Requests for work

Open Observatory is built by people who donate their coding agent's attention for a session. This page sets out the goal, how far the project has got, and the work, research and purchases that would move it forward. Every item is written so an agent can pick it up with no other context.

The goal

Full monitoring of data-centre utilisation, at the highest time resolution public data allows. For every data centre in the world: where it is, when each building goes up, when it starts running, how much of its capacity is in use, and, ideally, whether that use is training or inference.

China is the priority. Public data there, official figures above all, is notoriously unreliable and cannot be checked from outside, so independent measurement matters most. It is also where the project sees least today.

What utilisation means here. The share of installed computing capacity in use. Public data reaches it only indirectly:

• Commercial utilisation is the share of built capacity that customers have leased, as colocation operators report it. It says nothing about how hard the equipment runs.
• Electrical utilisation is average power drawn divided by capacity. It needs a load and a capacity for the same period.
• Activity signals such as construction, energisation and on-site fuel burning bracket utilisation without measuring it.

Training or inference. The planned method is a deep-learning model trained on detailed thermal imagery of campus substations and transformers, with labels from AI labs such as OpenAI, which know when and where their training runs happened (T-05). A transformer's losses rise with the square of its load, so its temperature follows the site's load with a lag of hours, and the same imagery can count transformers to give capacity. Separating workloads assumes, untested here, that inference follows daily demand cycles while training runs near-flat for weeks, so it needs several images a day at a few metres.

How close we are

Load without a same-period capacity does not give utilisation. Meta reports electricity for 18 campuses and Google's water use gives an approximate load for 12, but among them only Luleå in Sweden has a capacity figure for the same years: it ran at 0.25 to 0.45 of its 120 MW supply in 2022 to 2024.

The biggest gaps

• China has no utilisation measurement. Operators publish commercial rates, per data centre only in Chindata's filings up to 2022 and mostly without a city. Construction dating is the only independent signal.
• Capacity is the weak half of utilisation. Two sites have a load and a capacity for the same period, and most capacity figures are third-party estimates or connection limits rather than installed computing.
• Nothing is measured often enough to separate training from inference. The finest load series is monthly, at one campus that burns its own fuel. The planned transformer method needs thermal imagery at a few metres several times a day, which no free source provides, plus training-run records from AI labs. Hourly NO₂ from geostationary satellites over North America (TEMPO) and East Asia (GEMS) is an untested alternative where fuel is burned.

Donate a session

1. Read "Results so far" at the end of this page, so you do not repeat a method that has already failed.
2. Pick one open item that fits your time and credentials, from Priority 1 if you can. On the website, claimed items are marked live; otherwise search open pull requests for its ID, for example RFW-07.
3. Claim it early. Fork the repository, push a branch named rfw/07-short-name with an empty commit (git commit --allow-empty -m "Claim RFW-07"), and open a draft pull request titled [RFW-07] Short title. Fill in the template's Claim and Plan. Theories are claimed as [T-NN] and paid-request preparation as [PAID-NN].
4. Deliver against the acceptance line. A negative result with numbers is a full delivery. Follow CONTRIBUTING.md.
5. Hand off. Fill in the template's Handoff section: what changed, the validation numbers, what remains uncertain and where you stopped. Append a dated entry to docs/LOG.md and mark the pull request ready for review.

A claims workflow labels each claim. The earliest open pull request for an item holds it, later ones are marked duplicate, and a draft with no new commit for 72 hours lapses so the item is open again. A pull request ready for review never lapses.

Copy this into your agent to start:

You are donating a session to Open Observatory: https://github.com/recozers/openobservatory
Fork and clone it, then read README.md, CONTRIBUTING.md and REQUESTS_FOR_WORK.md, including "Results so far".
Pick one open item that fits about <N> hours and needs only these credentials: <none | Earth Engine | EPA | Earthdata>.
Prefer Priority 1 (China). Check it is not claimed on https://openobservatory.info/requests.html or in open pull requests.
Claim it at once: push a branch with an empty commit and open a draft pull request titled "[RFW-NN] <title>" from the template.
Keep every number traceable to a public source and a script. Before you finish, run:
python -m unittest discover -s tests; node --test tests/*.test.cjs; python tools/refresh.py --dry-run
Record the result in docs/LOG.md, including negative results, fill in the template's Handoff section with what changed,
validation numbers, remaining uncertainty and where you stopped, then mark the pull request ready for review.

Keys: none needs nothing; EE needs Google Earth Engine, free for non-commercial research with your own Cloud project; EPA needs a free api.data.gov key; Earthdata needs a free NASA Earthdata login. Keep keys in a local .env and never commit them. Size is agent time: S under 2 hours, M 2 to 6, L more than 6, delivered as a first slice. Answers names the goal question an item serves: Where, Built, Running, Capacity, Utilisation or Workload (training or inference).

Ground rules

• Every public number traces to a free public source, a script in this repository, and a polygon with a stated confidence. One exception: confidential training-run labels from AI labs may train the workload model (RFW-23 gives the conditions).
• Official national statistics are inputs to test, never evidence.
• State results at the scale they were tested: how many sites, against what truth, with what uncertainty.
• Negative results stay visible. Document a failed method; do not delete it.
• Respect data licences and site terms. Do not scrape services that forbid it, never bypass a CAPTCHA, login or other access control, and follow the Nominatim usage policy for geocoding.
• A document's figure keeps its own quantity. Do not turn kVA into IT megawatts, add facility and IT figures, compare a load with a capacity from a different period, or treat a forecast as an observed load.

Requests

Status shows standing reservations by the project's own agents. Live claims come from open pull requests, and the website marks them in these tables when the page loads.

Priority 1: China

RFW-01 Confirm or reject the radar-detected structures in China

• Why: 39 inventory entries named cn_<hub>_r<rank> are new hall-like structures found by radar, dated, and unconfirmed. Each needs a verdict before it counts as a data centre.
• Do: For each entry, look at its chip (results_s1/<hub>_candNN.png, or render one with tools/chip.py) and its classifier score in results_cand/scores.csv. Classify it as data-hall complex, not a data centre, or unclear, with a one-line reason naming what you see: long parallel halls, cooling yards, generator rows, a substation. For confirmed halls, check the outline in data/polygons/<site_id>.geojson against the chip.
• Deliver: data/cn_radar_review.csv with site_id, verdict, reason and evidence path. Rejected entries leave the public inventory but stay in the review file. The site rebuilds and the tests pass.

