Drill — Pick the Right R
Naming the 7 R’s takes an afternoon. Choosing the right one, for a system you have known about for four minutes, in a room full of people who already have an opinion — that is a different skill, and it is the one this drill exercises. Below are ten systems from ten different fictional organizations. Every one of them has an obvious answer, and in every one of them a single buried fact makes the obvious answer wrong: a clause in a licence, a line in an audit log, a cost that only starts existing after you move, a regulator, a support matrix, somebody’s retirement date. You get three minutes per card — one R and one sentence, then move on, no going back. Then you score yourself against a key that gives the right R, the defensible second-best R, and the exact fact that would make the second one correct. The whole thing takes about fifty minutes and costs nothing.
Imagine ten boxes in a hallway, and for each one you have to say straight away: bin it, leave it, carry it, buy a new one, or rebuild it. Easy — until you pick a box up and find a note taped underneath. “This one belongs to Grandma.” “This one only opens with a key that Uncle Jim is taking with him when he moves out in June.” “This one is free to carry, but costs a pound every single time you open it in the new house.” The note is always there, on every box, and it is never on the outside where you can see it. This drill is ten boxes with ten notes, and the game is to find the note before you decide.
You need a timer and something to write on. There is no cloud account, no credit card, and no infrastructure anywhere in this drill; disposition is a decision, and decisions are made on paper long before anything moves. Nothing here carries over from the capstone either — these are ten unrelated organizations, and Brambleside Veterinary Group does not appear on this page at all. That is deliberate. The capstone teaches you to hold one estate in your head across five parts; this drill teaches you to walk into an estate you have never seen and be useful in three minutes, which is what an assessment week actually feels like. If you want the continuity version, it lives at Move Brambleside — Start Here and its Part 1.
What this drill assumes, and what it produces
☺ Like you’re 10: You already know the seven choices. This is practice at choosing fast — and then finding out why you were wrong.
What it assumes: that you can already define all seven R’s without looking them up. If you cannot — or you would like a refresher before the clock starts — read The 7 R’s first, or use the one-line reference table below. Nothing else is assumed. You do not need the capstone, a cloud account, a discovery tool, or any other page in this labs track.
What it produces: one artifact — a ten-row disposition scorecard in which every row carries an R, a one-sentence rationale, the fact you believe drove it, and a confidence score. It is deliberately the same shape as the matrix you build in Capstone Part 1, minus the columns that only make sense inside a single estate. After you score it against the key, the second and more valuable artifact is three lines of your own: the rule you would apply to the next unfamiliar system, written by you, in your words.
Theory this drill is built on — you do not need to read these first, but they are where the answers come from:
| Read | For |
|---|---|
| The 7 R’s (primary) | What each strategy means, the five decision criteria, and the quick decision path this drill is timing you against |
| What & Why We Move | TCO and the business case — several cards turn on a cost that does not exist until after the move |
| Case Studies | What these decisions look like when real organizations got them right, and when they didn’t |
None of these ten companies exist. Every currency figure, data volume, utilisation percentage and licence cost was made up to make the exercise bite, and the numbers are internally consistent with each other and with nothing else in the world. Real cloud pricing changes monthly and varies by region, commitment and negotiation, so no figure on this page should ever be quoted at anyone or carried into a real business case. Carry the method instead, which is the thing being practised: find the fact, do the arithmetic, write the sentence.
The seven R’s, one line each
☺ Like you’re 10: The seven choices, on one card, so you never have to leave this page mid-drill.
Keep this table where you can see it. Every card below gets exactly one of these seven — not two, not “Rehost now and Refactor later,” not “it depends.” The discipline of committing to one is most of the drill: a disposition matrix full of hedged rows is a matrix nobody can plan a wave from.
| R | In one line | The question it answers |
|---|---|---|
| Retire | Switch it off for good | Do we still need this at all? |
| Retain | Deliberately leave it where it is, for now | Is there a hard blocker, or no case to move yet? |
| Rehost | Lift-and-shift — move it unchanged | Does it work fine, and do we mostly need to be out of here? |
| Relocate | Move a whole virtualized platform at once | Is this a big block of VMs that should travel together, untouched? |
| Repurchase | Drop-and-shop — buy a ready-made replacement | Does an off-the-shelf product already do this better? |
| Replatform | Lift-tinker-shift — small tweaks on the way over | Would one or two contained changes unlock a big win? |
| Refactor | Rebuild it properly for the cloud | Is this critical and actively held back by its own design? |
How this drill works
☺ Like you’re 10: Three minutes a box. Write your answer in pen. Don’t peek, don’t go back, don’t change your mind.
Read one card. Start a three-minute timer. Write one R and one sentence saying why. When the timer goes, move to the next card whether you are happy or not — and do not go back to change an earlier answer, even when card 6 teaches you something that makes card 2 look wrong. Going back destroys the only thing this drill measures, which is what you actually do under time pressure with incomplete information, not what you can reason your way to with an hour and a second pass.
