One place for proposals, cases, approvals and the asset registry
A case study in prototyping the internal systems of a government disposition office, starting from its rules and its real paperwork
At a glance
| Situation | An office that sells and leases state assets tracked proposals in a hand-kept workbook, deadlines by memory, and approvals in letters, while its asset registry could only be read through exports |
| Approach | Three connected prototypes built from the office's guidelines and checked against real documents: an asset registry importer with reports, a proposal register, and a case tracker with approval routing and statutory deadlines |
| Window | 23 September to 22 October 2026 |
| Output | 13 registry reports, a proposal register, a case tracker with approval, broker, blacklist and reporting modules, named staff accounts with an audit log, 135 automated tests, and a set of staff training materials |
| Status | Prototypes on invented demo data, with a written list of what production would still need. None is deployed |
The situation
The office handles the disposition of government assets: public auctions, negotiated sales, unsolicited proposals and other modes allowed by its guidelines. Three kinds of information were scattered:
- The asset registry. A large national registry exports to files, but there was no easy way to ask basic questions of it, such as which assets have a valid appraisal, or which titles are in what state.
- Unsolicited proposals. A working workbook held proposals received by letter. Its summary sheet was already out of date against its own list: 76 on the summary and 90 on the list.
- Cases and approvals. Each disposition has statutory deadlines and must go through the office, then an inter-agency technical committee, then the council that approves. Tracking where a case stood, and how long each step took, depended on people.
The guidelines are specific about time. Deadlines run from milestone dates, brokers have a rating and sanction ladder, and a blacklisting procedure has its own fixed periods. That makes the rules suitable to compute rather than remember.
How the work was run
- Load the registry first. An importer for the registry's export files produced a reference dataset, and 13 HTML reports were generated from it: overview, regional summary, title status, appraisal health, a compliance dashboard, land use, data quality, maps and registry viewers. Every report page has a light, dark or automatic theme.
- Encode the process as data. The guidelines' flow, thresholds and deadlines were written down as a register of rules. A deadline engine computes every applicable deadline from a case's milestone dates, so nobody tracks them by hand.
- Build the case tracker around those rules. Every disposition mode is a case. Cases move through the office, the technical committee and the council, with a status model for that chain. Further modules cover broker accreditation (applications, authority to sell, annual ratings, the sanction ladder and a fee calculator by bid bracket) and a blacklisting procedure with each step's deadline computed.
- Add the proposal register. Every proposal appears in one list whether it came through the web portal, by letter or by email. Each has stages, a completeness check that reuses the portal's own requirement rules, matching of its lots to the registry, appraisal tracking, a reply log and one-click opening of a case.
- Check the design against real paperwork. Scanned approval letters for two real disposition packages and the proposal workbook were read and compared with the prototype. The findings produced a list of numbered requirement changes, each marked as built, partly built or not built.
- Make it multi-user. The shared password was replaced by named staff accounts with salted password hashes, a minimum length, lockout after five wrong attempts, two roles, and an audit log of sign-ins and every change to data. Every history row records who acted.
- Add a management report. Live figures cover cases by track, status and deadline severity, proposals by stage and channel, time to first reply, expiring appraisals, and council cycle times.
- Train and brief. A staff training manual, a slide deck with a presenter's briefer, and a concept deck for the finance department were generated from the same sources, with reading editions in Word and EPUB.
What the real documents changed
- Approval is by case, not by package. The first build grouped lots into packages, because the approving council's resolutions list several lots at once. The documents showed that bids, failures, acceptances and awards happen per lot, and that one bidder could win two lots. The design was replaced: each lot is its own case, and a resolution is entered on every case it listed.
- The time spent waiting between steps is what leadership will ask for. The approval letters carried enough dates to measure how long a case waits between the office's endorsement, the council's meeting and the council's letter. That became a set of cycle-time figures in the management report.
- Two folders were the same package. Two case folders held the same four files, byte for byte, so they were one package and not two.
- Names for the same title differ. Nine of 19 reference lots failed an exact title match because the registry and letters write titles differently. A normaliser for titles, lots and survey numbers is now shared.
- Registry status lags the council's record. Of three lots the council had accepted, the registry showed two as disposed and one as undisposed, with no field for the date or basis. Making status follow council outcomes is listed as a requirement.
- Appraisals expire. An appraisal stays valid for two years, with exceptions. Checked against the registry in October 2026, a large share of undisposed assets had a latest appraisal older than two years, and a smaller share had none. This is an upper bound, because an appraisal stays valid during an active disposition, but it showed that the validity check belongs at the front of the workflow.
- Coordinates are missing. No registry asset has stored coordinates, which limits what a map can show.
Where it could have gone wrong, and how it was handled
- Design from assumptions. The early build was based on how the process was described. Checking it against real letters and a real workbook is what exposed the package error.
- Demo data was clearly separated. Seed data was scaled up to hundreds of cases for training and reporting, and the training materials use only invented data. Real documents are kept out of version control and described in the notes only in aggregate.
- Open door by default. Without a shared password configured, login always fails rather than accepting anything.
- The last administrator. The last active administrator cannot be deactivated or demoted.
- Production settings. The notes require a long random secret key and secure cookies over HTTPS for any real deployment, and sessions end after a set idle time.
What was left open
- Sign-in through the agency's corporate identity system and finer roles for approvals.
- Moving submissions and uploaded files from a public cloud portal to the on-premises system.
- Documentary readiness checklists per stage, sales contracts, instalments, collections and title-transfer tracking.
- A live registry database instead of generated reports, with coordinates and a flag separating mass appraisals from property-specific ones.
- A parallel run with real cases, and independent security testing.
- A consolidated list of open questions, grouped by who can answer them: the office head, the IT unit and others.
Takeaways
- Compute the deadlines. Rules that depend on dates are better calculated than remembered.
- Test the design against real paperwork. The most important model change came from reading two real approval packages.
- Keep the registry as the shared key. Titles, lots and survey numbers needed one normaliser or nothing would join.
- Report the waiting time. The cycle time between steps was the question leadership would ask, and it was cheap to measure once dates were recorded per case.
- Say what is a prototype. Separating what is built, partly built and not built made the next decisions easier.
About these numbers
Test and report counts, dates and requirement changes come from the project's repository and design notes. Registry figures were read from real documents in October 2026, are rounded, and are aggregate only; no party, buyer or lot is identified here. The systems described are prototypes on invented data. They have not been security tested or used on live cases.