RFW-02 Locate every national computing cluster and scan it

• Why: China's national computing network names ten data-centre clusters. Of 12 radar boxes so far, six around digitised campuses and hub centroids found the 39 entries, six placed from approximate coordinates found nothing, and some clusters, such as the Yangtze River Delta demonstration zone, have never been scanned.
• Do: Find each cluster's data-centre parks and their locations or boundaries in park plans, environmental impact documents or land transfer notices, with URLs. Scan each park with RFW-26's command and score the candidates.
• Deliver: data/cn_hub_parks.csv with cluster, park, coordinates or boundary, source URL and page; candidate files per park; and a one-line verdict per cluster in docs/LOG.md.

RFW-03 Quarterly construction index for each Chinese hub

• Why: The simplest test of whether building continues or stops is new hall area per hub per quarter, and the radar entries already give each structure a month and an area.
• Do: Aggregate new hall area by hub and quarter from the radar entries and their outlines, keeping confirmed, unclear and rejected structures separate once RFW-01 has run. Report an uncertainty from the dating spread, and extend to new parks as RFW-02 adds them.
• Deliver: results_cn/construction_index.csv and a short note in docs/MVP.md stating what the index can and cannot show, with the counts behind each quarter.

RFW-04 Land transfer results for operators and start dates

• Why: Land transfer result notices name the parcel, its area, the buyer and the date. They can attribute radar-found structures to operators and date the start of each project.
• Do: Search the national land market network and provincial natural-resources bureaus for industrial parcel results in the hub parks. Match parcels to radar entries and inventory campuses by location and area.
• Deliver: data/cn_land_transfers.csv with parcel, area, buyer, date, URL and the matched site, and a note on match confidence. An operator is attributed only when a notice names it.

RFW-05 Procurement tenders and awards, including training and inference servers

• Why: Procurement notices for servers, racks, UPS systems and cooling name campuses, quantities and dates, which time each phase of fit-out. Some Chinese server tenders state whether the servers are for training or for inference, which would be the most direct public evidence of workload at a named campus.
• Do: Search the China Government Procurement Network and the three telecom operators' procurement portals for tenders and awards naming campuses at Ulanqab, Horinger, Zhangbei, Zhongwei, Qingyang, Gui'an and Chongqing. Record campus, item, quantity, units, stated purpose (training, inference or general), date and URL. These are company procurement records, not statistics.
• Deliver: data/cn_procurement.csv and a note on what each portal exposes. Do not convert kVA or rack counts into IT megawatts, and record a stated purpose only where the notice gives it.

RFW-06 Stated workload roles and cloud regions at Chinese campuses

• Why: Before any signal can separate training from inference, each campus needs its stated role to test against. Public cloud regions are a start: a campus that hosts a region serves outside customers, which suggests inference and general cloud work.
• Do: For each Chinese inventory campus and hub, record public cloud regions or zones hosted there from providers' own region lists, and the operator's stated purpose from company announcements or filings. Record national and provincial planning statements separately, marked as official inputs to test.
• Deliver: data/workload_roles.csv with site, stated role, source type, quote, URL and date. A role is recorded only where a source states it.

RFW-07 Chindata's per-data-centre table and an EDGAR name search

• Done so far: a first slice was delivered in pull request 15. tools/chindata_filings.py put Chindata's per-data-centre tables for mid-2020, end-2021 and end-2022 into data/cn_operator_disclosures.csv, 166 rows whose sums match each table's totals. It relabelled the 324 MW Zhangjiakou and 308 MW Datong figures as total capacity, not capacity in service. A hub-name search of 12 annual reports from VNET, GDS and Chindata found no capacity figure for Ulanqab, Zhangbei, Horinger, Gui'an or Zhongwei; GDS reports four data centres in Zhangbei. See docs/cn_operator_disclosures_notes.md, section 6.
• Why the rest matters: the tables give most data centres no city, so their figures cannot yet be tied to a campus. VNET's Ulanqab orders appear only in quarterly releases, and the Internet Archive holds none of VNET's releases from 2025 on.
• Do: Run EDGAR's full-text search for the five hub names across VNET, GDS and Chindata filings, including 6-K current reports. SEC requires a contact email in the User-Agent, so agree the address with Stuart first. Find the city of each remaining Chindata data centre in its other public documents, such as quarterly releases or Chinese-language land and environmental records.
• Deliver: New per-campus figures in data/cn_operator_disclosures.csv and locations in data/chindata_dc_locations.csv, each with its source URL and a verbatim quote.

RFW-08 Substation and transmission projects serving the hub parks

• Why: Environmental impact documents and approvals for new substations and lines often state capacity and the load they serve, such as a named data-centre park. That bounds the power a park can draw, the capacity half of utilisation, and dates when it could draw it.
• Do: Find substation and transmission project documents near each hub park. Record voltage, transformer capacity, stated served load, approval and commissioning dates, and URLs.
• Deliver: data/cn_grid_projects.csv and a note per hub. Transformer MVA is supply capacity, not IT load; keep the units.

RFW-09 Water-withdrawal permit notices for hub data centres

• Why: Water-withdrawal permit notices for data-centre projects can state permitted annual withdrawal, which bounds cooling water use. Most western hubs use little water, so absence is also informative.
• Do: Search provincial water-resources bureaus' permit notices for data-centre projects in the hub parks, and record the project, operator, permitted volume, date and URL.
• Deliver: data/cn_water_permits.csv and a note. Do not convert volumes to electricity until RFW-17 has a validated method.

RFW-10 Building-scale night lights inside Chinese parks

• Why: At park scale, VIIRS fails in China because the parks were already lit. The radar entries now give building outlines, which may be enough to see a single building energise inside a lit park.
• Do: Run tools/ntl_timeline.py on the radar entries and the three digitised campuses, with the comparison ring drawn inside the park. Test the same set-up first on a US campus inside a lit district, Prometheus at New Albany.
• Deliver: Lit-up months where they pass the existing rule, and a verdict on whether building scale works. Stop if the New Albany test shows no step.

RFW-11 Multi-storey data-centre positives for the classifier

• Why: The classifier learned from single-storey hyperscale halls. Many data centres in eastern China are multi-storey buildings it has never seen, which may be why six eastern hub boxes found nothing.
• Do: Assemble at least 30 verified multi-storey data-centre buildings from operator pages or filings, with coordinates, including Chinese ones where a document names the building. Extract features with tools/cand_features.py, retrain, and report leave-one-group-out results separately for single-storey and multi-storey positives.
• Deliver: A labelled positives file with sources, updated scores, and the per-class results in docs/LOG.md.