Three minutes is not arbitrary. It is roughly what a system gets in a real portfolio-assessment workshop, where forty applications are walked in a day and the room is a mix of people who own the system, people who fund it, and people who have never heard of it. You will not have complete information there either. The skill is not certainty — it is producing a defensible answer plus an explicit note of the one thing you would need to check to be sure.
Write a third thing on every row, and write it before the timer goes: the fact you believe drove your answer. Not the reasoning — the fact. “Vendor sells a hosted edition.” “Twelve users.” “Lease ends in five months.” When you score yourself, compare that column first, before you compare your R’s. Getting the right R off the wrong fact is worse than getting the wrong R off the right one, because the first is luck and the second is a correctable analyst. On a real programme, the first one gets you a matrix that falls apart the moment anybody asks a follow-up question.
Your scorecard
Copy this into a spreadsheet or a text file before you start. Ten rows, four columns you fill in on the clock, one you fill in afterwards:
card,organization,system,my R,one sentence why,the fact that drove it,confidence 1-5,scored
1,Hallowfield Housing Trust,TenantTrack,,,,,
2,Corvid Analytics,ParquetLake,,,,,
3,Meridian Diagnostics,PathologyPortal,,,,,
4,Ravensgate Borough Council,PermitDesk,,,,,
5,Calderhorn Freight,the Yard estate,,,,,
6,Sable and Finch,TimeLedger,,,,,
7,Puffin and Pike Books,Checkout,,,,,
8,Grimsby Marine Mutual,HullCalc,,,,,
9,Wexley Foods,ColdChain-Edge,,,,,
10,Larkspur Foundation,GrantHub,,,,,The confidence column earns its place at scoring time. A wrong answer you were sure about is a far more interesting result than a wrong answer you flagged as a guess — the first is a belief to correct, the second is just a gap in the brief.
The ten systems at a glance
☺ Like you’re 10: Here are all ten boxes, listed plainly — before you read the notes taped underneath them.
This table is the neutral shape of each system: what it is, what it runs on, how big it is, who touches it. Nothing in it will decide a single card. The deciding facts are in the ten write-ups below, buried in ordinary-sounding prose exactly where they sit in real life — in a paragraph somebody skimmed.
| # | Organization | System | What it does | Runs on | Data | Who uses it |
|---|---|---|---|---|---|---|
| 1 | Hallowfield Housing Trust 4,100 rented homes | TenantTrack | Tenancy records, rent accounts, repairs history | 3 VMs, Windows Server 2019, SQL Server 2017 Standard | 180 GB | 12 named licences; “business-critical” per the Head of Housing |
| 2 | Corvid Analytics 90-person market-data firm | ParquetLake | Analytics lake, plus the Spark cluster that queries it | 8 storage nodes, 12 compute nodes, own racks | 340 TB | 40 internal analysts and 41 partner organizations |
| 3 | Meridian Diagnostics private pathology lab | PathologyPortal | Vendor results portal — clinicians download reports | 2 VMs, PostgreSQL 13 | 600 GB | Clinicians at 300 GP practices |
| 4 | Ravensgate Borough Council planning department | PermitDesk | Planning-permit records, 2007–2022 | 1 VM, Classic ASP, Windows Server 2012 R2, SQL Server 2008 R2 | 240 GB | ~30 read-only lookups a month, by clerks |
| 5 | Calderhorn Freight 11-depot logistics operator | the Yard estate | Everything — ERP, file, print, line-of-business, dev | 420 VMs on 26 vSphere hosts, 3 racks | ~90 TB | All 1,200 staff; platform team of 4 |
| 6 | Sable & Finch 180-lawyer firm | TimeLedger | Vendor time-recording and client billing | 2 VMs, Oracle Database SE | 400 GB | Every fee-earner, daily. In support, well liked |
| 7 | Puffin & Pike Books online bookseller | Checkout | Checkout and order service | 6 VMs behind a hardware load balancer, MySQL 5.7 | 220 GB | Every customer; 3-person platform team |
| 8 | Grimsby Marine Mutual small marine insurer | HullCalc | Actuarial pricing tool | 1 Windows VM — Excel with VBA, plus an Access database | 4 GB | One person — the chief underwriter |
| 9 | Wexley Foods chilled-food distributor | ColdChain-Edge | Freezer telemetry and out-of-range alarms | 40 small on-prem servers, one per depot | ~1 TB total | Depot staff at 40 sites; unattended overnight |
| 10 | Larkspur Foundation ~600 grants a year, 60 staff | GrantHub | Grant applications, scoring, awards, donor reporting | 2 VMs, PostgreSQL 14 | 120 GB | All 60 staff, plus external applicants |
The ten cards
☺ Like you’re 10: Read one. Three minutes. One answer. Next.
