The slide exists. Somewhere.
Eight entities in one buy-and-build group. Several file systems. Someone needs a slide on bioinformatics for a pitch tomorrow, and it could be anywhere.
That is a familiar state for a roll-up. Head office runs on SharePoint. One acquired firm lives in Google Drive, another in OneDrive. Each has its own CRM, its own lab system and its own idea of how to name a folder. The group owns the knowledge. A company brain can connect it. Finding it is another matter.
The usual response is an IT migration: choose a stack, move everyone and wait for the combined estate to become usable. We think connecting knowledge across a portfolio can come earlier.
So we built Tarnovell Bioscience Group, a fictional group with archives we made messy on purpose, and published the data and expected answers. You can try the same questions on any system, including ours.
Why the integration programme keeps getting longer
Buy-and-build makes this a recurring problem. Juniper Square's summary of PitchBook's Q3 2025 US PE Breakdown put add-ons at 74.1% of buyout transactions, against a five-year average of 72%. Every acquisition brings another set of systems.
Cross-selling needs a shared view of the client. Pricing needs a shared view of what the group charges. Proposals need a shared view of work already delivered.
In 2011, McKinsey estimated from its merger work that 50% to 60% of synergy initiatives were tied to IT. That is an older practitioner estimate, but the dependency is familiar to anyone who has worked through an integration.
The systems being replaced often work well for their local users. A lab may have validated processes behind its software. Staff know their own folders. Meanwhile, the next acquisition closes and adds another workstream.
An investor who backs roll-ups told me the question with all of them is how fast acquisitions pay back. Waiting for the migration to finish before anyone can use the group's combined knowledge makes that harder.
Meet Tarnovell
Tarnovell is a fictional contract research group assembled over three decades. The eight entities are head office and seven operating companies. The firms, 200 people, 70 clients, 130 projects and 20 compound series are all invented.
| Entity | Location | Specialism | Joined |
|---|---|---|---|
| Head office | Nottingham, UK | Group functions; Tarnovell brand from 2020 | Group formed 1996 |
| A · Vellacombe Discovery | Nottingham, UK | Medicinal chemistry; founding platform | 1996 |
| B · Brindlewythe Biosciences | Cambridge, UK | In vitro biology, including hERG and CYP screening | 2001 |
| C · Merrivoss Bioanalysis | Madison, US | Drug metabolism and bioanalysis | 2008 |
| D · Quenverell Computational | Edinburgh, UK | Bioinformatics and computational chemistry | 2014 |
| E · Westravelle Analytical | York, UK | Stability testing and analytical chemistry | 2019 |
| F · Durnesswick Laboratories | Durham, US | Formulation | 2024 |
| G · Osterquill Toxicology | Portland, US | GLP toxicology | 1 October 2026; held back |
Each firm kept its own shared drive, email domain, client names and lab records. A laboratory
accession number at Vellacombe looks like VA20240012. Brindlewythe uses
B-ACC-2024-00012; Merrivoss uses M24/BIO/00012. Those are local
accession formats. None of them identifies a project across the group.
The archive has 340 files, including legacy PowerPoint, Excel and Word documents, 15 image-only scanned reports, draft decks, typo-ridden copies and a leaver's old working folders. Older records retain their original branding and date conventions. The Tarnovell brand did not exist before 2020.
The fictional integration programme, "One Group, Many Practices", began in April 2024. By the archive's reference date of 8 October 2026, price harmonisation covered three of six integrated firms. The lab systems had an agreed export dictionary but retained local accession numbers. The SharePoint migration was still in progress.
One client, four names, one trap
Vellacombe worked with "Corvenlea Pharma". Merrivoss knew it as "Corvenlea Ltd". Westravelle bills "Corvenlea Pharmaceuticals plc". The held-back acquisition has "Corvenlea (US)". The group's records identify these as variants of the same client.
There is a lookalike too: "Corvenlith Therapeutics", a separate client at Quenverell and Durnesswick. Merge it with Corvenlea and the group's client history becomes wrong.
Head office wants the full Corvenlea history and the services it has never bought. The archive includes a client identity bridge and a service-gap matrix to establish which names belong together. Similar spelling alone is not enough.
Use the group's knowledge before moving the systems
With a company brain, anyone in the group can ask one question and get an answer drawn from every firm's systems, while each firm keeps working in the tools it already uses. The original files stay where they are, and answers link back to them.
That means a first use case can go live without waiting for the migration. Connecting each source still needs scoped credentials, a security review and a clear owner for access decisions.
The public Tarnovell exercise uses uploaded ZIP archives. It tests whether the right answers and client matches come back from exported files. It does not by itself prove live connections or permission enforcement.
See one client's history under every name
Ask about a client, person, project or compound and you get the group's full history, even where each firm named it differently. Matches rest on evidence, such as a shared identifier or an approved identity bridge, and ambiguous records go to review. The Corvenlea example tests that judgement.
Check the source in one click
Every answer links the analyst to the slide, workbook sheet or report section that supports it.