RFW-12 Dedicated plants and public stack monitors in China

• Status: reserved by Astra (B8 in docs/PHASE2_TODO.md). A first pass was merged on 14 September; see docs/cn_stack_monitors.md.
• Why: Plants built for hub parks may publish hourly stack monitoring on provincial platforms, the counterpart of EPA plant records, and hourly readings could show a park's daily load cycle. So far only annual reports have been reachable, and hourly readings at the one linked plant sit behind a CAPTCHA.

RFW-13 Operator utilisation priors from Chinese filings

• Status: reserved by Astra (B9 in docs/PHASE2_TODO.md).
• Why: VNET, GDS and Chindata report company-wide commercial utilisation, a sourced prior for campuses whose operator is known.

Priority 2: Utilisation and workload, at higher time resolution

These methods are developed where ground truth exists, mostly in the US, so they can be carried to China once proven.

RFW-14 Make the plume test robust to start date and season

• Why: Abilene reaches 2.8σ with its documented start and 2.3σ with a start one month earlier, and several sites with spring starts show strongly negative z-scores. The test compares unmatched seasons and depends on one date.
• Do: Using the saved daily series in results_no2/<site>.csv, add a season-matched version of the test and report each site's z-score across start dates within three months of the documented one. Re-run the 62 control points the same way.
• Deliver: The revised test in tools/plume_batch.py, per-site date sensitivity in a results file, and a recommendation on whether the 2.5σ rule should change, with the control-point false-positive rate under the new test.

RFW-15 Near-field plume test for plants on a city's edge

• Why: The box method cannot separate a campus plant from a nearby city's plume, which rules out Dublin and similar grid-constrained markets where campuses run gas plants.
• Do: Build a sector test within 1 to 5 km that keeps only days when the wind blows from the side away from the city and excludes the city from the upwind sector. Validate it first on a US plant at a city's edge with EPA hourly emissions, then apply it to Grange Castle.
• Deliver: The validation result against EPA truth and, only if validation passes, the Dublin result. A failed validation is a full delivery.

RFW-16 Qualify gas-turbine NOx calibration plants

• Why: The NOx calibration rests on five tall-stack coal plants, but turbine exhaust leaves from low stacks. Three gas candidates gave factors from 0.89 to 4.83 and are not qualified (docs/low_stack_calibration.md). Until turbines are calibrated, NOx gives relative change but no utilisation ratio.
• Do: Verify physical stack heights from EIA-860's environmental-equipment data or permits, check isolation from other sources with the EPA's point-source inventory, and align the TROPOMI overpass time with the hourly records.
• Deliver: Each candidate qualified or rejected with its reason, and the calibration comparison updated for the qualified ones.

RFW-17 Climate-aware water efficiency for water-derived loads

• Why: Water use per unit of energy depends on climate, but the current method applies one fleet-wide figure. Against Meta's electricity it puts individual campuses anywhere from 0.35 to about 3 times their reported load.
• Do: Meta publishes both water withdrawal and electricity for 18 campuses. Model withdrawal per unit energy as a function of local climate, such as ERA5 wet-bulb temperature, and test it leaving one campus out at a time. Apply it to Google's campuses only if it beats the current method.
• Deliver: Leave-one-out errors for the current and new methods side by side; updated derived loads and a narrower band only if the new method is better.

RFW-18 Per-site electricity, capacity and commercial utilisation from more operators

• Why: Utilisation needs a load and a capacity for the same period, and only Luleå and Frontier have both. US colocation operators also publish commercial utilisation, which the project has not collected.
• Do: Survey sustainability reports, data appendices and investor filings from Apple, Equinix, Digital Realty, NTT, Iron Mountain, Switch, CyrusOne, QTS, Aligned, and the Chinese operators GDS, VNET and Chindata, for per-site electricity, capacity, PUE, water and utilisation, at the finest period published. Add rows to data/operator_disclosures.csv with source URL, table and period, add missing campuses with sourced coordinates, and run tools/ingest_disclosures.py.
• Deliver: A table of which operators publish what, at what resolution and period; every new figure with its source; and the list of sites that now have a load and a capacity for the same period.

RFW-19 Microsoft's metro electricity table to campuses

• Why: Microsoft's 2026 data fact sheet reports fiscal-2025 electricity and water by metro area. Where a metro holds one campus, that is a campus figure.
• Do: Map each metro row to inventory campuses. Attribute a figure only where the metro is effectively one campus, and record the rest as context.
• Deliver: Rows in data/operator_disclosures.csv with the attribution reasoning, and a note listing the metros left as context and why.

RFW-20 Generator fleet and dedicated plant discovery from EIA-860M

• Why: The generator watch list was assembled by hand from permits. The US Energy Information Administration's monthly generator inventory lists planned and operating generators by plant, which could flag new fleets built for data centres and plants that serve one campus. A dedicated plant that reports hourly to the EPA would give the first hourly load shape, the data needed to tell training from inference.
• Do: Parse the latest EIA-860M. Flag planned or new gas turbine and engine plants whose owner, name or location links them to a data centre, or that sit within a few kilometres of an inventory site, and check which report hourly to the EPA. Verify each flag against a permit or company statement.
• Deliver: A script, a candidate table with verification notes, and verified candidates proposed for the watch list or as dedicated plants. Coordinate with the RFW-21 claimant before adding watches.

RFW-21 Evidence for the pending generator watch records

• Status: reserved by Astra, working in astra/b7.
• Why: Seven planned on-site generation fleets need filing-located coordinates, a first-fire target and a valid emission factor before they can be watched monthly. See docs/generator_watchlist.md.

RFW-22 Stated workload roles for the rest of the inventory

• Why: The site's classes for AI-training campuses were assigned by hand from press coverage. Any training-or-inference signal needs sourced labels to test against.
• Do: For each non-Chinese inventory campus, record the operator's stated purpose and any public cloud region hosted there, with the quote, source type, URL and date, as RFW-06 does for China. Keep company statements, press reports and official planning documents distinct.
• Deliver: Rows in data/workload_roles.csv, and a list of hand-assigned site classes that no source supports.

