Browse the reference ⌄
Worked Example published

Worked Example: A Migration Through the Factory

A composite data-centre exit walked as a reusable decision trail: purpose sheet, recipe boundary, inventory as a discovery control, the control loop, transform review with adversarial cases and business reconciliation checks, rehearsals run by role, and the effort test that tells a factory from a pile of tools.

Version
0.1.0
Owner
DEIXEI
Last reviewed
Reading time
26 min

For Cto · Ai Architect · Platform Engineer · Product Leader · Risk And Assurance · Internal Audit

Sources and details
Source works
Book 2 · Book 09
Source chapters
11, 22, 28
Refresh trigger
source change

Full source list ↓ Source index →

Case status. Mary’s firm is a light composite: a growing, mid-sized regulated firm, deliberately distinct from Asset Manager X and every other composite in the books. Every figure below (counts, days, euros, rows) is the composite’s illustrative number, not a benchmark. The page separates source claims (what the books say happened or recommend), page recommendations (syntheses labelled as such) and the composite facts themselves.

Two tellings. The same migration is told twice. Book 2 Chapter 28 gives the method (six parts, the evidence package, five anti-patterns) and describes a nine-month exit. Book 09 Chapter 11 retells it with a ten-month window, concrete counts and three transform rejections. The details differ (duration, which policy-versus-actuals findings surfaced). This page uses Book 2 for the method and Book 09 for the counts and the transform review, and says which it is using where they diverge.

The decision this page supports

You have a business-critical application that mostly works, a fixed exit date, and coding agents that can generate migration scripts, tests and infrastructure faster than anyone can review them. The question is not whether agents can do the typing. It is which decisions must stay with named people, which artefact records each one, and which reference page governs it, so that speed does not recreate the sprawl you are migrating to escape. The trail below is reusable: replace the composite facts with yours and keep the decision order.

The starting position (composite)

FactFirst migration (2023)This migration
First governed environment (access approved, secrets stored, logging connected)53 working daysOne night, provisioned from the recipe
Who carried itOne engineer on every night of a sixteen-name rota; twelve of the last fifteen weekendsNine names on the rehearsal rota, each against a role
Knowledge of the systemOne engineer’s memory; a runbook corrected out loud during incidentsPurpose sheet, inventory, review comments and the evidence record
Outcome signalEvery dashboard green on cut-over morning; seventeen incidents in six weeks traced to undocumented dependencies, filed as “teething”Rehearsals with findings caught before cut-over

The lesson the books draw from the 2023 column: a firm that needs an extraordinary engineer to survive an ordinary migration has found someone to absorb the problem, not solved it. Agents cannot work from what is in someone’s head, so getting it out becomes a precondition of the work rather than a second job with no hours (Book 09 Ch 11).

The decision trail

Each row is a decision, its owner, the artefact that records it, and the reference page that governs it. Rows 0 to 3 happen before any migration code exists.

#DecisionOwnerArtefact / evidenceGoverning page
0Keep the system agent-recoverable before the migration is fundedTechnology management, via the readiness reviewCharacterisation tests, traffic capture, resurrected buildAgent readiness
1What is the migration for, and what may not change?Product owner (not the migration engineer)Purpose sheet committed as intent.mdAgent-speed SDLC
2Recovery point and recovery timeOwner of the operations the application servesSigned decision on the purpose sheetAccountability
3Which recipe, and what agents may and may not generateProduct team selects; platform owns the recipeRecipe reference in the BOM; agent boundary statementBOM and recipes
4What does the system contain and touch, and who owns each part?Product team; humans validate ownership and meaningInventory (BOM) with an owner or a blank per entry; findings listBOM and recipes
5Provision the linePlatform (automation factory)Repositories, pipelines, environments, secret stores, logging, dashboards, evidence package, each with an ownerApplication foundation, landing zone
6Compare rules with what is running, weekly then continuouslyFirst line runs it; second line adjudicatesRed rows; recipe or identity rule changesPolicy, audit, actuals
7Who reviews which change?Rule routes to a role, not a personRouting rule and reason on each changeChange governance, accountability
8Accept or reject each generated transformReviewer for data transforms (by role)Approval or rejection record with count query attached; new regression testAgent-speed SDLC
9Unrecoverable meaningBusiness owner of the dataAccepted limitation with review date and compensating controlAccountability
10Rehearse and roll backNight lead per rota slot; runbook names the call chainFreeze, extract and approval entries in the evidence recordRelease engineering
11Is the new system used the way the intent assumed?Product owner / adoption leadPer-screen usage view read from the productFactory measurement
12Renew the vendor product or rebuild on the foundation?Portfolio decision against a frameworkSourcing decision recordBuild, configure or consume
13Show where a human overrode the machineAnyone, by queryRejections in the evidence recordAudit evidence

