TL;DR. After Urda, a four-hour Renfe map went viral beside 34-year-old signalling. We have reached an inflection point where a finished-looking demo can be believed as a real product, even next to safety-critical infrastructure.
Contents
- What "legacy" means on a high-speed line
- What Vía Libre actually did
- The inflection point vibecoding just hit
- What the criticism actually said
- Human error or system error, through a legacy lens
- What juniors should take away
- Sources
On 28 October 2026, a Renfe AVE collided with a stationary Iryo train near Urda, in Toledo, Spain. Sixty-nine people lost their lives. My thoughts are with their families and with everyone affected. The investigation is still open, and nothing written here treats any cause as settled.
Two days later, a different story went viral. Álvaro Hernández published Vía Libre, a web map of Renfe trains built in about four hours from public GPS data. Colours marked delays. Screenshots travelled faster than the nuance.
Those two stories sit on the same week for a reason juniors keep hitting at work: legacy systems look slow and ugly until you learn why they are hard to change, and AI has made it so easy to ship something that looks finished that people can believe a demo can stand in for a real product.
This piece uses Urda and Vía Libre to hold both ideas at once, with Vía Libre as the heaviest example of that belief so far.
What "legacy" means on a high-speed line
The Madrid-Seville AVE corridor still runs under LZB (Linienzugbeeinflussung), a German continuous train-control system installed for the line's opening in 1992. That is 34-year-old railway software still sitting in the safety path.
In plain words, the track and the train keep talking so the system can supervise speed and brake automatically if needed. When LZB drops, traffic can fall back to BSL (bloqueo de señalización lateral: the side signals that organise blocks of track) with ASFA as onboard protection that reacts when the train passes beacons. That stack is old, layered, and certified for lives, not for demos.
Press coverage after Urda repeated the same timeline: LZB arrived because Europe's later standard, ERTMS, was not ready in 1992; Adif has been migrating Madrid-Seville to ERTMS Level 2; the physical work finished, but putting the new system into service still needs tests and authorisation from the rail safety agency.
A railway signalling engineer told La Tribuna de Ciudad Real that LZB gear is robust and near the end of a useful life of about three decades. The former head of the State Agency for Railway Safety had already called LZB obsolete in the Senate weeks before the crash. Replacement is underway. It is not a git merge.
On top of that world sits a thinner layer: Renfe's open data, including near-real-time vehicle positions in GTFS-RT, updated every 15 seconds for AVE and long distance (20 for Cercanías), under a CC BY 4.0 licence. That feed is for passengers and builders. It is not the signalling stack. Confusing the two is how vibecoding stories start.
| Layer | What it is | What it is not |
|---|---|---|
Signalling (LZB / ERTMS / ASFA) |
Safety path that supervises speed, routes, and braking | A passenger map |
Open data (GTFS-RT) |
Public positions for builders and riders | A control desk or safety case |
What Vía Libre actually did
Álvaro said he wondered how long a Renfe monitor would take, noticed that each train's GPS position was public, and had something working in three to four hours. Vía Libre plots Renfe services on a map of Spain with a delay semaphore and Adif incident notices on one screen.
Screenshot of Vía Libre. Source: El Español / Omicrono, 30/10/2026.
Mobile view of Vía Libre shared on X, 30/10/2026. Post: x.com status 2116271528396677542.
That is a real win for open data and for fast prototyping. Anyone can see how far a public feed plus a clear UI can go in a morning.
It is also where the new risk starts: a browser reading GTFS-RT does not supervise braking curves, does not interlock routes, and does not carry a safety case, yet it can look finished enough that people treat it as more.
Screenshot of Álvaro Hernández's original thread, reshared on X, 31/10/2026. Post: x.com status 2116679379468120546.
The inflection point vibecoding just hit
Vibecoding means asking an AI for code and shipping what comes back while it "feels" right, without owning the failure modes. Readers already know a four-hour map cannot retire LZB or finish ERTMS. That part is settled.
What is new is the inflection. AI has made it cheap enough to ship something that looks finished that a builder, or an audience watching screenshots travel, can honestly believe a demo-grade build can stand in for a real product. Not as a joke: as something that could run in a real business, or sit next to a system people already trust.
Most vibecoding stories stop at a SaaS landing page or an internal tool. Vía Libre is the heaviest case yet, because the thing it seemed to rival, in the public conversation after Urda, was safety-critical railway infrastructure.
That belief is the new risk. A map that paints delays in green, yellow, and red looks like monitoring. Screenshots travel like a product launch. The speed of the build can read as proof that the old stack was optional. None of that turns an open-data viewer into a control system, but the look of finish is strong enough that people can confuse the layers without trying.
Why legacy stays slow
Legacy railway systems stay slow to change for boring, expensive reasons:
- Safety certification. Every real change has to be re-tested and signed off before passengers ride on it. Think of a release process where "ship" means an outside auditor stamps the build, not a green CI job.
- Interlocking. Trackside logic that stops two trains from claiming the same block of track at once. Like a lock or mutex on a shared resource, except the resource is a stretch of rail and the failure mode is a collision.
