How MicroFire knows what it knows

The challenge asks for an interactive, AI-powered dashboard that summarizes, ranks and interprets microgravity combustion findings to deliver fire-safety insights for human space exploration. Each word maps to one place in the product. See all of it answer one mission question.

• ✓ FindSearch structured NASA evidence across experiment families. Evidence Atlas
• ✓ CompareSee conditions side by side: what matched, what differed, what was not reported. Compare Lab
• ✓ SummarizeRead verified NASA findings, quoted word for word with their pages. Top relevant NASA findings
• ✓ RankRank experiments and findings for mission conditions, with ranking robustness. Mission Analyst
• ✓ InterpretTurn an observation into a carefully bounded fire-safety implication. Evidence-backed insight
• ✓ AI-poweredEvidence is retrieved first; AI synthesizes; every factual claim is checked. Ask, with a verified example
• ✓ Evidence-bounded MLA validated model on NASA tests, gated so it abstains outside the conditions NASA tested. AI Model Lab
• ✓ Human space explorationNASA-studied exploration atmospheres, orbit, simulated lunar gravity and planned Moon-surface tests. Research Frontier
• ✓ UncertaintyEvidence Ladder, ranking robustness and named gaps show where the evidence stops. Methodology

The 56 test records are typed by hand from three tables in NASA's BASS-II Summary Report: Table 7.1 (SIBAL fabric, including four tests from the original BASS experiment), Table A.1 (PMMA films) and Table A.2 (Nomex). BASS-II Summary Report, Table 7.1, PDF p. 104

The transcriptions live in data/curated/*.csv with NASA's original units and wording. A Python build step converts units (cm to mm, µm to mm, atm to kPa), codes outcomes and writes the JSON this site reads. Unit tests check the conversions and spot-check values against the PDF.

The 2 LUCI burns, in lunar gravity simulated on a spinning New Shepard rocket, come from NASA's 2025 LUCI results. Every value is stored with the exact sentence and PDF page it came from, and the build fails if that sentence is not on that page or does not contain the value.

The 20 Saffire runs come from the test-matrix and results tables of three NASA Saffire reports. A build check finds every transcribed number on the exact table line it came from, and fails if one is missing. Each run links to its table and PDF page on the Saffire page.

The 62 quoted findings are checked automatically: the build fails if any quote does not appear word for word in its source's abstract or PDF text, and records the PDF page where it was found.

19 NASA documents are listed with their NTRS record, publication type and SHA-256 hash on the Sources page.

NASA's fire experiments burn different things in different physical regimes, so they are never merged into one table. Each record keeps its own columns and is also described along shared dimensions: fuel phase, material, size, gravity, oxygen, pressure and airflow. A mission question then sorts the evidence onto four rungs:

Direct

Same fuel phase, same material and same gravity, with every condition you set within tolerance.

Analogous

Solid-fuel tests that differ in named ways, listed on each card. Fewer differences rank higher; a gravity difference counts double, because it is a different physical regime.

Mechanistic

Other regimes, such as droplets or gas flames. They explain how flames behave, never how a material burns.

Gap

What no record covers, written as the matched-condition test that would fill it.

Tolerances: oxygen ±1.5 percentage points, pressure ±10 kPa, airflow ±50 %. A value a record does not state never counts as a match. These are MicroFire Atlas choices, not NASA criteria. Every ladder card has a “Why is this evidence shown?” path that walks through each check.

Worked example, computed live: A Moon base in exploration atmosphere A. Direct: 0 tests. Analogous: 78 tests, the closest being Saffire VI sample VI-3 (microgravity, not lunar gravity; 30.3 % O₂, not 34 %). Missing: oxygen. See this ladder

Each value on a test page carries one of four labels:

• Recorded for this test: a value in that test's row of NASA's table.
• Stated for the whole test series: a value NASA gives once for a group of tests, quoted with its page, such as the SIBAL sample thickness or the 99.2–99.5 kPa pressure range of one investigator's series. It is never presented as a per-test reading.
• Derived by MicroFire Atlas: our coding of NASA's words, such as an outcome code or the first numeric value of a flow column, with the rule shown.
• Not stated in the source: left blank. We do not fill gaps with typical values.

Airflow in Table A.1 mixes velocities in cm/s with fan potentiometer settings (“pot”). We keep the column word for word and only use the numbers that are velocities.