The rest of the page explains how each decision was taken in the composite and what to check.

0. Keep the system above the event horizon

Source claim (Book 2 Ch 28 §4; Book 09 Ch 11 note). Agents could carry this migration only because the old system’s behaviour was still extractable: the system itself acted as the specification and verification could be comparison. When the firm first graded its estate for agent readiness, a year before the hosting notice, this application scored low on every count because nobody had touched it for years. It received three small investments while its migration was still unfunded:

  1. Characterisation tests generated by agents from observed behaviour and reviewed by people, pinning down what the system did, including behaviour nobody could explain any more.
  2. Traffic capture at the boundaries, so the new system could be tested against what the old one actually received and sent, not what the documentation said.
  3. Resurrected build, assembling deterministically from source, so a generated change could be checked rather than believed.

Acceptance check. Before an exit date is real, each legacy system on the watch list has the three holding actions or a recorded decision to let it fall below the horizon. The method and scoring are on agent readiness.

1–2. Write the purpose before choosing tools

The first meeting, a week before anything was provisioned, opened on what the migration was for. The product team (not the migration engineer) wrote it on one sheet.

Purpose-sheet fieldComposite contentWhy it matters
Business outcome / why it cannot waitData-centre exit; resilience not currently held; faster change; better evidence; a foundation for later AI-assisted workThe anchor against which “clever shortcuts” are tested
Users and operating hoursWho uses it and whenDrives cut-over windows and who to warn first
Risk classificationThe application’s risk classSelects the recipe’s control set
Data domainsWhat data it holdsDrives data-class and lane decisions
Regulatory constraintsApplicable obligationsLinked to the regulatory map, not restated
Recovery expectationsRecovery point and recovery timeA shortcut that quietly changes these is a design change
Integration dependenciesWhat depends on it, and what it depends onSeeds the inventory

Book 2 names the artefact intent.md and lists the fields above; Book 09 calls it a single sheet with the same content. The recovery point was a decision signed by the owner of the operations the application serves, after working out with the product owner what an hour of lost data would cost her team at month-end. It was not a number the engineer assumed for want of anyone telling her.

Page recommendation: treat any field on the purpose sheet that only the migration engineer could fill in as a finding. The point of the sheet is that someone other than the hero can read it.

3. Select the recipe and draw the agent boundary

The team chose the firm’s recipe for a regulated internal-workflow application: cloud-hosted API, managed database, identity integration, audit logging, controlled deployment, staged migration. The recipe fixed repositories, environments, identity, logging, secrets, tests, deployment stages and evidence requirements before anyone argued about them.

The boundary (source claim, both books).

Agents may generateAgents may not
Migration and transform scriptsInvent the delivery pattern
Test casesChoose how identity is wired, where secrets live, how logging is structured or what a deployment stage looks like, outside the recipe
API / interface clients (including the replacement for an ungoverned feed)Decide ownership or the meaning of data
Infrastructure modules within the recipe’s constraintsApprove their own output
Documentation drafts

Without a recipe, a capable agent makes a hundred reasonable small decisions per service, none of them the firm’s; across a year of migrations that manufactures the same heterogeneous estate the migration is meant to escape, at machine speed (Book 2 Ch 28 §4). Every agent thread began from the committed intent and specified its work within the recipe; the thread’s wrangler approved the plan before any code was written.

