Network data couldn’t load. Check your connection and try again.
Simulation connecting…
REAL NETWORK / SIMULATED OPERATIONS
WORLD 02 / METRO / London
AI decisions · simulated London
Jev is an AI by TypeSafe, running a simulated London Underground.
Disrupt your copy. See how it responds.
Think you can do better?
Challenge this exact moment.
EVIDENCE / LONDON
CURRENT RESULT / LONDON / SIMULATOR V6 / 22 SEPT 2026
+0.5 health points on average, and ahead in 48 of 48 matched tests. Each decision takes about 0.3 s and costs $0.0001.
Final health: Jev 86.5 · Forecast-only script 86.6 · Fixed rules 84.8 · No intervention 81.6
How it works. Every 3 seconds Jev gets the city’s state and a 15-second forecast for each option, then chooses. A script that always takes the best forecast scores the same, so the forecasts do most of the work; Jev’s part is reading them, and the situation, in plain language instead of hand-coded rules.
Precisely: Jev was slightly but consistently better than fixed rules: +0.51 points on average (95% interval +0.35 to +0.66). Against fixed rules: 48 wins · 0 ties · 0 losses across 48 matched conditions. Jev was consistently better than no intervention: +2.23 points on average (95% interval +1.99 to +2.46). No reliable difference between Jev and the forecast-only script (always takes the best 15-second forecast): +0.01 points on average (95% interval −0.02 to +0.04). Jev’s briefing includes the same projections; the baseline simply picks the best one.
The score. Average health is the network health index averaged over the whole 90-second run, the complement of the game’s damage score. It rewards a fast recovery, not just a good final second. Higher is better for the operator. It is a designed congestion index, not an official transit metric.
The ceiling. In 4 of 8 scenarios, Jev and the rulebook both score between 90 and 100: the city has little room to get healthier, so differences there are fractions of a point. Operators separate in the hard scenarios (Station closure, Blackout, Compound, Late pressure), which are highlighted below.
The interval. Differences are paired: each operator faces the same seed, arrivals and disruptions. The 95% interval treats each of the 6 seeds as one observation. An interval that crosses zero means no reliable difference.
Timing. Every operator gets the same decision opportunities; the earlier protocol’s extra final-second action for fixed rules has been removed.
Prompt. metro-choice-4 is live. On these same conditions it changed average health by +0.48 points versus metro-choice-2 (nyc-jev-5 options, no projections) (95% interval +0.31 to +0.64; 48 wins · 0 ties · 0 losses).
Scope. One model run per condition (48 conditions: 6 seeds × 8 scenarios). Trains, riders and disruptions are simulated on real TfL geography; this is not real-world transit performance. “nyc-jev-6” is the version name of the current simulator (simulator v6), used for both cities.
THE DIFFERENCE / PAIRED BY SEED AND SCENARIO
| Operator | Average health | Final health | Collapses | Reserves left |
|---|---|---|---|---|
| Jev | 88.2 | 86.5 | 6 | 3.3 |
| Fixed rules | 87.7 | 84.8 | 6 | 1.0 |
| Forecast-only script | 88.2 | 86.6 | 6 | 4.9 |
| No intervention | 86.0 | 81.6 | 6 | 9.0 |
EVERY SCENARIO / HARD ONES HIGHLIGHTED
In 4 of 8 scenarios, Jev and the rulebook both score between 90 and 100: the city has little room to get healthier, so differences there are fractions of a point. Operators separate in the hard scenarios (Station closure, Blackout, Compound, Late pressure), which are highlighted below.
| Scenario | Jev | Fixed rules | Forecast-only script | No intervention |
|---|---|---|---|---|
| Quiet | 100.0 (100.0) | 99.9 (100.0) | 100.0 (99.8) | 99.6 (98.5) |
| Signal | 91.7 (100.0) | 90.8 (99.8) | 91.7 (100.0) | 87.7 (98.7) |
| Oxford Circus crowd | 98.5 (100.0) | 98.4 (100.0) | 98.5 (99.8) | 97.7 (98.5) |
| Stratford crowd | 99.1 (100.0) | 99.0 (99.8) | 99.1 (100.0) | 98.6 (98.5) |
| Station closure · hard | 94.1 (98.8) | 93.4 (96.5) | 94.2 (99.0) | 92.5 (93.3) |
| Blackout · hard | 92.6 (100.0) | 91.9 (98.7) | 92.5 (100.0) | 90.9 (96.5) |
| Compound · hard | 43.5 (18.2) | 42.8 (17.0) | 43.5 (18.2) | 39.4 (10.7) |
| Late pressure · hard | 86.2 (75.2) | 85.4 (66.3) | 86.2 (76.2) | 81.6 (58.5) |
SPEED & COST / CURRENT EVALUATION
785 measured Jev responses from this evaluation. Typical and slow requests both count.
785 model requests · $0.079 estimated inference cost · 1,891,781 input tokens · jev-1.13.0. Inference only: hosting and failed-request usage are excluded. A fixed test sample, not a live service SLA.