How close a NASA test is to the scenario you describe, from 0 to 100:

relevance = Σ wᵢ · simᵢ / Σ wᵢ      over the variables you set
coverage  = Σ wᵢ (reported by the test) / Σ wᵢ

• Numeric variables use sim = exp(−|Δ| / scale). The scale is the standard deviation of that variable across all 56 tests: oxygen 1.81 %, airflow 3.48 cm/s, width 4.6 mm, and log-thickness 0.43.
• If your airflow falls inside the range a test ramped through, the airflow match is 1.
• Pressure uses a fixed 10 kPa scale, because every test ran near 1 atm and the data's own spread is too narrow to be fair. A value inside a reported range matches fully.
• Material: 1 for the same material, 0.5 for the same broad class (PMMA: thermoplastic; SIBAL fabric: fabric; Nomex: fabric), otherwise 0. Flow direction and gravity: 1 if equal, otherwise 0.
• A variable the test does not report adds 0 but keeps its weight, so missing data lowers a score and is listed as “not reported”. It is never treated as a match.

Ties are broken by coverage, then by test ID, so the ranking is deterministic. A scenario outside the tested range of oxygen, airflow or pressure, or at a gravity level no test used, triggers the notice “Direct evidence under these exact conditions is limited”.

The weights, the pressure scale and the ladder tolerances above are our choices. A fair question is whether the answer depends on them. So every ranking is recomputed 1,000 times with those choices varied over a stated range, from a fixed random seed so the result is repeatable:

• each relevance weight multiplied by 0.75 to 1.25;
• the pressure scale between 5 and 15 kPa, and the similar-material credit between 0.25 and 0.75;
• for the Evidence Ladder: oxygen tolerance ±1–2 points, pressure ±5–15 kPa, airflow ±30–70 %.

For each test we report the median rank, the rank range (5th to 95th percentile) and how often it stays in the top three. “Stable rank” means a range of 3 places or fewer, “fairly stable” up to 7. Robustness is shown beside relevance and never merged into it: a stable rank means the test stays among the closest under many reasonable assumptions, not that a fire is likely or unlikely.

For A Moon base in exploration atmosphere A, the direct rung stays empty in 100 % of tolerance variations, and Saffire VI sample VI-3 stays the closest record in 100 %. Loosening our tolerances cannot turn orbit tests into lunar-gravity evidence.

Separate from relevance: how solid one test's record is, whatever your scenario. Seven equal-weight checks: test-level NASA row; outcome stated by NASA; oxygen reading not flagged by NASA; pressure documented; CO₂/CO readings present; fresh (not reused) sample; at least two related tests in the same family within ±1 % oxygen and ±2 cm/s.

Strong is 80 % or more of checks, Moderate 55 % or more, otherwise Limited.

The gap map bins tests by oxygen and airflow. A ramped-flow test counts in every airflow range it passed through. A cell with 3 or more tests is an observed region, 1–2 is sparse, and an empty cell is outside available evidence. The site makes no claim about empty cells and never interpolates across them.

Finding Relevance ranks verified NASA statements for a question; experiment relevance compares individual test conditions. Both are MicroFire heuristics. NASA did not create, validate or endorse this ranking.

Requested topics come from the question selector plus explicit triggers: airflow at ≤5 cm/s; oxygen above 21 % or below 19 %; quench below 19 %; pressure below 95 kPa; partial gravity when lunar or Martian gravity is selected. These thresholds organize retrieval; they are not combustion limits.

Relevance = 100 × (5 × topic recall + 3 × material match + 2 × gravity match) / requested-feature weights. Topic recall is the fraction of requested topics found in the curated tags. Exact material and gravity matches are 1, otherwise 0. Unrequested features leave the denominator; requested features with unknown metadata remain in it and contribute zero. A finding must match a topic or material to enter the list.

Evidence type constrains sorting first: direct, analogous, mechanistic, context; only then relevance descending, with finding ID as the stable tie-break. Direct requires an observed finding with explicit row links and all linked rows matching the selected conditions under the existing Evidence Ladder tolerances, without platform caveats. Publication-level solid-fuel findings remain analogous. Liquid/gas evidence stays mechanistic; planned FM² work and background remain context. One exception (ranking v1.1): when no material is named, several topics are requested and no observation addresses all of them, findings that address every requested topic sort first, still in evidence-type order and still labelled. A mission-context question such as “which exploration atmospheres has NASA studied?” then shows the atmosphere studies first. Whenever an observation addresses the whole question, observations stay first.