Timer on. Try not to read ahead — several cards are harder if you have already been taught by the card after them, and taking them cold, in order, is the whole point.
Card 1 — Hallowfield Housing Trust · TenantTrack
TenantTrack was built in 2009 by a contractor and has behaved impeccably ever since: a supported operating system, a supported database, 180 GB, nothing exotic, no integrations anybody can name. It holds tenancy records, rent accounts and repairs history for 4,100 homes. Asked in the assessment workshop how critical it is, the Head of Housing says, without hesitating, that the Trust cannot function without it — and points out that they pay for twelve named user licences every year.
The application keeps its own login audit, running since 2019. Over the last eighteen months it records an average of three interactive sessions a month, every one of them from a single shared account called reporting, and every one of them doing the same thing: running one saved query and exporting the result to CSV. The CSV is emailed to Finance, who paste it into the rent-arrears model. Tenancy records themselves have been mastered in the housing-management SaaS platform the Trust bought in 2021; nobody has typed a new tenancy into TenantTrack since the parallel run ended eighteen months ago.
Your call: one R, one sentence, and the fact that drove it.
Card 2 — Corvid Analytics · ParquetLake
The board has committed publicly to “everything in the cloud by year end,” and ParquetLake is the largest single line on the hosting bill: about $95,000 a year of rack space, power, hardware amortization and support, carrying 340 TB of columnar market data and the Spark cluster that queries it. The economics look outstanding on first inspection — the cluster is idle roughly 70% of the week, which is exactly the profile that ephemeral cloud compute eats for breakfast, and nobody enjoys replacing failed drives at two in the morning.
Discovery also measured what leaves the building. Forty-one partner organizations pull extracts from the lake continuously — roughly 90 TB a month — and internal analysts pull another 15 TB down to their laptops. Priced at the egress rate the shortlisted provider quoted in its proposal, that traffic works out at about $310,000 a year. Today that number is zero, because moving bytes out of a rack you own costs nothing beyond the circuit you already pay for.
Your call: one R, one sentence, and the fact that drove it.
Card 3 — Meridian Diagnostics · PathologyPortal
PathologyPortal is a vendor product: clinicians at 300 GP practices log in and download pathology reports. The vendor is healthy, the product is fully supported, and the relationship is good. The data centre lease, however, ends in eight months and will not be renewed. Conveniently, the vendor now sells a hosted SaaS edition of exactly this product; it costs less than Meridian’s current run-rate and it removes an upgrade cycle the two-person infrastructure team openly dreads. The account manager is enthusiastic. The migration lead has already written “Repurchase” in the spreadsheet and moved on.
Meridian’s regulator requires that identifiable patient data is processed and stored in-country. The vendor’s SaaS edition runs in two regions, neither of them in-country. An in-country region appears on the vendor’s published roadmap for “next year,” with no date attached, no clause offered in the draft contract, and — when pressed by Meridian’s counsel — no willingness to commit to one in writing before the lease expires.
Your call: one R, one sentence, and the fact that drove it.
Card 4 — Ravensgate Borough Council · PermitDesk
Everything about PermitDesk says rebuild-or-bin. Classic ASP. Windows Server 2012 R2. SQL Server 2008 R2. No source control, no tests, no living author. It holds every planning permit the borough issued between 2007 and 2022, and the council has a statutory duty to keep planning decisions queryable for fifteen years.
Live permitting moved to a SaaS planning platform in 2022, and nobody has raised a new permit in PermitDesk since. What remains is about thirty read-only lookups a month, almost all of them conveyancing solicitors doing local searches, and each one is answered the same way: a clerk runs one of two saved queries and prints a decision notice in the layout the search response requires.
Your call: one R, one sentence, and the fact that drove it.
Card 5 — Calderhorn Freight · the Yard estate
The lease on Calderhorn’s three racks ends in five months, with no extension available at any price — the facility itself is closing. Inside those racks are 420 virtual machines on 26 vSphere hosts. The platform team is four people, two of whom also staff the helpdesk. There is a CMDB; it was last accurate in 2019. Nobody in the building can produce an application-level inventory, and the people who could have are three reorganizations ago.
About 300 of the 420 VMs belong to a single vendor-supported ERP suite. Its support matrix certifies vSphere and nothing else, and the vendor’s written position — obtained last month, in an email the migration lead has now read twice — is that running the suite on the target cloud’s native compute would place the estate out of support. The other 120 VMs are the usual sediment: file servers, print servers, a dozen small line-of-business apps, and eleven dev machines nobody has logged into since 2021. The migration lead has proposed doing this properly: full discovery, an application-by-application disposition matrix, then per-app rehost and replatform.
Your call: one R, one sentence, and the fact that drove it.