Not an AI recipe. The migrated application is not itself an AI product, so it used a conventional-application recipe from the standard foundation, not one of the AI recipe families. The agents were workers on the line, not features of it. If the firm later wants a grounded assistant over the migrated data, it starts from the nearest recipe family and applies the least-agency rule (Book 2 Ch 22); see choosing the AI product shape and the research-assistant example.

Failure signal. Two services built in the same month do the same thing in two different ways. A generated artefact that departs from the recipe is a rejection that becomes evidence, not a clever exception.

4. The inventory as a discovery control

Source claim (Book 2 Ch 28 §4). The first bill of material is uncomfortable because it exposes what nobody knows. Its job is to move discovery from the incident’s schedule to the firm’s. Agents summarised the old documentation and reconciled it against what the systems actually did; people validated ownership and meaning, because those are exactly what a model will confidently guess wrong.

The inventory recorded every component, dependency, identity, secret, certificate, API, data domain, environment and evidence requirement, each with an owner or a blank (Book 2 lists the categories; Book 09 adds “an owner or a blank”).

What it found (Book 09 Ch 11 counts): 214 components, 47 findings.

FindingWhy it was dangerousAction takenResidual
Seven interfaces in no diagramWould break silently after cut-overEach given an owner and brought into the inventoryNone recorded
Four scheduled jobs with no ownerOne was the 02:00 vendor batch file, set up by a leaver, feeding a reconciliation the business depended on without knowing it; it ran on a server due to be switched offOwner assigned; rebuilt as a governed integration; put into the recipeNone
One service account reaching 23 systems; the application needed 3Worked for years, accumulated grants one at a time; carried into the cloud it would have a larger blast radiusRe-provisioned with access to 3 systemsNone
Two report fields whose meaning had driftedDownstream reports depended on themFirst field’s meaning recovered in a weekSee below
Runbooks accurate only because one engineer corrected them aloudKnowledge lived in one headAgent-drafted documentation reviewed by that engineer, corrections attached line by lineSee below

The migration is the moment to fix identities. Everything is re-provisioned anyway, so the firm has both the excuse and the leverage. Miss it and the excess is carried into an environment where a compromised identity reaches further (Book 2 Ch 28 §4). Scoping is governed on identity and secrets.

Unrecoverable meaning becomes an accepted limitation. The second drifted field could not be recovered. Two upstream feeds had long ago diverged on what they put in it (one a settlement date, the other a trade date), nobody at the firm could say when that started, and three years of historical reports mix the two conventions. The business owner signed it off as an accepted limitation with a review date, and an analyst now spends two days at every month-end reconciling old against new by hand. Page recommendation: record such a limitation with the signer, the review date, the compensating manual control and its recurring cost, so the cost stays visible rather than becoming normal. Even a green dashboard entry for it should mean “accepted”, not “fixed”.

Turning tacit knowledge into reviewable text. An agent drafted the application’s first documentation, about forty pages, from the code and the inventory. The engineer who had corrected the 2023 runbook out loud reviewed it; each correction became a sentence attached to the line it corrected (for example, that the nightly reconciliation could not be re-run after six, because a second run doubled downstream figures). Thirty-three comments in a day; about a dozen the migration engineer had heard during 2023 incidents, four she had never heard. Asked why he had not said so before, he answered that nobody had asked. Page recommendation: the review of agent-drafted documentation by the person who holds the tacit knowledge is a scheduled task with an owner, not a favour.

5. Provision the line, and what the evidence package contains

The automation factory read the recipe overnight and built the repositories, branch policies and pipelines, infrastructure templates, two environments, secret stores, logging workspaces, dashboards and deployment workflows, each carrying an owner. The first entries in the evidence record were the provisioning itself (Book 09 Ch 11). The landing zone was a standing asset, not a per-migration build (Book 2 Ch 28 §6).