Material comes from explicit row links, material-specific sources, or literal material names in the quote. Gravity comes from linked records or a known experiment family. LUCI always retains its simulated-lunar-gravity limitation. These associations do not establish matched pressure, oxygen, geometry or flow history.

Only explicit curated record IDs count as supporting records; a shared family or publication never creates row links. Support count is not a replication count. Coverage is the average fraction of requested condition fields reported in linked rows, not the fraction matching. Publication-level coverage is unknown. Only curated abstract labels establish source role; an unspecified PDF section stays unclassified.

For example, weak-flow PMMA questions retrieve low-airflow findings. A lunar question also retrieves partial-gravity findings, with the simulated-platform limitation. A droplet result cannot outrank relevant solid-fuel evidence through topic count alone. A high relevance number can coexist with an analogous rung and unknown coverage.

Finding sensitivity uses 200 repeatable variations (seed 19), independently multiplying each weight by a uniform factor from 0.75 to 1.25. Evidence-type ordering stays fixed. Top-3 frequency is reported separately from relevance; it measures sensitivity to project weights, not measurement uncertainty or scientific confidence. Existing experiment sensitivity still uses 1,000 variations.

Fire-Safety Insight uses a small set of source-bound interpretation templates. NASA observations retain their exact quotes; MicroFire interpretations are labelled and preserve mismatches. Without a reviewed template, the panel abstains. This is neither mission certification nor a crew procedure.

Research planning supports spacecraft-fire researchers and early-stage mission evidence analysts before formal engineering assessment. It uses the current curated atlas; it does not determine whether a mission or material is safe.

Mission Scenario Registry

The finite registry reuses eight Mission Analyst presets and three additional Research Frontier questions. The duplicate Moon-base frontier and Challenge questions reuse existing presets. No Cartesian product is generated. Each question retains its origin, rationale, category and source context. Unspecified conditions remain unknown. Questions with identical normalized wording and conditions count once even if repeated under different IDs; distinct questions at identical conditions share one gap.

Gap normalization, deduplication and shared gaps

Normalize material, gravity, oxygen, pressure and airflow into a fixed key order; reject unsupported dimensions, invalid numbers and malformed values. The full canonical condition object determines the stable gap ID. Exact values define identity: 34 and 34.1 % remain different gaps even though they can be close under coverage tolerances. Missing pressure never equals a specified pressure.

Run the existing Evidence Ladder for each scenario against completed BASS, Saffire and LUCI records. A direct condition match prevents a coverage-gap entry; all platform caveats still apply. Findings and future FM² plans never become rows. Uncovered scenarios sharing exact conditions merge, retaining distinct scenario links and categories. The missing dimensions and next-experiment text come from the Ladder. A joint-condition gap remains even if each condition was tested separately.

Analogous support counts nonmatching rows with the requested material and at most two differing condition dimensions. It is a transparent proximity filter, not evidence strength. Links show the five closest rows even if more distant. Cross-family breadth uses those links and five selected non-context findings; mechanistic support remains a separate count of non-context liquid/gas findings.

Candidate generation and Evidence Gain

Each unresolved condition combination becomes a hypothetical research question. Only exact identical candidate specifications merge; no extra condition is borrowed from an analogue. Candidate objects have no measured outcome, NASA citation or experimental record family, and never enter the evidence corpus.

For every candidate–gap pair, reuse the Evidence Ladder's condition comparison: exact material/gravity, oxygen ±1.5 percentage points, pressure ±10 kPa, airflow ±max(1 cm/s, 50 % of requested flow). Missing requested candidate values fail. All requested dimensions matching means potential direct-condition coverage. Gravity or material mismatch excludes even partial coverage. Partial overlap requires a matched missing dimension (or two matched dimensions for a joint-combination gap); it remains a research lead, never counted as direct coverage or demonstrated gap reduction.

for each unique curated question:
  classify current completed records with Evidence Ladder
  if no direct condition match: group gap by EXACT conditions
for each unique gap condition object:
  create hypothetical candidate; keep unknown fields absent
  compare candidate conditions with every registered gap
  count matching gaps and UNION their distinct question IDs
potential direct questions = current direct questions + new unique matches

“Evidence Gain” means potential coverage if a valid completed record existed. It is not expected scientific information gain, entropy, Bayesian design, active learning or a prediction of experimental outcome. A broad question can be covered by a more specific candidate, but the reverse fails where required values are missing.

