload + extraction.	Facets too coarse to be useful. 19 of 54 settings unanswered. gd0/gd7 still `todo`.	Yes — tab 2 only
llm-game.ukacq.com building	One complete benchmark with all three legs. Real median-based step forecasts.	20-min runs. Thin citations. 9 of 23 steps record no timing at all.	Yes — never click Run
map.ukacq.com holding page	Nothing. No service on :8001, no systemd unit.	Everything.	No — skip entirely

What to fix — ranked by demo exposure

	Item	State
done	9 dead app links on the landing page → dimmed, non-clickable, "Coming soon"	deployed today
done	Employee provenance label — "Filed accounts" only when actually filed	deployed + verified today
done	`map` status corrected to "planned"; its false `port=8001` assertion nulled	done
done	llm-game asserted a retracted claim in a live API field — `"gold_standard": "…hand-scored"` for `run_id=0`. Corrected to `reference_set` + `reference_is_human: false`	deployed today
done	`app_registry` was not a faithful index — `llm-game` and `race-game` were deployed and serving but unregistered; `assist-board` missing. Fountainhead renders from this table, so the front door misrepresented the estate	deployed today
done	Dead `app.ukacq.com` links → `apps.ukacq.com` (0 references remain)	deployed today
done	llm-game bench loading verified — query filters `engine_report IS NOT NULL ORDER BY bench_id DESC`, so the empty `__perm_test__` row cannot be selected	verified today
P1	Revenue units — 551 large companies show impossible figures	filed to Pixel with measurement (his lane; a display patch would hide it)
P2	Add attendee emails to `beast.ukacq_users` if they want their own login	needs your say-so — I won't add accounts unasked
P3	9 of 23 pipeline steps never write `wall_sec` — forecast panel is blind on them	after the demo
open	`select-game`'s registry purpose promises four ranking modes, interactive filtering and saved shortlists. The app has five GET routes and zero writes. Either the implementation is scoped to games where the purpose describes companies, or the purpose is wrong	needs your ruling

What to enrich next — and why, in order

#	Area	Today	Next
1	Engine citations	Reports run from model memory; near-zero real sources	Wire search. This is the single biggest quality gap — it is what turns a plausible report into an evidenced one, and the whole pitch is "evidencing, not opining".
2	Entity identity	CH number is the only key	`entity_registry` + `entity_identifier`. Unlocks non-UK entities, LEI, FIGI. Three other agents are blocked behind this one table.
3	Financial grid	16 `fin_` tables built, all empty*	Populate from CH statutory columns. Fixes the revenue-units defect at source rather than patching the display.
4	Employee provenance	Only 8.4% traceable to a filing	Either raise the trusted share or stop carrying untraceable series. Right now the card is honest about a weak dataset — better is a stronger dataset.
5	Taxonomy	SIC codes only — hence the coarse facets	Adopt Librarian's ratified 7-axis taxonomy. Directly fixes define-game tab 2.
6	Digital twins	1 of 10 built	Game card next — it has the most data behind it already.
7	A human benchmark	None exists	Without it no quality claim from either instrument is defensible. Needs a person to score games by hand.

Settings — showing consequence without drowning people

Three beats, in this order

Consequence first, mechanism never yet. "That's five years of financials." Don't open a settings panel.
Then the control. Change it to three. The card re-renders. One setting, one visible effect. This is the whole idea in ten seconds.
Then the governance. "I can also hide this from users, or show it locked, or let them move it only between 1 and 10." That is the part competitors don't have.

User settings vs admin settings — the clean split

User = your view. Density, which blocks show, how many years, light/dark. Preferences. Reversible. No blast radius beyond the screen.
Admin = what the system does. Enrichment depth, pipeline concurrency, embargo rules, evidence thresholds, which model runs which step. These have consequences that outlive the session.
The four axes are what connects them: On · Shown · Change · Scope. A setting can be on but hidden; visible but locked; changeable but bounded.

Do not show the full 54-setting define-game schema in a first demo. It reads as complexity rather than control. Show three settings and the four axes. Send the technical audience to demo 2.

