EVIDENCE RETRIEVAL / NASA SPACE APPS
MicroFire Atlas — NASA Microgravity Fire Evidence, With Its Limits Shown
Ask what NASA knows about a material under a future lunar-habitat atmosphere. It finds the records across BASS, Saffire and LUCI, compares conditions, ranks experiments and verified findings separately, checks every AI claim against the evidence, and names the combination nobody has tested. It never gives a safety rating or a probability.
Public demo · Team project — NASA Space Apps Challenge 2026 (Flame in Freefall)
9 — computed stages from question to research gap
The problem
NASA has studied fire in microgravity for decades, but the results are spread across test tables, figures and reports. A crew-safety researcher cannot quickly ask what NASA actually observed for this material and this cabin atmosphere — or how far that evidence reaches.
The approach
Challenge Mode answers one question in nine computed stages: find the evidence families and the role each can play, compare the closest test records condition by condition, quote NASA findings word for word with page numbers, rank experiments and findings separately with their stability across weight variations, interpret one bounded implication from a reviewed template, retrieve evidence before any generation and check each AI claim, then name the untested condition combination as a research question. Asked for a probability, it returns the closest evidence and the mismatches instead of a number.
The results
Live site with a challenge mode, an explorer mode for ages 8–13 and five versioned data exports carrying corpus hashes and limitations. The repository states that no human impact study has been run and no performance improvement is claimed.
Stack: Next.js, React, TypeScript, three.js, Anthropic API, Node test runner, Vercel