Inspiration
Environmental debates often contain a real observation followed by a conclusion that is much larger than the evidence. Earth Evidence was built to teach a disciplined question: does the evidence support the environmental claim, contradict it, or remain insufficient—and what evidence should come next?
What it does
A learner enters an environmental claim and a piece of evidence. Earth Evidence classifies the relationship as:
- SUPPORTED
- CONTRADICTED
- INSUFFICIENT
It then identifies a broad environmental category such as Climate, Energy, Water, Waste, Biodiversity, or Transport and produces a Next evidence needed prompt. The event-window V2 update also adds an Evidence Context Coverage panel that checks whether the supplied evidence contains six kinds of decision context: baseline, comparison, time window, spatial scale, measured outcome, and source/provenance. Each field is reported as PRESENT, MISSING, or explicit NOT_APPLICABLE.
MISSING is intentionally conservative: it means that context was not detected in the supplied evidence text, not that the fact is false or absent in the real world.
NextStep / Earth Forward development disclosure
This entry continues work from the earlier ProofPath project. It is not represented as wholly new from zero.
Before NextStep
The inherited ProofPath baseline already contained:
- generic claim/evidence input UI;
- three-class SUPPORTED / CONTRADICTED / INSUFFICIENT classifier;
- browser-side softmax/logistic-regression engine;
- generic training and holdout examples;
- generic educational evidence-assessment claim ceiling.
Baseline repository commit: 537e276724a75b9f5f45404c550712b84f56d046.
Developed during NextStep
On branch nextstep-earth-evidence-v1, the NextStep build adds:
- a new 32-pair Earth-specific training corpus;
- a separate 10-pair Earth holdout/demo set;
- environmental category detection;
- the new Next evidence needed guidance layer;
- an Earth Forward-specific interface and examples;
- Earth-specific benchmark and UI tests;
- explicit before/during documentation.
Event-window delta V2
During the still-open NextStep submission window, the project adds:
- deterministic Evidence Context Coverage for baseline, comparison, time window, spatial scale, measured outcome, and source/provenance;
- PRESENT / MISSING / explicit NOT_APPLICABLE output per field;
- a visible coverage panel in the interface;
- a negative test where a context-poor evidence statement must remain MISSING on the absent fields;
- a guard test proving that a bare source/year citation is not treated as a measured outcome;
- positive and explicit-NOT_APPLICABLE coverage tests;
- preservation of the original fixed 10/10 Earth holdout result.
The classifier itself and the 10-pair holdout set were not changed by this V2 delta.
The changes are isolated on a separate branch rather than modifying the four already-submitted ProofPath entries.
What we learned
The biggest learning step was that adapting an evidence-reasoning model to environmental questions is not just a matter of changing examples. Environmental claims often depend on missing context such as baselines, comparison groups, time windows, spatial scale, measured outcomes, source provenance, or independent replication. That led us to build the Earth-specific corpus, category detection, Next evidence needed guidance, and the new context-coverage panel rather than simply re-skinning the earlier ProofPath interface.
We also learned that a strong benchmark result can still be narrow. The separate 10-pair Earth holdout currently passes 10/10, but we keep that result explicitly bounded to a small curated implementation benchmark instead of treating it as evidence of general-world accuracy.
How it works
Earth Evidence runs entirely in the browser using JavaScript, HTML, and CSS. It reuses the compact three-class softmax/logistic-regression engine from ProofPath and trains it against the new Earth-specific examples. The model uses interpretable text-pair features including lexical overlap, negation mismatch, numeric conflict, antonym conflict, contradiction cues, uncertainty cues, overclaim cues, and evidence/claim length relationships.
The V2 coverage panel is separate from the classifier. It uses deterministic text cues to report whether six contextual evidence fields are detectable in the evidence supplied by the learner. It does not infer real-world truth from a missing cue.
No external AI API, paid model, account, CDN, model download, or cloud inference is required.
Verification
The fixed untouched Earth holdout remains 10/10. The V2 coverage tests include a context-poor negative canary, explicit NOT_APPLICABLE handling, a fully populated positive case, and a guard against treating a bare source/year citation as a measured outcome. Static UI checks also require the Evidence Context Coverage panel. These are bounded implementation tests, not evidence of general-world accuracy.
Live demo
https://turbo-light-fmf7q67.shipstatic.com
Demo / pitch video
4:47 demo video (public GitHub Release MP4): https://github.com/lbkapcsolat-dotcom/ProofPath/releases/download/earth-evidence-nextstep-demo-v1/Earth_Evidence_NextStep_Demo_V1.mp4
Release page and SHA-256 asset: https://github.com/lbkapcsolat-dotcom/ProofPath/releases/tag/earth-evidence-nextstep-demo-v1
Exact public MP4 SHA-256 reported by GitHub:
1668ccb54c9d31824b077d5a8ca7d4461e9d84e27d6cfce28023cc8b4f6944e6
The dedicated Devpost video field still requires a public YouTube or Vimeo URL; the GitHub MP4 remains linked here until that platform-specific bind is available.
Source
NextStep branch: https://github.com/lbkapcsolat-dotcom/ProofPath/tree/nextstep-earth-evidence-v1
Before/during disclosure: https://github.com/lbkapcsolat-dotcom/ProofPath/blob/nextstep-earth-evidence-v1/NEXTSTEP_CHANGELOG.md
Claim ceiling
EDUCATIONAL_ENVIRONMENTAL_EVIDENCE_ASSESSMENT_ONLY
Earth Evidence is not a climate model, scientific validator, policy recommendation engine, or automatic fact checker. Model scores are not calibrated scientific confidence.
Built With
- css
- html
- javascript
- machine-learning
- softmax-regression