Quick reference

Surface	URL	Login	State
Landing	ukacq.com	public	live
Fountainhead	apps.ukacq.com	yes	live
Company Cards	company.ukacq.com	yes	live
Select Game	select-game.ukacq.com	yes	live
Define Game	define-game.ukacq.com	yes	building
LLM Game	llm-game.ukacq.com	yes	building
Race Game	race-game.ukacq.com	yes	not linked publicly
Map	map.ukacq.com	—	502 — skip

Demoing to Nick Ashford (Fordhouse)

Who he is, from our own records: Fordhouse. Aerospace & defence. He supplies the Airbus QSPL ExSitu export privately — 18,218 rows / 1,065 suppliers / 40 countries, and our own registry says "Nobody else has this. It arrives privately from Fordhouse/Nick, not from a public register." The 13-bucket civil/defence classifier was built for him. He has an FH-1 archetype card whose weights are recorded as "Seth's best-shot, NOT Nick-validated" and which awaits his reaction. And the standing correction on work aimed at him is Adam's: "we are evidencing, not opining — that is Nick's job."

🔴 The thing he will most expect to see is down. map.ukacq.com — UK aerospace mapping — returns 502. Nothing is listening on its port and there is no service unit. For an aerospace counterparty this is the single most conspicuous gap in the estate. Either stand it up before he sees it, or open by naming it: "the map isn't running on this box yet — it's the next thing I want to show you." Do not let him click it.

What we can genuinely show him — verified today

SIC	What	Companies
`33160`	Repair & maintenance of aircraft	1,431
`51210`	Freight air transport	573
`30300`	Manufacture of air & spacecraft	459
`25400`	Weapons & ammunition	52
`51220`	Passenger air transport	33
`30400`	Military fighting vehicles	6

That is a real aerospace & defence universe — roughly 2,550 companies — not a generic slice. The top of it by headcount is exactly the list he'd expect: Airbus Operations (7,842), GKN Aerospace (3,456), AgustaWestland (2,876), Airbus Defence & Space (2,199), Safran Seats, Safran Landing Systems, Goodrich Actuation, GE Aircraft Engine Services.

🔴 But do not open a card on the Headline block without reading this first. The monetary figures are stored at an inconsistent scale, and it bites hardest on exactly the companies Nick knows cold:

Company	Staff	Card shows	Reality
AIRBUS OPERATIONS	7,842	£69K	Airbus UK is ~£3bn
GKN AEROSPACE	3,456	£709K	~£710M
AIRBUS DEFENCE & SPACE	2,199	£532K	~£532M
BAE SYSTEMS (OPERATIONS)	16,609	£7K	wrong at any scale

BAE Systems (Operations) is the top result by employee count in aerospace, so this is the first thing he'd hit. Mitigation deployed today: the Headline block now carries "monetary scale unverified" in its source line, so the card no longer asserts a figure it cannot stand behind. That turns a wrong number into a declared unknown — but it does not make the number right. Root cause is filed with the data owner.

Suggested flow for Nick — 20 min

N1 · 3 min

Open on the aerospace universe, not the landing page

https://company.ukacq.com → search AIRBUS OPERATIONS

"This is 3.2 million UK companies, and inside it about 2,550 in aerospace and defence — 459 manufacturers, 1,431 in MRO. Let's start with one you know."

Show Identity, Ownership (PSC → Airbus SE), Classification. These are registry-sourced and correct.
Point at "monetary scale unverified" yourself before he squints at the revenue. "That line is the product. We know the scale on these financial figures is inconsistent in the source, so we say so rather than print a confident wrong number."

This is the strongest possible move with Nick specifically. He is the person who sends us data privately; his question will be whether we handle data honestly. Volunteering a known defect on a company he knows is worth more than a clean demo on a company he doesn't.

N2 · 5 min

Ownership chains — where the aerospace estate gets interesting

10.77M PSC records, 340,170 parent/subsidiary edges. Show Airbus Operations → Airbus SE, then GKN Aerospace → Westland Group.
This answers the question a supplier list cannot: who actually controls this entity, and what else do they control.

Be straight that multi-hop traversal isn't built — the immediate corporate owner renders, the full tree doesn't. The data supports it; the query doesn't exist yet.

N3 · 6 min

Selection matrix — evidencing, not opining

https://select-game.ukacq.com

23 games, 19 equally-weighted criteria, 437 scored judgements each with written rationale.
Show the excluded criteria: shown, but deliberately not in the arithmetic.

"The engine orders the field. It does not tell you what to buy — that's your job. Every score carries its reasoning so you can disagree with a specific one rather than the whole ranking."

This phrasing is deliberate — it is Adam's own correction applied ("we are evidencing not opining; that is Nick's job"). It also pre-empts the objection that a machine is making investment calls.

If he asks whether the scoring is validated: no. The reference set turned out to be model-generated and we retracted the claim in writing today. Say it plainly.