Card 6 — Sable & Finch · TimeLedger
TimeLedger records every billable six minutes of a 180-lawyer firm and turns them into invoices. It is a vendor product, fully supported, on a version the vendor still patches, and the partners like it. The only pressure is hardware: the support contract on the servers it runs on ends this year, and nobody wants to buy more tin. Rehost looks like a five-minute decision.
Two clauses complicate it. The licence agreement’s hosting clause prohibits running the software on third-party multi-tenant infrastructure without the vendor’s written consent; the vendor will grant consent, at a licence uplift larger than the entire hosting saving. The same vendor sells a hosted edition of the product for roughly what Sable & Finch already pays in annual maintenance, with the database included in the price. Separately, and on an unrelated page of the brief, the firm must be able to produce seven years of billing history on demand during a costs assessment.
Your call: one R, one sentence, and the fact that drove it.
Card 7 — Puffin & Pike Books · Checkout
Checkout is a well-tested Java monolith that the three-person platform team is, honestly, quite fond of. It has one nasty shape: for about three days every November it carries roughly forty times its baseline load, and the hardware was bought for that peak, so for the other 362 days most of it does nothing but draw power. MySQL 5.7 underneath it is past its upstream end-of-life. The last serious production incident was caused by the load balancer’s own firmware. Deploys take twenty minutes and go out twice a week without drama; nobody has ever asked to deploy more often than that, including during the November peak.
The CTO has spent the winter reading about microservices, has a diagram, and has written “Refactor” on the whiteboard next to Checkout’s name.
Your call: one R, one sentence, and the fact that drove it.
Card 8 — Grimsby Marine Mutual · HullCalc
HullCalc is a spreadsheet with ambitions: Excel with a great deal of VBA, an Access database behind it, four gigabytes in total, sitting on one Windows VM. Exactly one person uses it — the chief underwriter, who wrote the first version in 1998 and who retires in June. On the face of it this is the easiest decision on the estate: one user, four gigabytes, gone by the summer, and the migration gets smaller.
Underneath, HullCalc encodes twenty-six years of pricing rules — vessel-class loadings, trading-area factors, a bespoke deductible curve — that exist in no other document anybody has been able to produce. Three monthly management reports consume its output, and the annual capital submission to the regulator quotes figures it calculates. Nobody else in the business has ever opened it.
Your call: one R, one sentence, and the fact that drove it.
Card 9 — Wexley Foods · ColdChain-Edge
Forty small, individually-managed, inconsistently-patched servers — one per depot — is exactly the kind of thing a cloud migration exists to delete, and the target provider’s managed IoT service ingests telemetry beautifully and costs almost nothing at this volume. Each depot box reads temperature probes across the chilled and frozen chambers, logs readings to a central historian, and trips a physical alarm relay when a chamber drifts out of range.
Two facts sit underneath the obvious answer. The alarm must trip that relay within about two seconds of a chamber going out of range. And the depots’ broadband links drop for hours several times a year — one rural depot lost its line for a day and a half last winter. A chamber failure that goes un-alarmed is a six-figure stock write-off plus a food-safety notification the group would very much prefer not to make twice.
Your call: one R, one sentence, and the fact that drove it.
Card 10 — Larkspur Foundation · GrantHub
GrantHub has been built and rebuilt over twelve years by an in-house developer everyone likes and everyone depends on. It runs the whole grant cycle: applications, scoring, awards, payment schedules, donor reporting. The standard line in the building, offered before anyone asks, is that nothing off the shelf does what the Foundation needs.
A feature-by-feature comparison run during discovery found that thirty-one of GrantHub’s thirty-four features are standard in each of three mature off-the-shelf grants platforms. The three that are not are a bespoke scoring formula and two report layouts, and at least two of the three vendors can reproduce all three as configuration. The developer currently spends about 60% of her time maintaining GrantHub. The Foundation’s genuinely distinctive work — long-run impact measurement across a twenty-year portfolio, which no vendor sells and no peer does well — has no software supporting it at all.
Your call: one R, one sentence, and the fact that drove it.
Go back through your ten rows and, for each one, add a single line: “I would change my mind if ___.” Do it now, before you see any answers. That sentence is the most useful thing on a real disposition matrix — it is the column a steering committee actually interrogates, it is what turns “Repurchase” from an assertion into a decision somebody can challenge, and it is the only part of your answer that stays valuable after the facts change. If you cannot complete the sentence for a row, you have not decided that row. You have guessed it.