Planned NASA overlap and sorting

Verified FM² plan findings identify lunar-gravity and PMMA/SIBAL material overlap. This deliberately conservative mapping does not infer full atmosphere, flow or geometry compatibility. Plans count only as planned overlap and never increase observed coverage.

Board sorts expose separate dimensions: distinct questions, direct count ascending (then analogous support ascending), distinct categories, analogous support, planned overlap, or maximum candidate gap coverage. Descending sorts break ties by stable gap ID. Candidate order uses directly addressable gaps, then affected questions, then stable ID. There is no composite or official NASA priority score.

Limits and non-goals

Geometry, scale, duration, ignition method, confinement, orientation, hardware and sample history are not jointly represented. LUCI's lunar gravity is simulated. Even a direct match only covers represented dimensions under project tolerances. Counts depend on a small curated question set and do not measure researcher demand; repeated or selectively added scenarios can bias apparent breadth.

The planner does not estimate information entropy, outcome value, experiment cost, hardware feasibility, crew risk, TRL, schedule, program/funding/political priority, safety impact or probability of experiment success. NASA has not validated or endorsed these planning heuristics. Human usability impact study pending; no performance-improvement claim is made.

Open the research landscape. Versioned JSON exports retain corpus identity, assumptions and hypothetical/planned status. Source updates require human review before the evidence layer changes.

The AI Model Lab trains a regularised logistic regression on 41 BASS-II tests (SIBAL fabric and PMMA; reused samples, suspect oxygen readings and the all-negative Nomex family excluded). It predicts one thing: whether a flame was established after the ignition attempt. Final outcomes are not modelled, because most BASS-II flames were extinguished on purpose by turning the fan down. Saffire and LUCI are not pooled: different scale, different gravity regime.

Validation leaves one crew session out at a time, with bootstrap intervals, a repeated stratified check and a leave-one-material-out stress test. Five models are benchmarked against a baseline; the simplest one that beats it on proper scores is deployed.

Every query passes a domain gate first: gravity must be microgravity, the material must be in the training data, pressure must be ISS-cabin pressure, oxygen and airflow must sit inside each material's tested range, and at least three similar tests must lie nearby. Otherwise the answer is “out of domain” or “insufficient evidence”, and no number is shown. Reproduce it with npm run model-lab.

Ranking, the Evidence Ladder, comparison and the gap map are deterministic code. The Ask page uses a language model only to phrase answers from an evidence package of these records and verified quotes. Each claim in an answer is checked before it is shown: cited IDs must be in the package, numbers must appear in the cited record, and the page flags unit mix-ups, microgravity results described as lunar, causal or safety wording, and predictions. Without the model, every other page works unchanged.

The Verified Example on Ask is a saved synthesis rechecked against the current evidence and checker at build/test time; it makes no live request. The checker detects defined citation, numeric, unit and wording errors. It cannot prove full semantic entailment or real-world transfer.

Flame Vision is classical OpenCV segmentation, with no trained fire-prediction model. Not yet validated against hand-annotated real frames. Pixel measurements describe the image, not temperature or a calibrated physical flame size.

MicroFire-Eval v1.1 is 100 questions written before the measurements and then frozen: direct lookups, numbers, comparisons, synthesis across reports, mission scenarios, deliberately unanswerable questions and misleading premises. Gold answers are NASA record IDs taken from the data, not from the system. When the system failed a question, the system was changed, never the question. When the evidence itself changed (two LUCI lunar-gravity burns were added), the expectations that depend on it were updated, and each change is listed with its reason in the changelog. Read every question

Play: fool the checker

Before the model: retrieval and the claim checker (recomputed on every build)

The broken claims cover invented citations, wrong numbers, swapped units, microgravity results told as lunar, causal wording, predictions and uncited facts, ten of each. Still failing: “Can a flame jump a gap to another surface?” and “How long did a fabric sample take to burn at 5 cm/s?”.