N4 · 6 min

The engine vs a one-shot

https://llm-game.ukacq.com — never click Run

Loads a completed benchmark. Engine 20 m 35 s vs one-shot 2 m 49 s, measured.
The point for Nick: the extra 18 minutes is the difference between an essay and a worked analysis.

Volunteer that citations are currently thin — search isn't wired, so it reasons from model knowledge. For someone whose job is diligence, being told the evidence layer is incomplete is far better than him finding a thin bibliography himself.

Two more assets worth showing — verified

N5 · 3 min

The A&D supply-chain tier taxonomy

employee_data.priority_sic — 44 SIC codes across 6 tiers, built for aerospace & defence rather than generic sector coding:

Tier	Group	SIC codes
1	AD-core (25400, 30300, 30400)	3
2	precision-metal	9
3	castings	10
4	electronics	10
5	mechanical	5
6	MRO-services	7

"This isn't a generic sector filter — it's a supply-chain tiering. Tier 1 is the primes, tier 2 precision metal, tier 3 castings, down to MRO. It's the shape you'd actually use to walk a supplier base."

Why this matters for Nick specifically: the generic sector column is useless here — it has only four values org-wide and every aerospace company collapses into "industrials" or "services". The tier taxonomy is the thing that makes the dataset aerospace-literate.

N6 · 4 min

The A&D thesis scorecard — ranked #2 of 23

There is a 4,974-word Aerospace & Defence Tier 2/3 Suppliers benchmark with a paired 19-criterion scorecard. It ranks #2 of 23 games. Strong on barriers to entry (9), mission-criticality (9), revenue quality (8), margins (8); weak on controllable forces (2) and competition (4).

The most credible detail in it: a revision note recording that the first pass undercounted competitors roughly 3× (6 → 12 consolidators) and that the "UK white space" hypothesis did not survive the second pass. Leave that in. Self-correction is the credibility.

Two things to fix before it goes near him:

Its header reads "Prepared for Tom Mather · Screened by: Claude (research assistant)". Rewrite it.
Its scorecard is attributed Tom Mather (via Adam) — that is the attribution we retracted today. Do not present the scores as human-authored.

🔴 Do NOT search "Fordhouse" in front of him. Company 08009227 renders as SITG LIMITED in one table and FORDHOUSE EQUITY LIMITED in another — same company number, two names, one table is stale and I don't know which. Three further Fordhouse rows exist in company_employees but not ch_universe. His own group is the worst possible place for a name-drift bug to surface.

🔴 Do NOT show the raw demo-aerospace-t23 engine report. It contains a paragraph reporting a prompt-injection attempt against the pipeline, and it self-labels "DRAFT — not yet board-committable". Both are fine internally; neither should go to a counterparty unreviewed.

On his QSPL data — we do not hold it, and I nearly told you we did. QSPL is the Airbus Qualified Supplier Process List. A search of the box surfaced a company literally called "Quality Springs & Pressings Limited" whose initials match, and that is a coincidence, not his data. There is no Airbus supplier, certification, NADCAP or AS9100 data on this box at all. Do not imply otherwise.

What he'd ask for that we do not have

He will expect	State	What to say
The aerospace map	502	Now serves a holding page, not a 502. The working map is Atlas's at `map.aubrains.com` — separate login, your ukacq session does not carry across.
His Airbus QSPL data	not on vpsm	Confirmed absent. No Airbus supplier, certification, NADCAP or AS9100 data anywhere on this box.
FH-1 archetype card	not built here	Archetype cards are 1 of 10 twin types, none built on vpsm. His FH-1 weights are recorded as not-yet-validated by him — that's an ask, not a deliverable.
Civil vs defence split	not on vpsm	The 13-bucket classifier was built for him but lives in the mapping estate, not here.
Certifications (AS9100, NADCAP)	absent	Zero certification tables. Confirmed by full schema scan.

The honest framing for the whole Nick session: what is strong here is the company universe, the ownership graph and the provenance discipline. What is weak is everything aerospace-specific — the map, his own supplied data, the archetype card and the civil/defence split all live elsewhere or aren't built. If you want the aerospace story to land, the map is the thing to stand up first.

Howard · 2026-07-21 · Every figure here was measured on vpsm today or is marked unverified. Timings from game.race_job (n=20) and game.step_run (n=15 complete runs) — two independent measures agreeing to 0.2%.
Presenter-time estimates are my judgement, not measurements, and are labelled as such.