This page is a real Arasense run, not a brochure: climate models scored against observed climate near Sydney, the skilful ones trust-weighted into a mid-century signal for the city’s worst rain day, with model agreement reported, because a projection without it is just a guess with confidence.
The trust-weighted screening signal for Sydney is -2.3% change in maximum 1-day precipitation by 2040–2059 under SSP2-4.5: 88.96 mm → 86.88 mm on the baseline worst rain day. Only 45% of trusted-model weight backs this direction. The number is reported with its doubt attached: a signal to deepen, not to quote alone.
Sydney is part of the atlas's global preview tier: the same pipeline as the ranked European cities, published unranked while coverage grows. One thing changes outside Europe: Long-record ground observations strongly constrain the ERA5-Land baseline here, so the trust scores rest on solid reference data.
Each climate model is compared against observed climate behaviour near Sydney and decomposed with the Aras Diagram: bias, variability, and phase alignment, not one opaque score. Models that fail screening are rejected, and you see why.
Skill-weighted models estimate the mid-century change in Sydney’s maximum 1-day rainfall, with the spread and the 55% directional agreement reported, never hidden.
Atlas signals are screening-level evidence. Where a decision depends on Sydney, the same pipeline runs the full CMIP6 ensemble: every trusted and rejected model named, weighted, and auditable.
The brief behind this page: the screening run in full, what a 34-model deep dive would add for Sydney, and how the evidence holds up in front of a technical reviewer. Prepared per request from the same pipeline; tell us where to send it.
Every climate model in the screening ensemble is scored against roughly 20 years of observed climate behaviour near Sydney (ERA5-Land) using the Aras Diagram, a peer-reviewed skill-decomposition method (Izzaddin et al. 2024). Models that track local reality earn weight in the projection; models that fail screening are rejected. The weights and rejections are reported, never hidden.
Trust scores are only as good as the observations behind them. The flag is a qualitative tier for how densely long-record ground observations constrain the ERA5-Land baseline in each region. For Sydney it reads “strong”. Long-record ground observations strongly constrain the ERA5-Land baseline here, so the trust scores rest on solid reference data. Publishing the flag, instead of pretending every baseline is equally solid, is part of the method.
As screening-level evidence, yes, always alongside its 55% model-agreement figure. It is built for prioritisation and planning conversations. For site-specific or engineering-grade decisions, the next step is a full-ensemble deep dive for Sydney, which names every trusted and rejected model with its weight.
Averaging treats a model that mistracks Sydney’s observed climate the same as one that nails it, and it buries how much credible models disagree. Trust-weighting measures each model’s local skill first, then lets the skilful ones carry the projection, and reports the remaining disagreement as part of the answer.
Maximum 1-day precipitation (rx1day) within a 50 km radius of the city, baseline 1995–2014 versus future 2040–2059, scenario SSP2-4.5, using the 5-model trust-screening ensemble shared by every Trust Atlas city. Bologna additionally has a 34-model deep dive, which is why its two published numbers differ, on purpose.
Every trust profile publishes its signal as JSON, so risk platforms, notebooks, and reviewers can pull the number with its uncertainty attached:
curl https://www.arasense.com/trust/sydney.json
Arasense