RFW-23 AI lab partners with training-run records

• Why: The planned training-or-inference method (T-05) needs labels: when each campus was running training, and when it was serving inference. The AI labs that use these campuses, such as OpenAI, Anthropic, xAI, Google DeepMind and Meta, know when and where their training runs happened.
• Do: Map each lab to the inventory campuses it trains or serves on, from its own statements and those of its cloud and data-centre partners. Record the training runs labs have announced publicly, with model, dates or duration, hardware scale and campus where stated, and the source URL; these are coarse public labels to start from. Draft a data-sharing request for Stuart to send, asking each lab for training-run start and end dates by campus, the share of capacity each run used, and inference-serving periods, and what the lab would allow to be published. The request sets out how labels given in confidence are handled, using the label condition below.
• Deliver: data/workload_labels_public.csv with lab, campus, model, period, scale, source type and URL; the lab-to-campus map; and the draft request in docs/ai_lab_partner_request.md. Nothing is sent without Stuart.
• Label condition (decided by Stuart, 14 September 2026): labels a lab publishes, or agrees to have published, are used openly. Labels given in confidence may train and validate the model. They stay in data/private/, which git ignores, and never appear in commits, pull requests, issues or site data; only Stuart and the sessions Stuart runs handle them, so a donated session never receives them. The model's code, its aggregate validation scores and its outputs for campuses the labels do not cover are published, each output marked as coming from a model trained partly on confidential labels. Nothing is published for the campuses and periods those labels cover, and the trained model is not released without the lab's agreement, because either could reveal the labels.

RFW-24 Utilisation ramp curves from Meta's 18 campuses

• Why: How fast utilisation rises after a building is finished is the key prior for estimating it between annual disclosures. Meta's electricity series from 2011 to 2024 across 18 campuses measures the load side of that ramp.
• Do: Date each campus's buildings from Sentinel-2, radar, or Landsat for older ones. Fit average load per building against years since roof-on, with uncertainty.
• Deliver: Fitted ramp parameters with uncertainty in docs/utilisation_model.md, and the per-campus fits.

RFW-25 Implement the utilisation model

• Why: The site uses fixed rules per evidence type. docs/utilisation_model.md defines one model in which every source moves the utilisation estimate consistently, with stated priors and posteriors per site.
• Do: Implement section 7 of that document: hall states, the evidence likelihoods, and Monte Carlo percentiles in a pure-Python utilmodel.py. Keep the timeline JSON shape so the site needs no change.
• Deliver: A first slice covering reported electricity, documented capacity and roofs-only sites, behind a switch, with tests showing it reproduces today's bands where the rules are already calibrated.

Priority 3: Coverage and tools

RFW-26 Radar candidate scan around every inventory site

• Why: Radar dates construction through cloud and finds unlisted buildings, but it has run in only 12 Chinese hub boxes, 9 US boxes and 40 night-light areas.
• Do: Start with Chinese sites. For each site in data/sites.csv without a results_s1/<site_id>_candidates.csv, run python tools/s1_timeline.py --chip LAT LON --candidates --half 6000 --early 2021 --late 2026 --out results_s1/<site_id>, then score with tools/cand_features.py and tools/cand_classifier.py. Work region by region.
• Deliver: A candidate file per site; a table in docs/LOG.md of recall at known halls and the false-positive types by region. Stop and report if Earth Engine quotas make a region impractical.

RFW-27 Date dark and grey roofs

• Why: Brightness dating misses dark membranes and grey roofs, as at Hyperion and at Chinese campuses such as Ulanqab. A vegetation and built-up index rule already failed (docs/dark_roof_validation.md).
• Do: Use radar structure-on months as labels and test other Sentinel-2 signals, such as shortwave-infrared change, texture or temporal variance. Validate on Hyperion, the Ulanqab long halls and all 12 Abilene halls, and check that fallow fields produce no false events.
• Deliver: A rule that dates Hyperion's two structures without a false early event and moves no Abilene date by more than a month, or a documented failure with the numbers.

RFW-28 Recalibrate the hall classifier with new labels

• Why: The classifier has 18 positives. RFW-01, RFW-11 and the 78 Epoch campuses with polygons can multiply that.
• Do: Add the new labels, retrain with tools/cand_classifier.py, and report leave-one-group-out results by region and building type. Refresh the scores of existing candidates.
• Deliver: Updated results_cand/scores.csv, results in docs/LOG.md, and a note on any threshold change.

RFW-29 Standby generator permits as a capacity bound

• Why: State air permits list the diesel standby generators at many US data centres. Their total rating bounds facility power, because campuses back up their full load, which gives a sourced capacity figure where none exists.
• Do: For inventory sites in Virginia, Ohio, Texas, Arizona and Georgia, find the air permits and record unit counts, ratings, fuel, permit date and URL. Compare total standby MW with documented capacity where both exist.
• Deliver: data/standby_generation.csv and a note giving the ratio of standby MW to documented capacity across at least 10 sites. Propose a capacity-bound basis only if the ratio is consistent to within about 30 %.

RFW-30 Construction timeline on the map

• Why: Every dated building has a month, but the map shows only the present.
• Do: Add a time slider to the landing map that shows sites as they were at the end of each quarter, using the dates already in site/data/timeline/. Keep it static with no new framework.
• Deliver: The slider working at desktop and phone widths, and Node tests in the style of tests/frontend_evidence.test.cjs.

RFW-31 Tests for the load precedence rules

• Why: The rules deciding which figure sets a site's load have been extended several times: reported electricity, carried averages, water-derived figures, third-party estimates, plant records, radar entries, generator watches.
• Do: Write synthetic cases for cap_in_force, utilisation_prior and the evidence kinds in build_status.py, covering each basis and each carried-forward case.
• Deliver: Tests that fail if any precedence rule changes silently, all passing on main.

RFW-32 Source link checker

• Why: Hundreds of source URLs back the numbers on the site, and links rot.
• Do: Write tools/check_links.py to read every URL in data/*.csv and the source JSON files, check each at a polite rate, and look up whether an archived copy exists without submitting new captures.
• Deliver: A report of dead links with their archived alternatives, and the script documented in CONTRIBUTING.md.

RFW-33 Crawl public records and satellite data for data centres in the rest of the world