Show the worked answers — all ten cards
- Card 1 · TenantTrack — Retire. Twelve licences is what somebody bought; three sessions a month is what somebody uses, and only one of those two numbers is evidence. The tenancy data has been mastered elsewhere since 2021 and the only living function is one monthly CSV, which the SaaS that already owns the data can produce as a scheduled report. Retire the application, keep a read-only export of the historical rent-account rows for the retention period, and switch off three VMs and a SQL Server Standard licence. Defensible second-best: Rehost onto the smallest instance that runs it, for a defined two-quarter verification window. The fact that would make the second one correct: evidence of non-interactive consumers. An application’s own login audit records people; it almost never records ODBC connections, linked servers, scheduled jobs or an API somebody wired up in 2016. Before you switch anything off, run a week of database-connection or network-flow logging — when a quiet system breaks something on the way out, it is nearly always a job, not a user.
- Card 2 · ParquetLake — Retain. The move creates a cost that does not currently exist: roughly $310,000 a year of egress against a $95,000 hosting saving. That is not a close call, and it is the textbook shape of data gravity — the data is not immovable because it is large, it is immovable because of what is attached to it. Write the reason into the matrix in exactly those words, because “Retain” with no reason reads as “we ran out of time.” Defensible second-best: Replatform — split the workload. Move the Spark cluster, which really is idle 70% of the week and really is cheaper as elastic compute, and keep the data on-premises or in a colo adjacent to a dedicated circuit, paying for the link rather than for per-GB internet egress. In plenty of firms that is the better answer. The fact that would make it correct — or make Retain wrong outright: who pays for the egress. If the 41 partners can be moved to a requester-pays arrangement, or billed for extraction under their contracts, or if the provider will write a committed-egress discount into the deal, the $310,000 stops being yours and Retain loses its entire justification. Ask that question before you build an architecture around a number somebody quoted you.
- Card 3 · PathologyPortal — Rehost. The lease ends in eight months, so Retain is not on the menu; the SaaS edition cannot hold identifiable patient data in-country, so Repurchase is not either — not yet. Move the vendor’s on-premises edition unchanged into an in-country cloud region, clear the lease, and keep the compliance position intact. Defensible second-best: Repurchase — and it is the right long-term destination. The fact that would make it correct now: a signed date. A published roadmap is a marketing artefact; a contract clause naming an in-country region, a go-live date before your cutover, and a remedy if it slips is a plan. Write the trigger into your migration plan explicitly — “Repurchase when clause 8.3 goes live; Rehost is the position until then” — so it becomes a scheduled review rather than an argument you have all over again in eighteen months. Treating a roadmap as a commitment is one of the most common routes from a repurchase decision to a compliance incident.
- Card 4 · PermitDesk — Retire, and note that Retire is not Delete. The statutory duty attaches to the records, not to the Classic ASP application wrapped around them. Export the record set to a read-only archive — a small managed database or an indexed document store — reproduce the two saved queries and the decision-notice layout that local searches actually need, get the records manager to confirm in writing that the archive satisfies the fifteen-year duty, and only then switch the application off. You also take a Windows Server 2012 R2 machine running SQL Server 2008 R2 off the network, which is very plausibly worth more than the hosting saving on its own. Defensible second-best: Rehost the VM as-is into an isolated subnet reachable only by the clerks. The fact that would make it correct: a records-management or legal opinion finding that the duty attaches to the system’s own rendering of a decision rather than to the underlying data, or a finding that the export cannot faithfully reproduce a decision notice. Get that opinion in writing before you build either answer — and if the answer is Rehost, be explicit that you have chosen to keep an out-of-support Windows estate alive for over a decade to serve thirty lookups a month, because that is a security decision, not an architecture one.
- Card 5 · the Yard estate — Relocate. Five months, four people, no application inventory, and 300 of the 420 VMs pinned to vSphere by a vendor support matrix. Move the vSphere estate wholesale into the cloud’s managed VMware service, preserve the ERP’s support position, clear the lease date, and then do the disposition work from a position where the deadline is no longer holding a gun to your head. This is precisely the brief Relocate was invented for: it is the one R that trades cloud benefit for schedule certainty, and here schedule certainty is the binding constraint. Defensible second-best: a split — relocate the 300-VM ERP block and rehost the other 120 in parallel, which is more work now and banks real savings sooner. The fact that would make it correct: the ERP vendor’s current support matrix. Check it this quarter, not in the 2022 project pack — if the vendor now certifies the target cloud’s native instances, Relocate’s central justification evaporates, and since Relocate banks no licensing saving, no rightsizing saving and no rationalization, the split becomes clearly better. Whichever you pick, write down what Relocate does not excuse: those eleven dead dev machines now cost you every month instead of once, and the discovery you deferred has to appear in the plan as a funded project with a date, or it will simply never happen.
