A dossier shows you exactly where a claim came from, graded for how much weight it can bear.
A claim with no citation does not ship.
That is the guarantee behind every Atlas dossier. A dossier does not tell you something is true and ask you to trust it: it tells you something is true and shows you exactly where that came from, graded for how much weight it can bear. If Atlas cannot show its work for a claim, the claim does not appear, or it is flagged as a gap instead of quietly smoothed over.
The evidence under a liquefaction judgment is largely shared and public; the judgments built on it diverge widely, and rarely show their working.
The evidence is shared. The judgments are not.
In an October 2025 study by the GARP Risk Institute, commissioned by the Climate Financial Risk Forum (convened by the Bank of England's Prudential Regulation Authority and the Financial Conduct Authority), thirteen physical-risk vendors scored one common portfolio of 100 real properties; pairwise correlations between vendors' flood damage-ratio estimates for identical properties ranged from 0.2 to 0.9.
Italy's three official seismic hazard models (MPS04, MPS19, ESHM20) diverge by more than 35% in predicted ground motion at specific locations, and formal testing against 50 years of records from 143 stations could not statistically distinguish which model is right (Iervolino et al., PLOS One).
PEER Report 2023/02 documents a 2022 workshop of 49 experts averaging 19 years of experience: the field's leaders do not share a common understanding of liquefaction susceptibility, and when asked to distribute 100 points of influence across 17 parameters, the single top parameter received anywhere from 0 to 80 points.
Atlas does not promise the agreement the field itself does not have. Its offer is a judgment you can re-run, ablate, and date: take one apart in the gallery of worked examples.
Liquefaction is a secondary effect of the earthquake primary peril, not a hazard that arrives on its own: it only happens where shaking meets ground already prone to it. That is the relationship every Atlas dossier is built to trace, evidence by graded evidence.
Eight guarantees, each pinned to one of these five stages.
Every stage is checked in code before a claim can ship.
A synthesized dossier is checked in code, not just by prompt, before it can ship. Any sentence that asserts a fact must attach a citation marker that resolves to a real record in that investigation's evidence ledger. A marker pointing at nothing, or an assertion with no marker at all, is a defect the checker catches and blocks on, not a matter of the model's discretion.
Every evidence record carries the source name, source URL, license, and retrieval timestamp alongside the content an agent is allowed to cite. A claim can always be walked back to where it actually came from, not to an intermediate paraphrase.
For the case-history corpus, every short verbatim quote is checked at ingestion time: an automated pass confirms the quoted text appears as an exact, whitespace-normalized substring of the source report before the record is ever added to the evidence index. A quote that cannot be matched this way is not committed.
Sources are graded A through F on a fixed reliability scale, set deterministically per connector rather than judged case by case, and that grade travels with the evidence into the dossier. A national survey and an index listing are never presented as equally weighty just because both produced a citation.
A separate reviewer pass reads the draft against its own evidence ledger, looking for the failure modes a confident-sounding draft can hide: conclusions the evidence doesn't actually support, ignored contradictions, weak inference dressed as certainty. Its findings drive a corrective retry. Whatever it still flags afterward goes into the dossier's own limitations section rather than a separate file, so the record shows what was challenged, not only what survived.
Which screening outcome a dossier reaches is decided by a frozen, founder-signed rule table applied deterministically to the evidence ledger, the same table for every run. The model is never asked to choose the outcome: its job is to narrate, in plain language, a decision the code already made. No hazard number is predicted by a model either; every numeric value comes from published models and measured data through cited connectors, not from a guess dressed as a figure.
A completed dossier is bound by an HMAC signature over its area of interest, its evidence ledger, and its content. A dossier that has been altered after signing fails verification rather than rendering silently. Export and display re-verify that signature before showing anything on GeoLiquefy letterhead.
A fixed battery of deterministic checks gates the prose before a dossier can ship: citation coverage, wording rules, and the narrate-not-choose contract with the rule table above, among them. When a draft still fails those checks after one corrective retry, the run is marked compliance-flagged and the violations are recorded alongside the output. A dossier is never quietly cleaned up to look compliant; it either passes for a stated reason or it carries a visible flag saying it did not.
An outcome is only asserted with confidence when the evidence actually supports it.
Thin evidence is stated, never rounded up.
An outcome is only ever asserted with confidence when the evidence actually supports it. When it does not, the dossier states plainly that the data is insufficient rather than rounding up to a conclusion. The same honesty drives the public evidence coverage map. Coverage is uneven by design, and we would rather show exactly where we are strong and say plainly where we are not.
Three limits that belong on the page, not in the fine print.
Atlas is a desk-study screening instrument, not a site investigation.
It reasons over evidence that connectors can retrieve remotely: published mapping, catalogued subsurface data, hazard models, case histories. It does not replace a boring, a CPT, or a site visit, and it is not a substitute for site-specific geotechnical engineering judgment.
A citation guarantees traceability, not correctness beyond its stated grade.
Citing a source means the claim can be walked back to that source and checked; it does not mean the source itself is error-free. That is exactly what the reliability grade is for: it tells you how much independent weight to put on a source, not just that one exists.
Coverage varies by region, sometimes sharply.
Atlas can only cite what its connectors can actually retrieve for a location, and that is not uniform across the world today. The coverage map states plainly where that evidence is strong, where it is partial, and where it is not yet there, rather than implying a global reach the connectors do not have.
Three limits Atlas states, not hides.
The model is never asked to choose the outcome, or to produce a number.
The model narrates. The rule table decides.