Finding-ranking extension — provisional source-reading labels

Seven existing benchmark questions have finding-ID labels read from verified quotes, plus six supplemental adversarial questions. These are engineering labels awaiting independent scientific adjudication, not LLM-generated reference answers. They measure the finding ranker using the existing question parser; Mission Analyst uses explicit condition and topic controls.

Gold finding Recall@3 / Recall@5

64.3 % / 71.4 %

Mean reciprocal rank

0.568

Wrong-regime ordering violations

0 / 13 cases

Source-role checks

8 / 8

Adversarial finding-gap handling

6 / 6

Template-rule unsupported implication rate

0 % across 10 outputs

Injected unsafe/directive/probability claims rejected

30 / 30

Wrong-regime checks detect mechanistic findings promoted ahead of relevant solid evidence. Implication checks enforce source-bound templates and reject injected prohibited statements; a zero rate is not an independent semantic or scientific accuracy measurement. Abstention requires a flagged request limitation and no direct finding. Source-role checks include correctly leaving PDF sections unknown. Labels and per-case output are in eval/finding-labels-v1.json and node lib/eval-run.ts.

Finding gold set v2: 42 hand-labelled cases

Each case names a scenario and topics directly, so this measures the finding ranker itself, not the question parser. Labels were assigned by reading each finding's quote, kind and topics before running the ranker: a primary answer, other acceptable findings, and findings that must never rank above the primary. 8 cases are questions NASA evidence cannot answer (safest material, probabilities, proof, predictions).

Recall@1 / @3 / @5 (a primary answer in the top k)

47.1 % / 82.4 % / 85.3 % of 34

Recall@3 counting acceptable findings

91.2 %

Mean reciprocal rank

0.653

Forbidden findings ranked above the answer

0 / 34

Mechanism-only evidence above a solid-fuel answer

0 / 27

Context or planned work above an observation answer

0 / 27

Unanswerable questions flagged, with no insight attached

8 / 8

Unsupported implications

0 of 25

Evidence-rung agreement with the label

94.1 %

Mean top-3 stability of the answer (200 weight variations)

81.8 %

Still missed in the top 3: fr-003 (rank 4), fr-010 (rank 10), fr-013 (rank 8), fr-018 (rank 8), fr-019 (rank 10), fr-028 (rank 6). The first run of this set found that mission-context questions (which exploration atmospheres NASA studied, what FM² will use) buried the context findings that answer them below observations that matched only one topic. Ranking v1.1 lets findings that address every requested topic lead only when no observation does and no material is named; observations otherwise stay first. That lifted Recall@3 from 71 % to 82.4 % with every guardrail above at zero. Labels are engineering labels awaiting independent scientific adjudication. File: eval/finding-gold-v2.json.

Historical live-model run

All 100 questions were sent to the live Ask pipeline with gpt-5-mini (server default; OPENAI_MODEL was not set). 5 matched no evidence, so the model was never called. The 481 claims in the other answers were checked by the verifier at that time. These are saved historical metrics, not a new run or a rescore of this hardening pass. Re-scored: the saved answers re-checked after we fixed verifier false alarms (chemical formulas read as numbers, the question's own numbers, arithmetic in derived claims, negated safety words), with no new model calls.

Flagged claims stay visible with the reason shown; most remaining flags are rounded ranges (“about 21–22 %”) or unit conversions the checker cannot confirm. These checks measure faithfulness to the cited NASA text, not whether NASA's results transfer to a mission. Every answer is in results-live.json.

• 56 BASS tests (thin samples in a small duct, near 1 atm) and 20 Saffire runs (large samples, some at reduced pressure, 54 to 73 kPa) ran in microgravity. Two LUCI burns ran in lunar gravity simulated on a spinning rocket, in normal air. No test row comes from the Moon itself or from Martian gravity.
• Many flows ended at fan settings with no recorded velocity, so exact quench and blowoff speeds are often unknown.
• Outcome codes are our reading of short crew and ground notes.
• Spread rates appear in NASA figures, not tables, and are not transcribed.
• Flame Vision measures published NASA media in pixels. Without calibration, it cannot report physical flame size or speed.
• Nothing here is a fire-risk prediction or a NASA safety rating.