• Why: Outside the US and China the inventory has 12 sites in 11 countries, eight of them Epoch AI campuses, and dated buildings at 6. No radar, optical or night-light discovery scan has run there. The only regional records are official totals for Ireland and Singapore in data/regional_dc_load.csv, which can test a lead list's completeness but never place a site. Planning, environmental and permit records name data-centre projects, locate them and date them before construction.
• Records: Build a lead list, starting with sources that cover a whole country, and record what each publishes per site. Sources to check: OpenStreetMap features tagged as data centres, through the Overpass API; reporting by data centres of 500 kW and above under the EU Energy Efficiency Directive, whose EU database is public only in aggregate, and the national registers that implement it, such as Germany's; the public registers of medium combustion plants in EU countries and the UK, whose entries give operator, location and rated thermal input and can include standby generators; industrial emissions permits for sites whose generators pass 50 MW of thermal input, such as Ireland's EPA licences; planning applications and zoning plans published as open data; environmental assessment portals such as India's PARIVESH, Chile's SEIA and the New South Wales major projects portal; PeeringDB facilities; and the property lists of listed operators and data-centre REITs.
• Satellites: Measure recall where the answer is known before searching. The radar scan compares 2021 with 2026, so it finds buildings that went up between those years. Five rest-of-world campuses have halls that Sentinel-2 dates to 2022 or later: DayOne Nusajaya, Google Waltham Cross, Oracle Batam, Southgate Melbourne and Start Campus Sines. Run RFW-26's command around each, reusing its files where they exist, and score the candidates. tools/s2_candidates.py runs only on five fixed boxes, so add a coordinate option before using it. Then scan boxes around located leads, and around announced campuses where no record turns up; announcements are leads, never evidence. The classifier has never seen a multi-storey data centre, common in dense metros, so report its scores by building type.
• Match: A lead is confirmed when a public record names a data centre at a location where a hall-like structure is visible on a chip. Add confirmed sites to data/sites.csv with an outline and coords_quality. A lead with a record but no structure yet stays in the leads file as a watch point with its record date, so later scans can date it. A structure with no record stays in the candidate files, unconfirmed. Keep each figure in its own unit: megawatts in a planning application, kVA, thermal input and floor area are not IT capacity.
• Rules: Crawl at a polite rate within robots.txt and each source's terms. At a CAPTCHA, login or ban on automated access, stop and record the block in the source audit. Commit extracted facts and links, not copies of documents whose terms forbid redistribution, and no personal details of people named in applications, such as objectors. Use commercial data-centre directories only by hand, as leads that must resolve to a public record. Keep the original-language text that names each project beside any translation.
• Deliver: A first slice for two areas, for example Ireland, where official totals exist to test completeness, and Johor in Malaysia, where DayOne Nusajaya is a known campus. data/row_record_sources.csv with country, source, what it publishes per site, access method, terms on automated access, URL, date checked and outcome; data/row_leads.csv with country, project, operator where a record names one, record type and date, stated figure and unit, location and how it was found, source URL and status; the crawler scripts under tools/; candidate and score files per scanned area; and a table in docs/LOG.md giving, per country, sources checked, leads, located leads, matched structures and new sites, plus the radar recall at the five campuses and the false-positive types. The site rebuilds and the tests pass. If the radar scan finds fewer than three of the five campuses, report the recall and stop before scanning areas without records. If Earth Engine quotas block an area, say so; PAID-08 covers paid compute.

Theories to explore

Research bets with a first test that fits a session and a condition for stopping. China first.

T-01 Chinese power-exchange market records

• Hypothesis: Provincial power exchanges publish market participant registrations and green-power transaction notices that name data-centre companies and, sometimes, volumes.
• First test: Search the Inner Mongolia and Beijing exchanges' public notices for companies operating at the hub campuses, and record what each notice gives. These are company transaction records to be tested, not statistics.
• Stop if: no volume can be attributed to a campus.

T-02 Hourly NO2 from GEMS over plants supplying Chinese parks

• Hypothesis: GEMS, a geostationary instrument over East Asia, measures NO₂ every daylight hour. Over a plant that supplies a hub park, such as Shengle at Horinger, it could show the plant's daily output cycle, and with it the park's daily load shape, where TROPOMI sees only one moment a day.
• First test: Check GEMS data access and terms with Korea's National Institute of Environmental Research, then compare GEMS NO₂ over Shengle at TROPOMI's early-afternoon overpass with TROPOMI itself, and look for a repeatable daytime cycle.
• Stop if: GEMS and TROPOMI disagree at the overpass hour beyond the calibration scatter, or no daytime cycle stands out from day-to-day noise. Plant output is not campus load without an allocation, as RFW-12 found.

T-03 Temporary site housing as a construction signal

• Hypothesis: Large Chinese construction sites house workers and site offices in prefabricated blocks with distinctive blue roofs, visible in Sentinel-2. They appear when work starts and go when it ends, so they may lead radar dating and mark completion.
• First test: Track blue-roof area within 1 km of five radar-dated Chinese structures by month, and compare its appearance and removal with the radar structure-on month.
• Stop if: it does not appear at least a month before radar structure-on at three of the five.

T-04 Radar backscatter after the structure goes up

• Hypothesis: Backscatter keeps rising after a hall's structure appears, as rooftop plant and yard equipment are installed, which would make fit-out visible through cloud.
• First test: At Abilene, compare each hall's backscatter after structure-on with the roof darkening that marks fit-out in Sentinel-2 six to nine months later, then apply the result to the Chinese radar entries.
• Stop if: Abilene shows no post-structure rise distinguishable from month-to-month noise.

T-05 Transformer heat as a load and workload signal

• Hypothesis: A transformer's load losses rise with the square of its current, so its tank and radiator temperatures follow the load it carries, with a lag of hours set by its oil and cooling. Imaged at a few metres several times a day, a campus substation would give a load series, and the number and size of its transformers would give capacity. A deep-learning model trained on substation imagery from campuses whose AI labs have supplied training-run records (RFW-23) could then estimate utilisation and tell training from inference at other campuses, including in China. This is the project's planned method for the workload question.
• First test, free: Use the standard transformer loading models (IEEE C57.91 and IEC 60076-7) with load profiles for training and for inference, taken from public cluster traces and published descriptions of training load, to predict surface temperature changes. Compare them with the noise of thermal sensors that exist or are planned. Then sample the simulated temperatures as a satellite would, add that noise, and check whether a classifier still separates the two workloads.
• Second test: image the substation of one US campus with training-run records, across periods with and without training (PAID-03), and check that the temperature pattern changes with the labels. RFW-23's label condition applies: with labels given in confidence, only the aggregate result is published, never a dated series for that campus.
• Stop if: the predicted changes are smaller than achievable sensor noise at a few metres, the simulation cannot separate the workloads at achievable sampling, or the pilot campus shows no change between labelled periods.

T-06 Hourly NO2 from TEMPO at turbine-fed campuses

• Hypothesis: TEMPO, a geostationary instrument over North America, measures NO₂ every daylight hour. At a campus that runs its own turbines, such as Colossus 2, it could show whether generation is flat through the day, as training would be, or follows a daily cycle, as inference would.
• First test: Recover the known hourly NOx cycle of a large EPA-reporting plant from TEMPO first, using the plant's hourly records as truth. Only then look at the hourly shape at Colossus 2.
• Stop if: the reporting plant's hourly cycle cannot be recovered.

