Joint Watch Daily Broken-Rail Risk for Main Line Railways
VERTICAL-DRIVEN ARCHITECTURES · RAIL TRANSPORT
Joint Watch
A system design for the chief engineer accountable for track safety, divisional executive engineers, permanent way inspectors and the engineers who would build it. It shows how a national railway with aging track can rank every kilometer post for failure risk each day from its own records, and leave every decision with the officer the law holds accountable.
THE CASE IN BRIEF
- A railway running overage track stays safe only while someone inspects it, and a joint can fail between patrols.
- The evidence of deterioration exists, but it is spread across paper patrol books, ultrasonic charts, speed restriction registers, welding records and inquiry files held in different divisions and formats.
- Ultrasonic inspection covers only part of the network each year, and no record joins a weld, a patrol observation, an ultrasonic indication and a restriction on the same kilometer post.
- As a result, no one sees the same kilometer post twice, and deterioration is found after the failure.
- Joint Watch reconciles every record to kilometer post identity in one model the railway owns and issues a ranked daily watch list with its evidence attached.
- The system is advisory by design: it ranks and evidences, the named officer decides, and nothing connects to signalling or train control.
WHY THIS MATTERS
Failure happens between inspections. Visual patrols see the surface and ultrasonic runs see inside the rail, but neither sees what a joint does between visits.
Records exist but never meet. A weld's history, a patrol note, an ultrasonic indication and a restriction in force describe the same place, yet no system holds them together.
Alerts without capacity become liabilities. When detections exceed what patrol and ultrasonic crews can work, defects sit on record unaddressed, so the alert budget has to match real repair capacity.
And public sector data must stay in the country. Data residency rules and export controls on the largest accelerators decide what hardware and models the railway can hold and run.
This creates a structural exposure.
- Patrol books are paper records that no system computes on.
- Ultrasonic charts are held apart from the patrols and restrictions on the same stretch of track.
- Speed restrictions and welding records sit in separate registers in different divisions.
- Inquiry findings are not linked back to the kilometer posts they concern.
- Long sections lose connectivity, so field data arrives late or not at all.
- Findings risk landing in a queue with no obligation to act.
A joint should fail on the watch list, not under a train.
Every kilometer post carries its history and its risk, and every daily decision stays with a named officer.
WHAT YOU'LL LEARN
Inside the architecture, you'll find:
- Why track breaks between patrols, and what answering the daily question requires
- How nine records each hold one slice of the track, and why the failing joint is visible only in the joins between them
- The four constraints and the statute that shape the design: data residency, named decision makers, intermittent long sections and staffing capacity
- Three scoping decisions and what each costs, from an advisory-only boundary to reusing the estate in hand
- What the design chose against for model serving, detection, forecasting, video ingest and positioning
- The layered stack from nine sources through one adapter family to four surfaces, including an offline patrol app
- The fourteen-object model, with the kilometer post shown in its recorded form
- How paper patrol books are digitized with the inspector who wrote them confirming every entry
- How store-and-forward buffering and time synchronization keep field records ordered across dark sections
- Why forecasting runs on divisional hardware and frontier reasoning on the railway's own GPUs in-country
- How export controls shape the hardware plan, and the memory arithmetic that fits the frontier model on available accelerators
- The three models and why their licenses let the railway hold every weight
- How the watch list runs in shadow until a flagged joint is found defective by an inspector before a train finds it
- The six named failure modes, lessons the design encodes and the questions still open
- How Praxis contextualized the requirement, read it through eight lenses and reasoned every choice in the design
DOWNLOAD THE ARCHITECTURE
Read the complete architecture, from the nine records that each hold one slice of the track, through the object model, ingestion, inference placement and model register, to a three-phase rollout proven in shadow and the reasoning Praxis recorded for every choice.
A railway that can see its weakest kilometer posts each day can send its inspectors where the risk is highest. Keeping that record and those weights in its own hands means the evidence stays with the officers accountable for the track.
Contributors

CodeNinja Engineering Team
Designed with Praxis