Climate & nature · evidence verification
Separating fact from fiction in climate and nature.
Evidence infrastructure for teams that need to know whether a claim stands up.
TerraSift traces a claim back to the primary science — assessment reports, peer-reviewed literature, agency monitoring data — and returns a verdict with page-level citations you can put in front of an editor, a board, or a regulator.
Evidence infrastructure · runs on your hardware, your private cloud, or ours in the EU
“The Atlantic overturning circulation will collapse before 2050.”
Why checks fail
Climate and nature claims break in three predictable ways.
A check that cites another check inherits every one of them.
Right number, wrong frame
The figure is quoted accurately — from a superseded assessment, or from a scenario no one now considers plausible. Nothing in the sentence is false, and the whole thing is wrong.
Confidence laundering
One preprint becomes “scientists warn”. Every caveat, confidence level and sample limit is stripped somewhere between the paper and the headline, and no downstream reader can see where.
Contested treated as settled
Live scientific disagreement — Antarctic sea ice, greening, carbon-removal permanence — gets flattened into a verdict in one direction or the other. Overstating the science damages trust exactly as much as denying it.
The method
How a claim is settled.
Five stages, in order. The same five whether the claim arrives from a newsroom at 4pm or from an assurance review with a three-week window.
Parse the claim
Separate the falsifiable assertion from the framing around it, and state plainly what would have to be true for it to hold.
Retrieve wide
Semantic search across the vetted corpus and live agency data — not a web search, and not a lookup of what other people concluded.
Rerank
A cross-encoder pushes the passages that genuinely address the claim above the ones that merely share its vocabulary.
Weigh, don’t count
Evidence is scored by tier. Ten aligned blog posts do not outweigh one assessment chapter, and volume never stands in for weight.
Answer with the trace
A verdict, its uncertainty, and every sentence tied to a page in a named source. Where the literature disagrees, the disagreement is described rather than resolved.
Then: the audit file
Query, corpus version, model configuration and retrieved passages are kept, so the check can be reproduced months later.
The evidence base
Weighed, not counted.
Every source in the corpus is placed in one of four tiers before it is ever retrieved. The tier travels with the passage, so you can always see what a verdict is resting on — and how far it is from bedrock.
Tier 4 exists to give context and locate the claim in public debate. It is never load-bearing on its own.
The platform
One system, from source to citation.
Not a wrapper around someone else’s model. TerraSift owns every stage between a source entering the corpus and a citation reaching your page.
Sources, tiered and versioned
Assessment reports, journals, agency feeds and your own documents, each placed in a tier before retrieval and pinned to a corpus version you can cite.
Search that reads, not matches
Dense semantic retrieval across the whole corpus, then cross-encoder reranking, so the passages that address the claim beat the ones that merely share its words.
The tier model, applied the same way every time
Weight comes from where evidence sits, not how often it repeats. Dissent is surfaced and described rather than averaged away.
An answer that shows its working
Passages, scores, page references, rejected candidates and model configuration are kept together, so any check can be reopened months later.
A desk to work in, an API to build on
A verification workspace for people who check claims all day, a REST API for your CMS or research pipeline, and exports your compliance team can file.
Wherever your evidence has to live
Your hardware, your private cloud, or our EU-hosted instance. The same system in each case — nothing is held back for the cloud edition.
What we offer
Six ways to buy the same rigour.
Start with a desk service; end with the method running inside your own organisation.
Rapid Check
A claim in, a cited verdict out, inside your deadline. Same-day turnaround on the beats we already hold a corpus for.
Corpus Build
We assemble, tier and maintain the source base for your beat — attribution, sea level, land use, biodiversity, carbon markets — and hand you the provenance record for it.
EvidenceGraph
The trace behind every verdict: which passages were retrieved, how they were weighted, what was rejected. Browsable, exportable, and reproducible after the fact.
NatureWatch
Standing watch on biodiversity and land-use claims, not just carbon: deforestation figures, restoration pledges, species status, offset integrity.
TrustMark
Documentation a compliance team can file: method statement, corpus version, model configuration and audit trail, aligned to EU AI Act transparency duties.
Academy
Training so your own desk can run the method without us — claim parsing, tier discipline, and how to write up a contested finding honestly.
Honest comparison
Why not just ask a chatbot?
For a first orientation, do. For anything you are going to publish, sign, or defend, the differences below are the whole job.
| General AI assistant | Fact-check aggregator | TerraSift | |
|---|---|---|---|
| Where the answer comes from | Whatever was in training data | Other organisations’ verdicts | A vetted, tiered corpus of primary science |
| Citation granularity | Often none, sometimes invented | A link to the check | Page-level, to the named source |
| Contested science | Picks a side, fluently | Inherits the aggregated verdict | Names the disagreement and both cases |
| Where your query goes | A third-party cloud | A third-party cloud | Your infrastructure, or ours in the EU |
| Reproducible six months later | No — the model has moved | Only if the check is still online | Yes — corpus and config are versioned |
Who we work with
Built for people who have to answer for the claim.
Newsroom verification desks
Check a claim against the science before it runs, and keep the trace in case it is challenged afterwards.
NGO policy and campaigns
Make sure your own numbers survive scrutiny — and know precisely where an opponent’s do not.
Sustainability and assurance teams
Substantiate environmental statements to the standard the Green Claims regime and your auditors expect.
Regulators and science bodies
Assess submitted evidence at volume, with a consistent weighting model and a record of how each conclusion was reached.
Sovereignty
Your sources and your questions stay yours.
What a desk is investigating is often more sensitive than what it publishes. TerraSift is built so that never leaves your control.
On-premise by default
The whole pipeline — retrieval, reranking, generation — runs inside your estate. No document, query or draft is sent to an external model provider.
An audit file for every check
Query, corpus version, retrieved passages, scores and model configuration are retained together, so a check can be reproduced and defended long after publication.
European, and documented
GDPR-compatible by construction, with transparency documentation prepared against EU AI Act obligations for the parts of the system that generate text.
Questions we get asked
Before you send us a claim.
Do you give true-or-false verdicts?
Only when the evidence supports one. Our verdicts are supported, contested, unsupported, or insufficient evidence — and the last two are not the same thing. A claim nobody has studied properly is not false; saying so would be its own error.
What if the science genuinely disagrees?
Then the answer says so, names the strongest published case on each side, and explains what evidence would settle it. We hold a dedicated corpus of contested questions for exactly this reason.
How fast is a check?
Minutes for a claim inside a corpus we already maintain. Days if the beat is new and the source base has to be built and tiered first. We will tell you which situation you are in before you commission anything.
Can we run it ourselves?
Yes. TerraSift installs inside your own infrastructure under licence, with corpus onboarding and configuration included. Teams typically start with the desk service and move in-house once the workflow is settled.
Is this an AI that decides what is true?
No. The model retrieves, ranks and drafts against sources it is given; it is not asked what it believes, and it cannot answer beyond the corpus. Every claim in the output is traceable to a page a person can read. The judgement stays with you.
Bring us a claim.
Send one your team is currently arguing about. We will run it through the full method and walk you through the trace — corpus, tiers, dissent and all — at no cost, so you can judge the work rather than the pitch.