T-07 Network presence as a sign of inference

• Hypothesis: Inference serves users, so it needs low-latency connections: campuses listed in PeeringDB with many networks, or hosting a public cloud region, are more likely to run inference than isolated training campuses.
• First test: Record PeeringDB facility entries and hosted cloud regions for inventory campuses, respecting PeeringDB's usage policy, and compare with the sourced roles from RFW-06 and RFW-22.
• Stop if: network presence does not separate campuses with stated training roles from those with stated inference or cloud roles.

T-08 Radar coherence over fan yards

• Hypothesis: Operating cooling and fan yards lose radar interferometric coherence faster than idle yards.
• First test: Compute 6 or 12-day Sentinel-1 coherence from single-look complex data at yards and roofs before and after documented energisation at three US sites, then at Chinese campuses if it works.
• Stop if: no step appears at the three US sites.

T-09 Cooling-tower vapour plumes

• Hypothesis: On cold humid mornings, plumes above cooling towers are visible in Sentinel-2 and Landsat, and how often they appear tracks operation.
• First test: Tally plume presence on clear winter scenes at evaporative-cooled sites before and after documented start dates.
• Stop if: weather explains plume presence better than operating status.

T-10 Wastewater discharge reports as a monthly cooling series

• Hypothesis: Where a campus discharges cooling-tower water under its own federal permit, the EPA's monthly discharge monitoring reports give a measured monthly flow that scales with heat rejected, and so with load.
• First test: Search EPA ECHO for permits held by data-centre operators at inventory sites, pull their monthly flows, and check the seasonality against wet-bulb temperature and against Meta's electricity where both exist.
• Stop if: no inventory campus holds an individual permit with flow reporting.

T-11 Building permits and occupancy certificates as energisation dates

• Hypothesis: US county permit records give the month each building was certified for occupancy, which marks the step between a finished building and one drawing load.
• First test: Match permit records from counties that publish them as open data, such as Loudoun and Franklin, to inventory halls, and compare occupancy dates with radar structure-on and with Meta's load ramp.
• Stop if: fewer than 10 inventory buildings can be matched.

T-12 Crane filings as construction-start signals

• Hypothesis: The FAA's public obstruction evaluations include temporary structures such as cranes, with coordinates and dates, which would flag construction starts months before radar sees a structure.
• First test: Pull filings near 20 US inventory campuses and compare the first crane date with the radar structure-on month.
• Stop if: filings match fewer than half the campuses or lead radar by less than a month.

T-13 Counting cooling equipment in public aerial imagery

• Hypothesis: USDA's public-domain NAIP aerial imagery, at 0.6 to 1 m, resolves chillers, dry coolers and cooling towers. Count times unit capacity from Epoch's equipment catalogues in data/epoch/ bounds the heat a campus can reject, and so its IT capacity. The same method in China needs purchased imagery (PAID-01).
• First test: Count units at five US sites with documented capacity and compare.
• Stop if: the estimate misses documented capacity by more than a factor of two.

T-14 Substation transformer bays as connection capacity

• Hypothesis: The number of transformer bays at a campus substation, visible in NAIP, indicates its grid connection capacity.
• First test: Count bays at campuses with documented connections, such as Luleå's 120 MW and New Albany's 250 MW, and compare bays times typical ratings, matching each figure's period to the imagery date.
• Stop if: ratings vary too much to beat a factor of two.

T-15 Utility retail sales where one campus dominates

• Hypothesis: For small utilities whose load is mostly one campus, the Energy Information Administration's annual retail sales by utility approximate that campus's consumption.
• First test: Identify serving utilities for inventory campuses, find those where recent sales growth matches the campus, and validate on a Meta campus with reported electricity.
• Stop if: no validated case lands within 30 % of Meta's figure.

Requests that cost money

These need funding rather than agent time, and they are where commercial data could close the gap in China. Nothing is bought until a pilot shows it is worth it.

• Who does what. An agent can claim the preparation as [PAID-NN]: a pilot specification, analysis code tested on free data, and a draft request for quotation. Agents never place orders, sign licences or handle payment; Stuart does.
• Free access first. Check research access before paying, such as NASA's commercial smallsat data programme for NASA-funded researchers, ESA's third-party missions programme, and providers' research programmes. Eligibility varies.
• Licence condition. Every public number must trace to free public data. A purchase helps only if its licence lets the project publish the derived numbers, and ideally the data, or if it is used purely to validate a method that then runs on free data. Buy nothing whose licence allows neither.
• Cost. No prices are listed because none has been quoted. Each request names what drives cost.

PAID-01 Sub-metre optical imagery of the Chinese hub parks

• Answers: whether the radar-found structures are data halls; counts of cooling units and generator rows, which bound capacity as T-13 does with free US imagery; phase dating at building level.
• Pilot: one recent cloud-free archive scene each over Horinger, Ulanqab and Zhangbei, then one more a year apart.
• Cost drivers: area, resolution, archive against new tasking.
• Success test: RFW-01's chip-based verdicts agree with the sub-metre view, and two independent unit counts match.

PAID-02 High-resolution radar over cloudy southwest hubs

• Answers: construction progress and equipment yards at Gui'an and Chongqing, where optical imagery is rarely clear and Sentinel-1's 20 m cannot resolve individual halls.
• Pilot: four spotlight-mode scenes a quarter apart over two parks.
• Cost drivers: imaging mode, number of acquisitions, archive availability.
• Success test: hall-level structure dates that Sentinel-1 cannot give, checked against later clear optical scenes.

PAID-03 Thermal imagery of substations and transformers at a few metres

• Answers: the data for the planned utilisation and training-or-inference method (T-05). Free thermal imagery is 70 m or coarser, while a campus transformer is a few metres across.
• Pilot: repeated acquisitions, day and night, over the substation of one US campus whose AI lab has supplied training-run records (RFW-23), spanning periods with and without training. Then a US campus without records, to test the model. Then one Chinese hub park by satellite, only if the pilot shows a signal.
• Cost drivers: acquisitions per day and per week, which decide whether daily cycles can be seen; satellite tasking or an airborne survey; area. An airborne survey is not an option over China, and flights near US campuses may need permission.
• Success test: substation temperature patterns differ between labelled training and non-training periods beyond the scatter between acquisitions. Stop if the pilot shows no difference. If the labels were given in confidence, only the aggregate result is published (RFW-23's label condition).

PAID-04 Frequent 3 to 5 metre optical monitoring of Chinese parks

• Answers: construction and fit-out pace by month at the fastest-building parks, sharper than Sentinel-2's 10 m.
• Pilot: near-daily imagery over five parks for three months, compared with the radar construction index (RFW-03).
• Cost drivers: area, revisit, subscription length. Check research access programmes first.
• Success test: monthly changes that the free index misses or dates later.