- Card 6 · TimeLedger — Repurchase. The hosting clause is the whole card. Rehosting to public cloud requires vendor consent at an uplift larger than the saving, which makes the “cheap, obvious” option the most expensive one on the table; the vendor’s own hosted edition resolves the clause instead of paying to work around it, and retires the Oracle SE licence at the same time. Defensible second-best: Retain on-premises until the licence anniversary in eighteen months, then repurchase at the natural break — genuinely correct if the firm’s own building is not closing, because everything you spend moving early is the uplift you were trying to avoid. The fact that would make it correct — and the check you must do either way: the export. Prove, before signing, that the hosted edition can produce seven years of billing history in a form a costs assessment accepts, and get a data-exit clause into the contract. A repurchase whose exit terms nobody has read is not a migration; it is lock-in with a nicer login screen. The general rule this card is really teaching: read the hosting clause of every vendor licence in the estate during discovery. On a real portfolio it changes several rows, and it is always cheaper to find in month one than in cutover week.
- Card 7 · Checkout — Replatform. Name the binding constraint before you name the R. Here it is two things, and neither is architecture: hardware bought for a peak that happens three days a year, and a database past end-of-life. Both are relieved by a lift-tinker-shift — the web tier into an autoscaling group behind a cloud load balancer, so November costs you three days instead of a year of idle iron, and MySQL 5.7 onto a managed MySQL 8, so the end-of-life problem becomes somebody else’s job — without touching a line of business logic. Defensible second-best: Refactor. The fact that would make it correct: evidence that the binding constraint is change speed rather than capacity — a deploy freeze that blocks the team for eight weeks around peak, or one component whose scaling profile has genuinely diverged from the rest of the monolith. Nothing in the card says that. Twice-weekly twenty-minute deploys with no drama is not a change-speed problem, and a three-person team is exactly the size a distributed microservice estate punishes hardest. The CTO’s diagram is an answer in search of a constraint.
- Card 8 · HullCalc — Refactor / re-architect, scheduled before June. One user and four gigabytes makes this look like the estate’s cheapest Retire, and it is the opposite: twenty-six years of pricing rules exist only inside it, three management reports and the regulator’s capital submission consume its output, and the only person who understands it leaves in June. Re-implement the rules in the pricing platform, with knowledge extraction as task one and the chief underwriter’s remaining months booked as a project resource rather than assumed as goodwill. The deadline on this card is not a lease and cannot be renegotiated, extended, or paid to move — it is a person’s last working day. Defensible second-best: Rehost — freeze the VM, move it as-is, buy time. Legitimate only if you write down what you have actually done: converted a people-risk into a permanent technical liability on a machine nobody can change, with a review date and a named owner in the plan. That is a tactical answer, not a resolution, and calling it one is how estates end up running a 1998 spreadsheet in production in 2035. The fact that would change everything: whether the rules are genuinely undocumented. Go and read the filed actuarial basis before you assume — if the pricing rules are already written up there, the knowledge is not trapped, the rebuild collapses into a configuration exercise in the existing platform, and the card becomes a Retire plus a ticket. That check costs one afternoon.
- Card 9 · ColdChain-Edge — Retain at the edge. A safety loop that has to keep working while the network is down does not belong on the far side of the network. A two-second budget and links that fail for hours settle it: the control-and-alarm function stays in the depot. Write that reason into the matrix in those words — a Retain with a stated physical reason is a decision; a Retain with no reason reads as a backlog item. Defensible second-best: Replatform — keep the loop local but stop running forty unmanaged snowflakes: move to a supported edge runtime or a managed edge-gateway product, centrally patched and centrally monitored, with the historian and all the analytics in the cloud. That is the answer most estates should actually reach, and it is the clearest demonstration that Retain and “do nothing” are not synonyms. The fact that would move the whole card to the cloud: whether the safety function genuinely depends on your software. Many freezer controllers can drive an alarm relay independently — if this one can, your system is pure telemetry, latency stops mattering, WAN outages stop mattering, and there is no argument left to have. One conversation with the refrigeration engineer decides which world you are in, and it is a conversation nobody on the migration team thinks to have.
- Card 10 · GrantHub — Repurchase. The test for building software is never “does it fit us” — a thing you built always fits you — it is “is this what we are for.” Grants administration is a commodity with three mature vendors and a thirty-one-of-thirty-four feature match; long-run impact measurement is the thing nobody can build for this Foundation but this Foundation, and it currently has no software at all. Repurchasing here does not just cut a run-rate, it returns the only developer to the only work that differentiates. Defensible second-best: Replatform — move as-is onto managed PostgreSQL now, repurchase after the round closes. The fact that decides between them: the calendar. If the data-centre exit lands mid-grant-round, cutting over to a new platform in the middle of an assessment cycle is a needless risk, and sequencing — replatform now, repurchase at the natural break — is simply the professional answer. It is also the question most likely to go unasked in a room full of people arguing about features. One more thing to check before committing: if the bespoke scoring formula is fixed by a trust deed or a funder’s contract in a way no vendor can configure, then the comparison that produced “thirty-one of thirty-four” was measuring the wrong thing, and the card changes completely.