- Layers added after earlier incidents. Extra checks and fallbacks piled on over decades. Like patches nobody wants to delete, because nobody is sure which one still holds a crash path shut.
-
Ageing kit and scarce spares. Gear from the early 1990s eventually leaves the vendor catalogue. Replacing a board is not
npm update; it can mean hunting a part that is no longer made. - Testing on a live corridor. A lot of the real proving has to happen on the actual line, often at night when passenger trains are not running. You cannot spin up a full high-speed clone in staging.
-
AESF authorisation. Spain's State Agency for Railway Safety (
AESF) has to authorise bringing a new or changed structural subsystem, including signalling, into service. Until that file is approved, the new stack does not carry passengers, no matter how ready it looks on paper.
That cost does not disappear because a morning build looks polished.
Álvaro drew the line himself. As quoted by El Español and La Razón, he did not want Spanish trains controlled by a "Paco-APP" built in a morning. His point was about political will to use what already exists for safety, not about swapping signalling for a weekend build.
Fair enough. The case still matters as evidence of the inflection: if even a careful builder's demo can travel as if it answered a railway, then the bar for "looks ready" has moved.
A four-hour map like this can still prove the feed is readable, shame bad UX, and start a useful argument about transparency. Keep those. What changed is the gap: "looks finished" and "is a product" used to be obviously far apart. Vibecoding shrank that gap enough that the belief is now part of what you have to manage.
What the criticism actually said
The pushback around Vía Libre clustered into a few themes. No need to pile onto individuals.
I do not question Álvaro's good faith in trying to help or put another solution on the table. Most of the criticism came from the sense that he had underestimated the infrastructure. My own take is milder: I think he simply had not informed himself well enough before framing it that way.
- Already existed. People said the map was not new: Renfe already had an official real-time viewer, and other private trackers existed. Putting open data on a map is not inventing the railway.
- Front end, not signalling. Others said it was mostly a nicer front end on a feed Renfe already publishes: useful to look at, not a monitoring or signalling product. A passenger map is not a control desk.
- Read the numbers. A third theme was about reading the numbers. Public feeds can lag or fail. Some trains are estimated from the timetable. Press pieces noted that "real time" on the site animates between updates, that operators such as Iryo and Ouigo are missing from that feed, and that the famous 93-of-119 delay count was one snapshot on a drivers' strike day, not an official punctuality score for the whole network.
There was praise too. People shared the site as a clear way to see delays, and some welcomed an independent view while still saying it should not replace the operators.
Official Renfe viewer (tiempo-real.largorecorrido.renfe.com). Source: Railway Gazette, 12/03/2026, Renfe screenshot.
Human error or system error, through a legacy lens
After Urda, public debate often split into "human error" versus "system error." Through a legacy lens, that split is too thin.
What official sources and the press have established so far:
- Around 08:30 the
LZBfailed on that stretch. - The control centre authorised the AVE to run under "marcha normal" instead of "marcha a la vista", which caps speed at 40 km/h when you are running with degraded protection.
- The preliminary report from
AITAT, the body that investigates accidents, talks about a "chain of failures, instructions and actions" without the needed safety barriers kicking in. - The technical and judicial investigations are still open. Every hypothesis still is.
There is a human decision in that chain. That it is the cause is not proven. Investigators also look at maintenance, protocols, and why backup layers did not stop the train.
That is how you read a decades-old stack: a person acts inside procedures, fallbacks, and machines that age. Naming one person without the chain is the same mistake as celebrating one map without the stack.
In software we call the careful version a blameless postmortem: after an incident, ask which layer was missing, not only who typed the command. People fail. Good systems assume that and put nets underneath.
Legacy railways already try to do that with continuous supervision, side-signal fallback, and automatic braking. When those nets tear, the lesson is to repair and replace the nets with the same seriousness that certified them, not to pretend a dashboard could have held them.
What juniors should take away
If you are early in your career, this week is easy to misread. Either "one app exposed the state" or "one person's mistake caused the crash." Neither is enough.
- Treat open data as a gift with limits. Credit the licence. Show how old each value is. Decide what the UI does when the feed dies.
- Respect legacy before you rewrite it. Ask what certification, interlocking, or incident history made it look "ugly."
- Use AI to learn and prototype. Notice when a finished-looking demo starts to feel like a product, to you or to your audience.
- When something breaks, look for the chain: inputs, fallbacks, monitoring, human procedures, and the layer that should have caught the mistake.
- Say out loud what your weekend build does not do. That sentence is part of the product, especially when the screenshots travel faster than the caveat.
Vía Libre shows how fast a readable view can appear when the data is public, and how easy it is now for that view to travel as if it answered more than it does. Urda shows why the systems underneath move on a different clock.
The useful junior skill is not dismissing legacy, and not treating vibecoding as a shortcut past it. It is learning to read both, and to name the gap between a finished-looking demo and a product that can carry real trust.
Sources
Full source list
Top comments (1)
Thank you for handling this carefully and not treating any cause as settled. The point for juniors lands: a four-hour map reads public data, while signalling has to be certified to never be wrong. Both can be good work, they just answer completely different questions, and a finished-looking screen hides that difference.