Aether Planetary Engine

Methane super-emitters and extreme heat are both quantification problems with a trust problem attached. A satellite sees a faint plume or a hot anomaly; turning that into a defensible number — a flux in tonnes per hour, a temperature anomaly in kelvin — means a chain of retrieval, calibration, wind, and baseline choices, each carrying uncertainty. The characteristic failure in this field is not the missing number, it is the over-confident one. Aether is built so a reader can follow any number on the screen back to the artifact and the assumption it came from, and so the dashboard says “cross-checked, not validated” where that is the honest ceiling.

Three real events across two phenomenon domains, end to end. Methane: two EMIT overpasses — the Goturdepe (Turkmenistan) and Permian (Carlsbad, NM) oil-and-gas plumes — detected with a per-column matched filter, quantified by integrated mass enhancement, and attributed against an open oil-and-gas infrastructure database. The matched-filter target spectrum is generated independently from HITRAN2020 line-by-line data; NASA's per-granule file is never read, and is kept only as a spectral-shape cross-check (r 0.993). Heat: the March–April 2022 northwest-India heat wave, with a 2 m air-temperature lane validated against ground stations and a separate land-surface-temperature lane that is never conflated with it. The pipeline runs on public data (EMIT, ERA5, MODIS); the dashboard serves committed artifacts only — no runtime fetching — and renders the results, and their caveats, verbatim.

The validation tier is earned by evidence, never asserted, and the failures ship next to the successes. For the India heat wave the pass/fail criteria were committed before the station data was read: the peak 2 m temperature (46.68 °C) and the regional anomaly (+5.67 K) passed against ground stations and the independent IMD product, and are tier VALIDATED — but the event's duration and extent FAILED their pre-registered criteria across two station datasets, and that criterion-edge fragility is reported as a finding, not hidden.

The methane flux is CROSS-CHECKED, not VALIDATED: it agrees spatially and in integrated mass with NASA's independent product, but absolute accuracy is unproven — and a +1.46× matched-filter amplitude systematic is reported and left uncorrected, because correcting it would forfeit the retrieval's independence, and it does not even transfer to the second scene. The Delhi daytime urban heat-island result is negative (−0.77 K — an urban cool island), carried as counter-evidence to the obvious prior. Event-level VALIDATED is deliberately held by no event. And the live demo proves this about itself: a committed verifier checks that the deployed service serves byte-identical committed results at the git commit it reports, and that non-redistributable raw data is provably absent — the dashboard cannot quietly drift from the repository.

Heat is reported in two physically distinct lanes the dashboard never merges: a 2 m air-temperature lane (validated against ground stations and the independent IMD product) and a land-surface (skin) temperature lane (capped at cross-checked — a late-morning Terra snapshot, never a daily maximum). Keeping them apart is the point: a reader cannot mistake a hot skin reading for an air-temperature claim, and the inspector shows the air-temperature lane before any land-surface number.

The heat wave's contributing factors are ranked from computed reanalysis diagnostics — presence and rarity, not a quantified causal share. The panel shows the synoptic ridge ranked first but capped, the popular dry-soil and urban priors argued against by the data, the negative daytime urban heat-island carried as explicit counter-evidence, and the published climate-change attribution result kept outside the engine's own scores as cited external evidence — never blended in.

Figures as of aether 91620d5, drawn from the repository's committed key-results snippet (docs/key_results.json) — each value traces to a Stage A/B artifact, not to this page.

This page is the case study; the working software is the live demo. They are different things — arkaneworks.co/ape is this write-up, aether.arkaneworks.co is the running dashboard — and the dashboard is where the numbers above are rendered with their full provenance and caveats.

The launch at the top of this page opens it in a new tab rather than embedding it, by design: the dashboard's API enforces a strict origin allowlist (the same hardening this case study describes), so framing it on another site would mean loosening the very guard that makes it trustworthy.

Aether was built with AI-assisted development under a human-architected, gated workflow. A human set the architecture, the sprint scope, and the acceptance criteria, and ran the gate reviews by reading the committed artifacts; AI coding tools did the implementation. One writer touched the repository at a time, and irreversible actions — deploys, pushes, anything outward-facing — required explicit human approval.

That discipline is the point, and it is the same discipline visible in the results above: every claim is required to trace to a committed artifact, validation tiers are not inflated (the methane flux is cross-checked, not validated; the pre-registered heat criteria where duration and extent honestly failed are reported, not hidden), and a deployed-integrity guard proves the live instance serves exactly the committed artifacts. The engineering maturity on display is not that every line was hand-typed — it is that the work was architected, reviewed, and held to a fabrication-resistant standard at every gate.

• Live demo: aether.arkaneworks.co
• Source & reproducibility: github.com/AnimeshRajvanshi/aether
• README (the dev-facing front door): README.md
• Scientific validation write-up (the rigor story in depth): sprint11_validation_writeup.md
• Thorpe et al. 2023, EMIT emission-source attribution: doi:10.1126/sciadv.adh2391
• Gordon et al. 2022, HITRAN2020: doi:10.1016/j.jqsrt.2021.107949
• Zachariah et al. 2023, India 2022 heat attribution (cited external): doi:10.1088/2752-5295/acf4b6
• OGIM v2.7 oil-&-gas infrastructure: doi:10.5281/zenodo.15103476