The evidence package the paved road emits as a by-product of use (Book 2 Ch 28 §6):

Element
Artefact version
Test results
Scan results
Infrastructure plan
Approvals
Deployment identity
Environment
Rollback method

Book 2 names four readers who stop chasing: the release manager no longer assembles screenshots (the pipeline produces the release record); security no longer chases scan results (they are attached to the artefact); audit queries the evidence instead of asking whether the process was followed; and operations reads the deployed version from the deployment record.

What did not hand over: understanding business rules, testing and re-testing the data migration, cleaning integrations, and users accepting changed workflows. That is the handover curve applied to one project: well-bounded provisioning moved to machinery first; data meaning, rollback design and user impact stayed with people.

6. Run the control loop from the first week

Source claim (Book 2 Ch 28 §6). The most consequential decision was to make policy, audit and actuals visible weekly and, once the line was running, continuously. The two tellings report different first findings; read them together as the kind of thing the loop finds first.

Policy saysActuals showedChange madeTelling
No production data in developmentTwo test processes still used masked-but-sensitive extractsThe recipe changedBook 2
Privileged access is time-boundThe migration support group held standing accessIdentity rules changedBook 2
Every API is registeredOne reporting tool read the database directlyGoverned API built; back door closedBook 2
Every connection is registeredAn analyst’s reporting tool connected to the test database at 07:10 every weekday, on a long-departed administrator’s credentialsRegistered interface with the analyst named as owner the following weekBook 09

The Book 09 case shows a useful disposition: the connection was not wrong, only unrecorded. Registering it under a named owner was the fix; the analyst kept her start-of-day routine.

Why the loop matters more at agent speed: agents introduce variation faster than a team typing by hand, so a calendar-based review (a quarterly access review over an estate that changes weekly) becomes ceremony. The loop keeps pace because it is continuous and compares intent, running state and drift directly. Mechanics, exception queues and the report-to-refuse switch are on policy, audit, actuals.

7–8. Data-transform review: adversarial cases and business reconciliation checks

This is the part of the migration where judgement mattered most. Three defects in the composite show three different catches: a reviewer reading for meaning, a different reviewer with different domain memory, and a business reconciliation check that no person ran.

The routing rule that sent the script to a person

Source claim (Book 09 Ch 11). A change to a transform over business-critical records, tested only by the agent that wrote it, goes to the reviewer for data transforms whatever its size. The firm’s routing could see where tests came from (test provenance; see change governance for the gate inputs). The rule named the role, “whoever owns transforms”, not the person.

Defect 1: the inner join that dropped rows silently

ElementComposite fact
ChangeAgent-generated transform carrying three years of settled-transaction history into the new database, converting each amount to the reporting currency
SpecificationConvert at the rate for the settlement currency on the settlement date
What the agent didExactly that: joined to the rate table on settlement currency and settlement date with an inner join
Hidden factThe rate table came from the old system, which always converted on the trade side, so it held rates only for currencies the firm trades. A few clients trade in one currency and settle in a currency the firm does not trade
EffectThose rows found no rate and were dropped. No error, no warning
Why tests passedForty-line script; 22 green tests generated by the same agent from the same specification; none sampled an untraded settlement currency (about thirty such transactions a month, too few to reach a sample)
How it was caughtThe reviewer read the code once for correctness (clean), then a second time looking for the place where two meanings of one thing meet. She ran a count against the first rehearsal’s extract: 1,140 rows of just over two million missing
DispositionRejected the same afternoon. The rejection stated the join, the effect, the count (query attached) and that the specification needed correcting before the script did. The specification was corrected the next morning
RegressionA new test feeds every transform a transaction settled in an untraded currency

Defect 2: reused identifiers

A later transform carried ten years of client fee history and attached it to clients by account number. The old vendor system reused account numbers once an account had been closed for seven years, a limit that appeared in no document. The second transforms reviewer, who had spent six years in fee billing, ran a count: 212 rows of history from closed accounts had been attached to the later clients who reopened those numbers. He rejected the transform with the count attached; the rule went into the inventory and a test the same evening. The first reviewer, reading after him, said she would have approved it.