Get the current version first
For the pitch question, the approved 2025 deck should come ahead of an obsolete 2012 version or an unfinished draft. The brain tracks when each document changed and whether it was approved, because newness alone does not make a document authoritative.
Each person sees only what they may read
A report restricted to Westravelle's quality team stays restricted. A group-wide question is answered from the evidence that particular asker may read. Our permissions guide explains that boundary for people and agents.
Five questions with checkable answers
The dataset includes an answer key listing expected sources and distractors. It gives slide and page locations where available, sheet names for workbooks and section headings for Word documents. Keep that key outside the material you ingest, so the system has to find the evidence.
| Ask | The trap | Look for |
|---|---|---|
| Find a bioinformatics slide for tomorrow's pitch. | A typo copy, an unfinished draft and an outdated deck compete with the current version. | Slide 3 of Quenverell's 2025 deck and its PDF export, ranked ahead of distractors. |
| Find hERG and CYP screening on about 40 kinase compounds, under three weeks. | Records span two firms and 13 years; three near-miss projects resemble the right ones. | Four matching projects at Brindlewythe and Merrivoss, with proposals, results and reports. |
| Show Corvenlea's group history and services it has never bought. | Name variants are separate in the CRM export; Corvenlith is a different client. | The supported client history and service gaps, with Corvenlith excluded. |
| Repeat the screening question after adding Osterquill. | The new archive brings different formats and another Corvenlea name. | Two additional matching projects and evidence for the Corvenlea (US) identity link. |
| Have we done any cryo-EM work? | The archive contains none. | No matching cryo-EM work in this archive, plus two related projects using X-ray structural evidence. |
The last test matters as much as the successful searches. A system that invents a cryo-EM project is worse than an empty result. Its answer should also make the search boundary clear: absence from the connected, permitted archive is not proof that work never happened anywhere in the group.
For the pitch question, the final check is a click through to the source. A partner needs to inspect the slide before putting it in front of a client. A fluent answer with no link to the slide doesn't pass.
Then the next acquisition arrives
In the fictional timeline, Osterquill closed on 1 October 2026. We held its 34 files apart on purpose. Load the 306-file main archive first, then ask the screening question. The answer key defines four matching projects, two at each of Brindlewythe and Merrivoss.
Add Osterquill's archive and repeat the question. The expected set grows to six projects. Its records also supply the "Corvenlea (US)" variant, which should extend the existing client history without pulling in Corvenlith.
This tests whether a system can extend an existing understanding when another archive arrives. In production, the new firm's knowledge becomes searchable once its sources are approved, processed and checked. It need not wait for the full IT migration.
Change the order of integration
Tarnovell still has pricing, branding and lab-system decisions to make. A knowledge graph cannot do that operational work for the group.
It can make existing evidence usable while those decisions are worked through. As each source system is replaced, its connection to the brain is updated and tested, so people keep access to the group's history throughout the migration.
Where operating companies need separate knowledge boundaries, each can keep its own brain, and head office can have a group view with its own authorisation. That view should contain only evidence approved for it. Head office's need for a combined picture does not grant unrestricted access to every subsidiary's records.
Try it with the public data
The portco-mock-dataset repository contains the main archive as eight ZIPs, 306 files in total, and a separate 34-file Osterquill archive. The eight ZIPs are a packaging choice and don't map to the eight companies. Each ZIP is under 30 MB. The answer key sits in its own file.
Synthetic data deserves scrutiny. The datasheet explains the construction and its limits:
- Invented names were screened against web search. That reduces collisions; it is not an exhaustive company-registry check.
- Scientific values have a defined method. Some assay values come from models trained on public Therapeutics Data Commons benchmarks and applied to invented compounds. Other results come from parametric models.
- Document text is synthetic. Language models drafted content from a fixed world specification, with planted answers, pilot files and selected background content reviewed and corrected. The datasheet records the review coverage and known limitations.
- Consistency checks run in code. They check employment dates, client-name eras, branding and planted facts. A public example makes those choices inspectable.
The dataset tests retrieval against designed answers. It is not real client data, a scientific validation study or a substitute for testing permissions on your own systems.
If you run a roll-up, start with the practical question: how long would your team take to answer those five questions today?
We install a first company brain use case in two weeks. Choose 60x Cloud with EU or US data residency, or Your Cloud inside your own Azure, AWS or GCP cloud environment. Book a demo and bring one cross-company question that your current systems make hard to answer.
Sources
- 60x, portco-mock-dataset: README, datasheet, world specification, architecture example and answer key, checked 14 November 2026.
- PitchBook, Q3 2025 US PE Breakdown, October 2025, as summarised by Juniper Square.
- McKinsey, Understanding the strategic value of IT in M&A, January 2011.
- Huang et al., Therapeutics Data Commons, NeurIPS Datasets and Benchmarks, 2021. The dataset's datasheet identifies the benchmarks used.
Tarnovell Bioscience Group is fictional. The investor anecdote is anonymised. 60x installs company brains for investment firms and professional services teams.