Scoring — and what your score actually means
☺ Like you’re 10: Getting the right answer for the wrong reason still counts as getting it wrong.
Score the fact column first, then the R column. Four separate results are worth knowing, and they mean different things:
| What happened | What it means | What to do about it |
|---|---|---|
| Right R, right fact | You found the note under the box. This is the only outcome that transfers to a system you have never seen. | Nothing. Go and do the harder version in Foxy’s challenge below. |
| Right R, wrong fact | Luck, or a good general instinct applied to the wrong evidence. It will not survive the first follow-up question in a steering committee. | Re-read that card and find what you skimmed. It will be in an ordinary sentence in the second paragraph. |
| Wrong R, right fact | The most correctable result there is — you found the deciding evidence and reasoned from it imperfectly. | Compare your sentence with the key’s. Usually one of the five decision criteria was weighted wrongly, not missed. |
| Wrong R, high confidence | A belief, not a gap in the brief. Far more interesting than a wrong guess. | Write the belief down in one line. That line is the most valuable thing this drill produces. |
Six or seven right out of ten on a first pass, with sound reasoning on the ones you missed, is a good result — these cards are built to be beatable only by someone who reads for the buried fact rather than the headline. And note the shape of the answer set before you draw conclusions from your score: four of the ten land on Retire or Retain, which is not a trick. Real portfolio assessments routinely find that a quarter to a third of an estate should not move at all, and the two R’s that cost nothing to execute are exactly the two that get decided last, or never, on the programmes that end up over budget.
What generalizes — the six facts that flip an R
☺ Like you’re 10: There are only about six kinds of note taped under the box. Learn the six, and you start looking for them.
Ten cards, ten buried facts — but not ten different kinds of fact. They fall into six families, and the reason this drill is worth repeating is that the families transfer completely: to an estate you have never seen, in an industry you know nothing about.
| The family | What it looks like | Cards | The question that finds it |
|---|---|---|---|
| Usage evidence versus claimed criticality | A system everybody calls essential that three people touch; a system nobody mentions that four others depend on | 1, 4, 10 | “What does the log say?” — and then, “what does the log not record?” |
| Contract and licence clauses | Hosting clauses, support matrices, uplifts on consent, exit and export terms | 3, 5, 6 | “Who has read the actual agreement, this year?” |
| Regulation, residency and records | Where data may live; what must stay queryable and for how long; whether the duty attaches to the data or to the system | 3, 4 | “Is that written down by a regulator, or by a vendor’s marketing team?” |
| Costs that only exist after the move | Egress, licence uplifts, commitments, support tiers, per-request charges | 2, 6 | “What is free today that stops being free?” |
| Physics | Latency budgets, and what has to keep working when the network does not | 9 | “What happens to this at 3 a.m. with the link down?” |
| Deadlines that are people, not contracts | The one person who knows; the retirement date; the contractor whose engagement ends | 8 | “If this person left tomorrow, what could we no longer do?” |
Card 7 belongs to none of them, and is a seventh habit rather than a seventh family: name the binding constraint before you name the R. Almost every wrong answer in this drill — and most of the expensive ones in real life — comes from matching the strategy to the most impressive-sounding option, or to whatever the room already wanted, instead of to the constraint that is actually hurting. Capacity is not change speed. A lease date is not a licensing problem. A person leaving is not a hardware refresh.
Every one of these ten cards was decided by a fact a competent person could have found in under a day: an audit log, a licence PDF, a support matrix, an egress estimate, a conversation with a refrigeration engineer, a retirement date sitting in HR. None of them needed a discovery tool, a consultant, or a workshop. Wrong R’s are common not because the deciding facts are hidden, but because nobody goes and looks — the obvious answer arrives so quickly, and feels so complete. Speed is the trap, and three minutes is plenty of time to ask one good question.
Three harder versions, in increasing order of discomfort. One: re-run the ten cards changing exactly one variable each time — give Meridian a two-year lease instead of eight months, give Calderhorn twelve people instead of four, give Corvid a partner contract that bills for extraction — and watch how many answers move. Most will. Two: take the four cards you found hardest and write the steering-committee sentence for each: one sentence a non-technical trustee or partner would accept, with no jargon and no “it depends,” that survives being repeated back to you slightly wrong. Three: do the ten cards alongside someone else, separately, then argue. The disagreements will cluster on cards 5 and 7, and the argument is worth more than either of your answers — because on a real programme the disposition matrix is not the deliverable, the agreement behind it is.
Milestones
☺ Like you’re 10: Tick each box as you finish a card. Your ticks are saved in this browser.
Ten cards, then two steps that turn a scored quiz into a skill you keep. Progress is stored locally on this device only.