Defect 3: caught by a business reconciliation check, not a reviewer

An analyst in the business team had built a check for the old system’s last year-end: every client’s fee balance must equal the closing balance on the old system’s final statement. On rehearsal data nineteen clients were out by a few euros each. The cause was a post-closure reversal: the legacy system posted fee reversals on the day after closure, the new system rejects entries dated after an account closes, and so the corrected transform lost those reversals without any error. Both reviewers had approved it. The check, running nightly, opened a defect against the transform’s owner (the second reviewer), who then went to tell the analyst that her check had caught something.

The transform review protocol (page recommendation drawn from the three cases)

Before review

  • Route by rule: any transform over business-critical records whose tests came only from the generating agent goes to the transforms role regardless of size.
  • Require a count query alongside each transform: rows in, rows out, and rows dropped or changed by each join or filter.

During review

  1. Read the code for correctness.
  2. Read it again for meaning: where two meanings of one quantity meet (trade versus settlement date or currency; booked versus value-dated balance; open versus closed account; old system’s convention versus the specification’s).
  3. For every join, ask what happens to a row that finds no match. Inner joins over reference data are a standing suspect.
  4. Run the count against rehearsal data and attach the query to the decision.
  5. Reject with a reason that names the mechanism, the effect and the count, and say whether the specification or the code must change first.

Adversarial case library (each case becomes a test every later transform must pass)

CaseComposite origin
Transaction settled in a currency the reference table does not carryDefect 1
Identifier reused after closure (account number recycled after seven years)Defect 2
Entry dated after an account’s closure (post-closure reversal)Defect 3
Recommendation: add a case each time a reviewer, check or incident finds a new meaning boundary

Business reconciliation checks

  • Ask the business owner’s team for the reconciliations they already trust (closing balances, control totals) and run them nightly against rehearsal data, opening a defect against the transform’s owner on any difference.
  • Tell the person who built a check when it catches something; it is the cheapest way to keep checks maintained.

Acceptance check. For each rejected transform, one query returns: the script, the agent’s version, the tests and who made them, the routing reason, the count and query, the rejection reason, the corrected version and the new test (the record contents Book 09 Ch 11 lists).

Route to the role, and read after

The day after the first rejection, the programme director asked for every transform to go to the reviewer who had caught it. She refused: the transforms went to the second reviewer, with her rejection and the new test, and she read his approvals after him for a month without touching the approve button. Queues ran about a day longer; one change waited two days that she could have cleared in twenty minutes. In that month he caught Defect 2, which she would have missed, and the nightly check caught Defect 3, which they had both missed. The director logged the slower queue as a risk with her name on it. The full rule is on accountability. Her reasoning: if a miss happens, fix whatever let it through; if the fix is “it must be me”, there is nothing to fix until she goes on holiday.

9–10. Rehearsals run by role, not by presence

The plan had three cut-over weekends: two rehearsals and the real cut-over. For the first rehearsal (Book 09 Ch 11):

  • The rota had nine names, each against a role; the migration engineer appeared once, on Saturday morning.
  • At the freeze, the record captured the time, the identity that made the old system read-only and the night lead’s approval. Extracts ran under the night lead’s name.
  • At 01:45 the night lead logged an unreconciled row count in the record, addressed to the reviewer for data transforms, not to a person, with the extract and query attached and marked for the morning, not urgent, load continuing. The morning reviewer found a double-counted table in ten minutes.
  • The rehearsal rolled back on schedule.

Page recommendation: the test of a rehearsal is that the person who used to be indispensable can go home, and the runbook, not habit, decides who is called. A rehearsal that only works because the hero is in the room has rehearsed nothing. Rollback design and the rehearsal sequence are on release engineering.

11. Watch usage against intent

Rehearsal one left a copy of the new system with real data for analysts to use for a week. The product owner sat beside them. The team had spent most of a month rebuilding the busiest screen, the month-end reconciliation view. Four of the six analysts he sat with ignored it, exported a plain list of settled transactions and did the month-end in a personal workbook. The platform recorded every deployment, its author and its evidence, but nothing about which screens anyone used.