LIMITS ARE PART OF THE RESULT
Jev: 6 collapses · Fixed rules: 6 collapses · Forecast-only script: 6 collapses · No intervention: 6 collapses in 48 conditions. 0 model request errors and 0 incomplete conditions recorded. A collapse means health stayed critically low; an API error is a separate failure.
| Scenario / seed | Jev | Fixed rules | Forecast-only script | No intervention |
|---|---|---|---|---|
| Compound · 1977 | 23s | 23s | 23s | 23s |
| Compound · 9896 | 23s | 23s | 23s | 23s |
| Compound · 17815 | 23s | 23s | 23s | 23s |
| Compound · 25734 | 23s | 23s | 23s | 23s |
| Compound · 33653 | 23s | 23s | 23s | 23s |
| Compound · 41572 | 23s | 23s | 23s | 23s |
LIVE SHARED CITY
Connecting to the shared London round…
OFFICIAL TFL / LIVE CONTEXT
Arrival predictions and line status refresh while this page is visible. They provide context and do not control simulated trains or demand.
Loading TfL arrival predictions…
Refreshes every 30 seconds. Stale predictions are withheld.
Connecting to TfL…
London Underground line boardings. Hammersmith & City and Circle are combined by TfL. Includes interchange boardings; not unique passengers or gate entries. Historical modelled reference, not live demand or simulator calibration. TfL NUMBAT combines ticketing observations with estimated route choices. The workbook was published 2026-07-01 and retrieved 2026-09-21. Major disruptions and exceptional events are excluded from its typical-day profiles.
SOURCES & LIMITS
TfL coordinates and ordered route sequences, retrieved 2026-09-21. Branches are included; reverse duplicate patterns are removed. Lines connect stations, not surveyed tunnels. Elizabeth line, DLR and Overground are not included.
Download the network ↗Demand, train capacities, frequencies, travel times and disruptions are designed simulation parameters. NUMBAT is a reference only. Health is a congestion index; confidence is not a recovery probability.
How the experiment works ↗Download current results ↗Arrivals use prediction timestamps. Line status shows when TfL was checked, because a separate current-status timestamp is not supplied. Missing feeds are shown as unavailable.
Powered by TfL Open Data ↗Map backdrop: simplified London borough boundaries and River Thames polygons from the Greater London Authority. Borough boundaries contain Ordnance Survey public sector information licensed under the Open Government Licence v3.0. Registered park boundaries: Crown Copyright 2025 and Ordnance Survey data, released under OGL, via GLA parks data. Major-road context: historical TfL Road Network via GLA. These layers are simplified and do not include every park or street. Some Tube stations lie outside Greater London. Download map geography ↗
Independent experiment. Not affiliated with Transport for London. Each city runs in its own shared world and has its own evaluation; histories and challenge links retain their city identity.
EARLIER STUDIES / KEPT FOR ACCOUNTABILITY
These studies were run before the current evaluation, mostly on the previous simulator, and informed today’s prompt. Their numbers are not the current result above and are not directly comparable with it.
FIRST LONDON EVALUATION / nyc-jev-5 / metro-choice-2 / 2026-09-21
| Jev | Fixed rules | No intervention | Jev vs rules |
|---|---|---|---|
| 96.2 | 96.3 | 90.2 | 7 better · 31 tied · 10 worse |
Paired mean difference −0.02 points. Measured before player disruptions were rebalanced; not comparable with the current result.
Download this study ↗CONTROLLER RESEARCH / NYC-JEV-5
We tested whether showing Jev approaching trains, blocked reserve approaches and recent passenger transfers improves its choices. The simulator and fixed rules stayed unchanged.
| Test set | Current Jev | More context | Fixed rules |
|---|---|---|---|
| Pilot · 24 conditions | 96.46 | 97.42 | 96.50 |
| Independent confirmation · 48 conditions | 96.42 | 96.38 | 96.38 |
Retain metro-choice-2. The pilot improvement did not reproduce in the independent confirmation set.
One model trial per condition and variant. Final health has limited headroom in these short runs. A richer context also increases input size.
Inspect both experiments ↗Jev is ready to decide when pressure builds. Start or return to a Jev-operated run to collect your own evidence.
Meet Jev, by TypeSafe AI ↗Latency includes the server’s TypeSafe round trip and any retries; it is not pure model compute time. Percentiles describe up to 1,200 recent valid responses. Cost uses documented pricing of $0.042 per million input tokens, verified September 21, 2026. Failed-request usage is unavailable; this estimate is not an invoice.
SAME CITY / THREE OPERATORS
Six matched scenarios: two seeds, three disruption schedules, 90 simulated seconds each. Jev, rules, and no intervention receive the same arrivals and events. Jev and rules choose from the same valid actions every three ticks.
Run real model evaluations and inspect survival, health, latency, and token usage. No invented results.
A small simulator diagnostic, not real-world transit validation. A 90-second survival result is capped by the test window. Jev may tie or lose; all completed scenarios are shown. No API calls are made just to display previously recorded results.
DECISION LOG / CURRENT RUN
0 of 0 matured decisions were followed by higher network health 12 ticks later. This is observational: later decisions and incident recovery also influence the result. It is not causal attribution or confidence calibration.
Choose Jev as operator, disrupt a station, then return to inspect its response.
The latest eight records are shown; downloads include up to 500 recent detailed records. Replays store all applied choices. Device-local records and share links are not independently attested benchmark submissions.