Work Case study My own product
Five Rivers.
A case, copy and custody system for a network of libraries, and the release discipline that has kept a 30,300-copy dataset and every account intact through every upgrade.
tested deployed independently checked
01 What it is
The system
- Kind
- Web application, invite-only
- Stack
- Django and PostgreSQL in containers
- Access
- Behind a Cloudflare Tunnel, two-factor sign-in
- Latest release
- 1.2.0, 14 October 2026
- Dataset
- 30,300 copies, preserved through every upgrade
Five Rivers runs the paperwork of a library network: intake and triage of cases, correction cases, custody shipments between libraries, an extraction review, and trackers with turnaround targets. It is the kind of application where a quiet data error costs someone a week, so the interesting part is not the feature list. It is how changes reach production.
Release 1.2.0 added access requests and a tracker report. What follows is what happened before it went live.
What is not here. The network’s name, its libraries, its people, and anything from inside the data. That is deliberate, and it is the same rule I apply to client work.
02 How a release ships
Nine phases, each one allowed to say no.
A guided upgrade script walks every release through the phases below. Each phase refuses to continue on any mismatch. Three kinds of rollback exist, and all of them are exercised in rehearsal. deployed
- preflight
- load
- backup
- static
- proof
- live
- deploy
- verify
- record
Proof is the phase that matters most: it runs the whole release against a copy of the live database before anything live changes. If the migration, the data or the code disagree, the release stops there, with production untouched.
03 Evidence before shipping
What ran before 1.2.0 was allowed out.
Different checks see different things, so several kinds run. Each line below is a run that is on file. tested
- A 73-case security suite, run on both PostgreSQL and SQLite, so a database-specific assumption cannot hide.
- 24 browser workflow checks through the journeys that matter to the people who use it.
- 848 layout checks: 4 screen widths × light and dark × 2 data windows. These caught two real display bugs, a phone-width overflow and a dark-mode contrast failure. Both were fixed before release.
- A full 1.1.1 → 1.2.0 upgrade rehearsal of 51 checks, plus a 22-check root-path rehearsal, with every rollback exercised.
- An independent review by a second AI model before release. independently checked
04 In production
Preserved: true.
Production upgrades have kept the existing 30,300-copy dataset and every account intact, and the verify phase records it in so many words: preserved: true. Outages are measured in seconds. deployed
None of that is luck. It is the proof phase catching problems where they are cheap, and rollbacks that have been run often enough to be boring.
05 What this means for your project
Mechanisms, not symptoms, in practice.
Why the layout sweep found what the other checks missed
The workflow checks and the security suite were green. They were looking at behaviour and permissions, not at pixels in dark mode at phone width. The two display bugs were only visible to a check that was deliberately looking across widths and themes. Coverage is a question of what is being looked at, not just how many checks there are.
Why “proof” comes before “live”
The most expensive place to meet real data for the first time is production. Running the whole release against a copy moves that meeting to where failure costs nothing. If your deployment process does not have an equivalent, that is usually the first thing I would suggest.
Why a second model reviews the first
The author of a change is the worst-placed reviewer of it, whether the author is a person or a model. An independent reviewer with no stake in the original answer catches a different class of mistake. Fixes go round until the reviewer says ship.
What I would bring to your release process
The same questions: where is your proof phase, which rollback have you actually run, which widths and themes do your checks see, and who reviews the author? Then the checks to answer them. Test automation with Playwright is where most teams start.
Next