The ten cards, three minutes each
Turning a score into a skill
scored column as right-R-right-fact, right-R-wrong-fact, wrong-R-right-fact or wrong-both — and you can say which of the four you had most of.Foxy: Three minutes a card? I could do these in thirty seconds each. Ancient app, out of support — Refactor. Tiny app, one user — Retire. Enormous VMware estate — Relocate. Next, next, next.
Professor Owl: Then do card three in thirty seconds, Foxy. Out loud.
Foxy: Easy. The vendor sells a hosted edition, it costs less than they pay now, and the team hates upgrades. Repurchase.
Professor Owl: And the regulator?
Foxy: …in-country. And the vendor’s in-country region is on a roadmap. With no date on it. So it’s a Rehost, and the Repurchase becomes something I schedule for the day somebody actually signs something. Fine. Thirty seconds got me the wrong answer with total confidence, which is worse than no answer at all.
Nutty the Squirrel: Which is why I count before anybody argues. Card one had twelve licences and three sessions a month. One of those numbers is what somebody bought and one is what somebody uses — and only one of them is evidence.
Gizmo the Gremlin: But the Head of Housing said it was business-critical! She said it with feeling! Are you calling her a liar? 😇
Timmy the Turtle: I’m saying a feeling isn’t a log line, Gizmo. And before anyone switches TenantTrack off, I want a week of connection logging — because that audit only records people, and the thing most likely to break is a job nobody remembers writing.
Professor Owl: Which is the whole drill in one sentence. The R is the easy part. The fact underneath it is the work — and it is almost never the fact the room brought into the meeting with them.
1. Two of these ten cards look like an obvious Retire and are not; two look like anything but a Retire and are. Name one of each, and the single fact that decided it. 2. Card 2’s entire argument rests on a cost that does not exist today — name it, and name the one commercial change that would make Retain the wrong answer. 3. Cards 3, 5 and 6 each turn on a document rather than a technology. Name the document in each case. 4. Card 7’s team wanted to Refactor and the evidence said Replatform. What is the general test, and what evidence would have made Refactor correct?
Check your answers
- Looks like Retire and isn’t: HullCalc (card 8) — one user, four gigabytes, and that user retires in June, which reads as a free win right up until you notice that twenty-six years of pricing rules exist nowhere else and the regulator’s capital submission quotes its output. Looks like anything but Retire and is: TenantTrack (card 1), which the Head of Housing calls business-critical and which the application’s own audit log shows is touched three times a month by one shared reporting account — or PermitDesk (card 4), which looks like a rebuild because of its age and is really a records-export exercise. The deciding facts, in order: the pricing rules exist in no other document; the audit log versus the licence count; the statutory duty attaching to the record rather than to the application.
- The cost is egress — roughly $310,000 a year to serve 41 partner organizations and the internal analysts, against a $95,000 hosting saving. Today it is zero, because moving data out of a rack you own is free. The commercial change that flips it is who pays: move the partners onto a requester-pays arrangement or bill extraction under their contracts, or negotiate a committed-egress discount into the provider deal. Change that and Retain loses its whole justification, and the split-the-workload Replatform becomes the obvious answer. This is the general lesson of data gravity — data is not stuck because it is big, it is stuck because of what is attached to it, and some of those attachments are commercial rather than technical.
- Card 3: the vendor’s contract — specifically the absence of a clause committing to an in-country region by a date, versus a roadmap promising one with no date and no remedy. Card 5: the ERP vendor’s support matrix, which certifies vSphere and nothing else, and whose current version — not the one in the 2022 project pack — decides between Relocate and a split. Card 6: the licence agreement’s hosting clause, which forbids third-party multi-tenant hosting without consent and prices that consent above the saving. None of the three is discoverable by looking at the technology. All three are discoverable in week one by somebody willing to read a PDF.
- The general test: name the binding constraint before you name the R, then check that the strategy you picked actually relieves that constraint. Checkout’s constraints are capacity elasticity (hardware bought for a three-day annual peak) and an end-of-life database — both fully relieved by a replatform, neither touched by a rewrite. Refactor would have been correct on evidence of a change-speed constraint: a deploy freeze blocking the team for weeks around peak, one component whose scaling profile has genuinely diverged from the rest, or a codebase where a change in one area routinely breaks another. Twice-weekly twenty-minute deploys, no drama, three people on the team, is the opposite of that evidence — and three people is precisely the team size a distributed estate punishes hardest.
That’s the drill. If several answers surprised you, the theory that explains them is in The 7 R’s — particularly the five decision criteria and the quick decision path — with the money side in What & Why We Move and real-world versions of all ten shapes in Case Studies. Ready for a different single skill on the clock? Try Drill — Size the Data Move, where the arithmetic gives an unambiguous answer for once, or Drill — Write a Rollback Plan, where it very much does not. To do all of this against one estate, with continuity across five parts, start at Move Brambleside — Start Here and take Part 1 — Discovery & the Disposition Matrix, which is this drill again with sixteen systems that all depend on each other.