Consequences: about a month of work to build what the analysts actually did into the new system, carrying the workbook’s checks across and retiring the workbook over the business owner’s signature; the second rehearsal moved by a month; the platform added a per-screen usage view read from the product itself, which the product owner read every morning. Book 09 calls it the most expensive finding of the migration, and it needed no agent and no rule, only someone watching. Book 2 states the role: the adoption lead reads usage against intent from day one, so success is measured in value collected, not a completed cut-over. Measures are on factory measurement.

12. The sourcing fork

The legacy application was near end of life, so the question under the migration was whether to renew the vendor product or rebuild the capability on the new foundation. Book 2 Ch 28 says the firm settled it with the factory’s buy-versus-build framework (sort the capability, price the reuse the foundation now gives, weigh network effects and proprietary data) as a portfolio decision, not a contest of the loudest voice. The book does not say which way it fell. Use build, configure or consume.

13. The rejected script as evidence

Source claim (Book 2 Ch 28 §6). The feared scenario (the model got it wrong) became proof of oversight: a generated artefact, a named reviewer, a documented reason and a corrected version, captured without anyone setting out to capture it. “Show me where a human overrode the machine on this migration” is answered by a query, not a reconstruction. The same record is a veto that cost something, and the quarterly count of vetoes is how the firm shows its stated values are lived ones. See audit evidence and the inspection walkthrough for how such records answer a supervisor.

The effort test: factory or pile of tools

Source claim (both books). Take the last hard change and sort where its effort went into two piles.

Pile 1: the firm has toolsPile 2: the firm has a factory
ProvisioningWhat the data means
Chasing approvalsHow the systems meet (integration strategy)
Reconstructing evidenceWhere the rollback point sits
Remembering undocumented dependenciesWho may decide what (product risk)

If pile 1 is larger, you have tools. In the composite spring, pile 2 held one inner join, one reused account number, one reversal a check caught that no person did, an unused screen, one unrecoverable field and an argument about a queue. The factory does not remove expertise; it stops wasting it.

The migration engineer’s own measure: her morning list of what she had to hold in her head shrank from about forty items in 2023 to four on day one, and three by May. The other items still existed; each now had an owner in the evidence record, and none was her.

Economics (composite, read with care)

ItemComposite figure
Delivery system cost over 18 months (platform team time, share of security and internal-audit time, tooling, cloud)About €2.1 million
Platform team sizeNever more than nine
2023 migrationAbout 1,600 contractor days, about €1.8 million, before incidents or the engineer’s weekends were priced
This migration (planned)About 600 contractor days
Applications whose changes the platform carried by springThirty-odd

Book 09 is explicit: on these numbers the platform did not pay for itself on one migration, and anyone claiming it could should be distrusted. A larger estate spreads the cost more thinly; a smaller one may not need all of it. Book 2 adds the sequencing rule: do not build the whole factory before using any of it. Build the parts against a real, hard change (a migration exercises all six at once) and keep only the parts that do work on it. The roadmap view is on the implementation roadmap.

Failure modes and recovery

FailureHow you recognise itRecovery (Book 2 Ch 28 §8 unless noted)
Cargo-culting the partsA BOM template and recipe document exist, but agents route around the recipe and the BOM is filled once and never queriedMake one part load-bearing before adding the next; a recipe the line enforces beats six artefacts nobody reads
The factory as documentEvidence is assembled for the auditor rather than queried by them; written policy and actuals drift apartMove the control into the paved road
Hero relapseUnder deadline the standing-access exception, the unowned batch file or the verbal runbook reappears “just for cut-over”; a director asks for all changes to go to one expertTreat each shortcut as a BOM entry and a loop finding; keep routing on the role and read after (Book 09 Ch 11)
Agent invents the patternTwo services built the same month do the same thing two waysHold the boundary: agents write code, the recipe decides its shape; departures are rejections recorded as evidence
Adoption blindnessGreen go-live, then a quiet unexplained fall in usage; or users working round a rebuilt screenAdoption lead reads usage against intent from day one; sit with users during rehearsal (Book 09 Ch 11)
Same-source tests trustedA generated transform passes its own generated testsRoute by test provenance; adversarial case library; business reconciliation checks (Book 09 Ch 11)
Green that means “accepted” read as “fixed”An accepted limitation shows green on the dashboardLabel the state; keep the review date and compensating cost visible