PAID-05 Dedicated satellite capacity

• Answers: sustained, independent monitoring of Chinese data-centre parks, the end state if a paid pilot proves an observable.
• Pilot: none until PAID-01 or PAID-03 shows which observable works. If PAID-03 passes, the likely requirement is thermal imagery at a few metres several times a day over the listed parks' substations. Write the requirements: resolution, revisit, day and night capability, spectral bands and the list of parks, and check whether a long-term tasking contract meets them before considering a hosted payload or a dedicated small satellite.
• Cost drivers: set entirely by those requirements, so no estimate is possible yet.

PAID-06 Native-language review of Chinese documents

• Answers: whether agent extractions from land notices, procurement records, stated roles and permits (RFW-04, RFW-05, RFW-06, RFW-08 and RFW-09) are correct, including stated training or inference purposes.
• Pilot: a Chinese-reading reviewer checks 20 extracted documents against the originals.
• Cost drivers: reviewer hours.
• Success test: an error rate per field that shows which extractions can be trusted without review.

PAID-07 Chinese corporate registry data

• Answers: which companies own the operating entities at hub campuses, and when they were set up.
• Pilot: ownership chains for the operating companies at 10 campuses.
• Cost drivers: subscription tier. These services usually forbid redistribution, so use them for leads and cite the free public document each lead points to.
• Success test: at least half the leads resolve to a free public source.

PAID-08 Compute for continental radar scans

• Answers: RFW-02, RFW-26 and RFW-33 across whole provinces or countries, if Earth Engine's non-commercial quotas block the scans.
• Pilot: one province by batch export to cloud storage, with the cost recorded per 1,000 km².
• Cost drivers: storage, processing and egress.
• Success test: candidate lists for the province that match the box scans where they overlap.

PAID-09 US public-records request fees

• Answers: utility, water and generator records for named US campuses, where agencies charge to process requests.
• Pilot: five requests to agencies serving inventory campuses, chosen to test T-10 and T-15.
• Cost drivers: processing and copying fees.
• Success test: at least two responses give an attributable monthly or annual series.

Results so far

Each result is stated at the scale it was tested. Numbers and corrections are in docs/MVP.md and docs/LOG.md.

What works, and its limits

• Roof dating from Sentinel-2 brightness (tools/s2_roof_timeline.py). Dates new bright roofs to the month. The only check so far is against radar: all 12 Abilene halls agree with Sentinel-1 dates within 0 to 4 months, 9 within 2. No dated construction records have been compared. It misses dark and grey roofs: across the imported Epoch campuses, 217 halls were dated, 34 left unresolved and 16 conflicted with reported construction timing; at Horinger, 4 of 27 radar-found structures got a brightness date. China: works for bright roofs only, and grey roofs are common.
• Structure dating from Sentinel-1 radar (tools/s1_timeline.py). Dates a hall when backscatter rises at least 4 dB and stays up for six months. It gave no false event over Hyperion's farmland and dated Chinese blocks that brightness could not. It has not been checked against construction records. China: works, including under cloud.
• Radar change scan for unlisted buildings (tools/s1_timeline.py --candidates). In US boxes the known halls were among the candidates at Abilene (ranks 1, 2, 4 and 5), Rainier (1 to 3), Fairwater (1) and Prometheus (8), and were absent at Hyperion, whose structures went up in 2025 and 2026. China: of 12 hub boxes, the six around digitised campuses and hub centroids produced 39 hall-like entries, dated and unconfirmed; the six eastern boxes, placed from approximate coordinates, produced none.
• Hall classifier (tools/cand_features.py, tools/cand_classifier.py). Trained on 18 positives from six campuses, three of them labelled by inspecting chips, and 1,159 negatives: 1,133 new industrial structures in 40 Chinese areas found by the night-light scan, 25 near four US factories and one solar compound. Leave-one-site-out AUC is 0.977; at a 0.5 threshold it keeps 14 of 18 halls and 48 of 1,159 negatives. Intel's Ohio fab under construction scores 0.99. It has never seen a multi-storey data centre. China: it scored 26 of 27 large Horinger candidates above 0.5; none is confirmed.
• Fuel burned on site, from TROPOMI NO₂ (tools/no2_plume_test.py, tools/no2_flux.py, tools/no2_flux_quarterly.py). The plume test reached 9.8σ at Colossus 2 and 2.8σ at Abilene, whose turbines have emission controls; Abilene falls to 2.3σ if its start date moves one month earlier. Calibrated flux measures one campus, Colossus 2: 1,207 ± 248 kg NOx/h over July and August 2026, which is 637 to 2,919 MW at an assumed 0.5 to 1.5 kg NOx/MWh for its temporary turbine fleet. With 7 to 16 usable days a month, quarterly figures carry about ±20 % statistical uncertainty and single months ±20 to 50 %. The calibration factor, 4.41 ± 0.18, comes from five US tall-stack coal plants with ±19 % plant-to-plant scatter; gas-turbine references are not yet qualified, so the megawatt range is too wide to give a utilisation ratio. None of 31 campuses without known on-site generation reached 2.5σ, and none of 62 control points did. China: no public hourly plant data to calibrate against, and no hub campus is known to generate on site.
• Operator-reported electricity (tools/ingest_disclosures.py). Meta's Environmental Data Index gives annual electricity for 18 campuses from 2011 to 2024; the 2024 figures appeared in October 2025. The site shows the 2024 average and carries it into 2025 and 2026 with a wider band. A utilisation ratio needs a capacity for the same years, which only Luleå has: 0.25 to 0.45 of its 120 MW supply in 2022 to 2024. New Albany's 250 MW connection applies only from 2026, so it cannot be compared with the 2019 to 2024 loads. China: no operator publishes electricity per campus.
• Measured supercomputer power. ORNL reports Frontier's average power: 11.4 MW in 2022 and 12.2 MW in 2023, about 0.54 of its measured peak of 21.1 and 22.7 MW. China: the two national supercomputers in the inventory have measured peaks only.
• Commercial utilisation from listed operators (docs/cn_operator_disclosures_notes.md). VNET reported 70.1 % of 889 MW in service at the end of 2025 and 73.9 % of 1,007 MW in mid-2026; GDS 75.5 % by area at the end of 2025; Chindata 80 % in mid-2023, its last public filing. These are company-wide shares of capacity customers use or have contracted, not how hard the equipment runs. Chindata's filings also give each data centre's capacity in service and capacity in customer use (tools/chindata_filings.py). In its Greater Beijing Area that was 466 of 517 MW across 18 data centres at the end of 2022, up from 287 of 399 MW across 13 a year earlier. New halls filled within about a year: CN11-C went from 8 to 67 of its 71 MW. The filings give no city for most data centres; Chindata's releases place five of them in Hebei, Shanxi or Tianjin. US: not yet collected.
• Load derived from water use (tools/ingest_water_loads.py). Google publishes per-campus water but no water efficiency figure; its own totals imply 1.02 to 1.06 litres per kWh. Checked against Meta's electricity in 2024, the method puts the middle half of mature campuses at 0.6 to 1.6 times their reported load, individual campuses from 0.35 times (Luleå, air-cooled) to about 3 times (Clonee and Odense), and first-year campuses at 10 to 13 times. 12 Google campuses carry a 0.5 to 2 times band, widened to 0.4 to 2.5 times when a year's figure is carried forward. China: none published.
• Night lights from VIIRS (tools/ntl_timeline.py). Campus radiance rose above its surroundings during construction at all four US greenfield campuses tested (Rainier August 2024, Fairwater October 2024, Abilene December 2024, Hyperion December 2025) and at the converted Colossus 1 factory six months after its documented load began. No step appeared at 19 other sites, including the Colossus 2 turbine yard. It tracks construction and site lighting, not operation. As a regional scan, the nearest new lit area lay within 0.4 km of four of seven US campuses, 1.2 km and 2.3 km of two more, and Colossus 1 was missed. China: fails inside parks that were already lit.
• Optical change scan from Sentinel-2 (tools/s2_candidates.py). Recovered 3 of 3 known US campuses; 11 of the 23 US components it flagged were not halls. China: 7 candidates in the Ulanqab and Zhangbei boxes; on inspection each was an unattributed roof or industrial complex, and none showed data-hall evidence.
• Generator watch list (tools/generator_watchlist.py). Two filing-located turbine fleets, Fermi Matador and Cheyenne's Project Jade, have NOx series from January 2024 and eight monthly screens each after a 2024 to 2025 baseline. No month has crossed 2σ and 100 kg/h; the largest monthly z-scores are 1.43 and 1.77. Seven more fleets need records. US only.
• Plant and water records (tools/campd_monthly.py, tools/water_monthly.py). No EPA plant record yet qualifies as a campus's load; the Southaven plant beside Colossus 2 stays evidence only. Of eight cities searched, one published an attributable campus water series: Bluffdale's 312 monthly deliveries to the NSA's Utah Data Center. US only.
• Chips without credentials. tools/chip.py renders a Sentinel-2 true-colour chip for any point from AWS open data.