Acceptance checks for your own migration

  • The purpose sheet is committed before any migration code, written by the product team, with recovery point and time signed by the business owner.
  • The agent boundary (what agents may generate, what they may not) is written down and referenced from the BOM.
  • Every inventory entry has an owner or an explicit blank, and every blank is a finding with an owner.
  • Every identity’s real permissions were compared with its need and re-scoped during re-provisioning.
  • The policy-versus-actuals comparison ran weekly from the first week, and each red row has a disposition.
  • Transform routing names a role; same-source-tested transforms over business-critical records always reach that role.
  • Each transform decision has an attached count query; each rejection produced a regression test.
  • Business reconciliation checks run nightly on rehearsal data and open defects against the transform owner.
  • Rehearsal rotas are by role and the formerly indispensable person is not on every slot.
  • A usage view per screen exists before cut-over, and someone has watched real users work.
  • One query answers “where did a human override the machine on this migration?”
  • The effort sort has been done and pile 2 is the larger.

Using this as a template

Copy the decision-trail table, replace the composite column with your facts, and keep three rules: no row is “owned” by the migration engineer alone; every row names the artefact that records it; and every row links to the reference page whose controls apply. A row you cannot fill is the first finding of your migration.

Source notes

  • Book 2 Ch 28 states a nine-month exit and dates the second migration 2026; Book 09 Ch 11 states ten months and gives no year. The page gives neither as fact; both are composite.
  • The policy-versus-actuals findings differ between the tellings; both are shown and labelled.
  • The transform review protocol, the accepted-limitation record and the adversarial case library format are this page’s synthesis from the Book 09 narrative; the books give the cases, not a checklist.
  • Book 2 references internal chapters for each part (intent and the artefact chain, recipe and BOM, automation factory, control loop, audit spine); those topics are covered on the linked reference pages.

Source material

  • Book 2 (The AI Delivery Factory), Chapter 28, Mary's Migration, Rewritten (§4 the same migration, a different firm; §5 Mary's method in six parts; §6 what the factory did; §7 what you do with the book; §8 risks and anti-patterns)
  • Book 09 (Everything Is Green), Chapter 11, Mary's Migration (purpose sheet, inventory counts, transform rejections, route to role, two piles, composite cost note)
  • Book 2 (The AI Delivery Factory), Chapter 22, The Shape of an AI Product (least-agency rule, referenced for the later-AI option)

← All AI Delivery Factory guidance

On this page

  1. The decision this page supports
  2. The starting position (composite)
  3. The decision trail
  4. 0. Keep the system above the event horizon
  5. 1–2. Write the purpose before choosing tools
  6. 3. Select the recipe and draw the agent boundary
  7. 4. The inventory as a discovery control
  8. 5. Provision the line, and what the evidence package contains
  9. 6. Run the control loop from the first week
  10. 7–8. Data-transform review: adversarial cases and business reconciliation checks
    1. The routing rule that sent the script to a person
    2. Defect 1: the inner join that dropped rows silently
    3. Defect 2: reused identifiers
    4. Defect 3: caught by a business reconciliation check, not a reviewer
    5. The transform review protocol (page recommendation drawn from the three cases)
    6. Route to the role, and read after
  11. 9–10. Rehearsals run by role, not by presence
  12. 11. Watch usage against intent
  13. 12. The sourcing fork
  14. 13. The rejected script as evidence
  15. The effort test: factory or pile of tools
  16. Economics (composite, read with care)
  17. Failure modes and recovery
  18. Acceptance checks for your own migration
  19. Using this as a template
  20. Source notes

All pages in Worked examples →