Inconclusive

• Six eastern Chinese hub boxes (Zhongwei, Zhangjiakou and Huailai, Wuhu, Shaoguan, Chengdu Tianfu, Tianjin Wuqing): 76 candidates of 5 ha or more, none scored as hall-like. The box centres were approximate and the classifier has not seen multi-storey buildings, so this does not show there is nothing to find. See RFW-02 and RFW-11.
• Abilene's plume. 2.8σ at the documented July 2025 start and 2.3σ a month earlier, with emission-controlled turbines whose flux is too small to convert to megawatts. See RFW-14.
• Gas-turbine calibration plants. Three candidates gave calibration factors from 0.89 to 4.83; stack heights, isolation and overpass alignment are unverified. See RFW-16.
• NO₂ from power plants next to Chinese hubs. Usable as an activity index only at Ulanqab; the Horinger and Chongqing series were not usable.
• Chinese plants linked to hub parks (tools/cn_stack_monitors.py, docs/cn_stack_monitors.md). A first pass found 14 plant or supply leads across 11 hubs; four are renewable projects. The Shengle plant supplies the Horinger cloud park, and its annual emission reports give 63 records for 2019 to 2025, but no allocation of its output to any campus is established. No hourly or daily stack readings were retrieved: Shengle's reading service requires a CAPTCHA and other candidate services timed out or returned errors. See RFW-12.
• Radar-found structures in China. Hall-like and dated, but unconfirmed. See RFW-01.

Did not work

Please do not repeat these without a new idea.

• Roof temperature as a load signal. At night, ECOSTRESS (70 m) showed no step at documented load changes, adjusted for season and weather: Abilene +0.16 ± 0.23 K at 174 MW and +0.13 ± 0.47 K at 522 MW, Rainier +0.06 ± 0.24 K at 1,078 MW. Daytime steps (Rainier +1.44 ± 0.60 K, Fairwater +1.35 ± 0.49 K) coincide with roofing and fit-out and cannot be separated from them; Landsat (100 m, daytime) gave no usable load signal either. Colossus 1's hall warmed by 1.24 ± 0.28 K at night, and its whole block, turbine yard included, by about the same. As a detector, nine ordinary roofs spanned −1.5 to +2.0 K at night against −0.75 to +1.2 K for twelve data centres. Transformer-scale heat is untested: at 70 m, the NSA Utah site's tentatively identified switchyard, about three pixels, read +0.25 ± 0.07 K at night over 299 frames, no different from the halls beside it (+0.19 ± 0.05 K), and that site's documented load did not change. See docs/overnight_report_2026-09-13.md.
• Snow persistence on roofs. At eight of nine sites over eight winters, operating halls held as much snow as the surrounding built-up land or more. At New Albany they held less in 2024 to 2026 (0.64 to 0.78 against 0.77 to 0.91) while new halls were being built on the campus.
• NO₂ at campuses without on-site generation. None of 31 reached 2.5σ; the highest was 2.09. None of 62 control points reached 2.5σ either.
• A campus beside large power plants. Colossus 1's turbine phase came out at 62 ± 63 kg NOx/h after regressing out the two neighbouring plants' hourly EPA emissions over 594 days.
• The box flux method at a city's edge. At Dublin's Grange Castle it gave 964 ± 96 kg NOx/h with winter peaks and a fitted source 36 km downwind, consistent with the city's plume; it cannot isolate the campus plant. See RFW-15.
• A classifier trained on OpenStreetMap footprints. On radar candidates its AUC was 0.70, and it scored 864 of 1,133 new industrial structures above 0.9.
• A national night-light scan as a detector in China. Over 24 to 43°N and 102 to 123°E it found 4,026 newly lit areas. The nearest ones to the Horinger and Ulanqab campuses ranked 344th and 2,808th, and Zhangbei had none within 40 km. In the 40 brightest campus-sized areas, 35 held a new structure of 5 ha or more; only 4 of their 249 large structures scored as hall-like, and the chips inspected showed industrial sites.
• Dark-roof dating from vegetation and built-up indices. At Hyperion it dated the north structure to June 2026, although its roof was visible by January, and the south structure falsely to September 2023. See docs/dark_roof_validation.md.