The Compute Commitment Register

Method, Challenge List Reference and Pilot Observations

Warmth Engine Observatory — Methodology Working Note

Warmth Engine

September 2026 | Version 1.0

CANONICAL SOURCE DECLARATION

This work is the authoritative version published at DOI: 10.5281/zenodo.22785950

The Compute Commitment Register, the Challenge List Reference and the Challenge Attribution relation are articulated here for the first time. The priority claim attaches to the articulation of these constructs and to the terms that name them, not to the underlying mechanisms — public disclosure of computing allocations, institutional risk lists, and the aggregation of lists by agreement count all pre-date this note. The finding that no instrument of the register's form was located is a dated search result.

The canonical version is the Zenodo publication. Derivative works, summaries, or adaptations should cite this canonical source.

Author: Warmth Engine | [email protected] | warmthengine.com

NOVEL TERMINOLOGY — PRIORITY DOCUMENTATION

Originated by: Warmth Engine

DECLARATIONS

The author operates the Warmth Engine Observatory and performed every verification recorded in §IX.3. The work is unfunded. Instruments were located with the assistance of research models, and every record in the register was located through Anthropic's model. Anthropic is an actor of record in the register (§IX.2), and the Observatory's only relationship with it is as a user of its models.

ABSTRACT

The Compute Commitment Register records, for any named actor, publicly documented commitments of computing capacity to a stated societal challenge, and records whether the instrument states that the capacity is for AI. Each record is one instrument, such as a budget line, an allocation record, or a filing. It is counted once, and its declared quantity is shown as published and never converted or summed.

Records are assigned to the Challenge List Reference. The Reference is derived from institutions whose published challenge lists meet a qualifying definition: a challenge enters when a threshold number of those institutions name it. The Reference is fixed for an epoch. At this first epoch it holds nine challenges drawn from four institutions.

The register is a layer in the Observatory's Analytical Layer Model. It touches the event corpus at one point: the Challenge Attribution relation, which is recorded only where an event's own source is a record's instrument.

This note publishes the method, the first Challenge List Reference with its run table and epoch record, the rules fixed during a three-phase pilot, the register's first observation, and the results of a four-model prior-art search. At publication the register holds eleven records from nine actors, nineteen document–challenge pairs, and an attribution edge of one record to one event.

The register's limits include a disclosure base concentrated in public allocation programmes; a challenge list drawn from three European-authored institutions and the United Nations; no energy item; and counts that are floors. As of the date of search, no instrument was located that records verified, actor-level commitments of AI compute against a challenge taxonomy. The nearest by subject is an applicant-facing directory of open opportunities. The nearest by method is a single-country promise tracker.

Keywords: compute commitments; compute allocation; societal challenges; challenge taxonomy; institutional risk lists; evidence standard; register methodology; AI for good; public research computing; coordination intelligence

OPERATIONALISATION STATUS

TABLE OF CONTENTS

ABBREVIATIONS

I. PURPOSE AND POSITION

I.1 The question

Governments, agencies, research programmes, and corporations state commitments of computing capacity to climate, health, food security, and other global challenges. Money directed to challenges is tracked by development-finance and climate-finance datasets. Compute supply is tracked by model databases, system rankings, and cloud-availability measurement. A public allocation programme that reports awards reports only its own. The search described in §X located no instrument that records, across actors, which bodies have committed compute to which challenges, and on what evidence.

The Compute Commitment Register records this from the documents themselves.

I.2 Position in the Observatory

The WEO Methodology Manual's Analytical Layer Model (§2.10) assigns each layer a unit of analysis and a defining chapter. The Compute Commitment Register is added as a layer to that model, and its unit of analysis is the record (§II). Adding or removing a layer does not modify the processing sequence architecture or the classification system count. The register changes nothing in the Observatory's other layers, including tier classification, domain classification, industry tags, and Coordination Connections. Every record is produced under the evidence standard set out in §III. The register touches the corpus at one point, the Challenge Attribution relation (§VII).

I.3 Scope of the measure

The register counts instruments and shows declared quantities as published. It is a record of presence, not magnitude: which actors have documented which commitments to which challenges. It does not weight actors or commitments, and it does not measure how much compute has gone to any challenge or how much of it has been delivered.

II. WHAT THE REGISTER RECORDS

A record is one publicly documented instrument in which a named actor commits computing capacity to a stated purpose that maps to a candidate in the run table of the Challenge List Reference. The run table is the full table of candidate challenges considered for the Reference, above and below its threshold (§V.5).

Each record carries AI compute: stated where the instrument states that the capacity is for AI, whether in the programme's name as printed, its mandate, or the commitment's wording. Otherwise it carries not stated. The value is transcribed from the instrument at record level. At publication, AI compute is stated in eight of the eleven records (§IX.2).

A record has two evidence slots. The first holds the commitment instrument. The second holds a document evidencing award, disbursement, or use, where one exists. It may be the commitment instrument itself or a later document. Status is Committed where only the first slot is filled, and Allocated where both are. Further documents may attach to a record as corroborating evidence (§III.5).

Any actor is admissible, including governments, agencies, supranational bodies, public research and computing centres, corporations, and foundations. Admission is decided by what the instrument states, whatever the actor's type.

III. EVIDENCE STANDARD AND ENTRY GATES

A source document is assessed in four steps: the recording rules (§III.2), the match test (§III.3), the two entry gates (§III.4–§III.5), and assignment (§VI). The evidence standard (§III.1) applies throughout.

III.1 Evidence standard

The register's evidence standard derives from the WEO Methodology Manual's source hierarchy (§3.1). It is narrower than the Manual's standard in what may corroborate a record and wider in the document types it admits (§III.4). Every document the register holds as evidence must sit on the issuing body's own domain or in an official repository, whether it is the commitment instrument, delivery evidence, or corroborating evidence. Reproductions, news coverage, and aggregators may locate a document. They never serve as evidence, even where the Manual admits them as corroborating at event level. Commitment evidence is further limited to the instrument in which the commitment is made, and that instrument must belong to one of the classes in §III.4.

Every evidence address is archived at entry, with its access date. The commitment passage and the purpose wording are copied into the record exactly as they appear in the archived copy of the source, with a marked English translation where the original is not English.

III.2 Recording rules

Five rules set passages aside before the match test:

• statistics describing capacity that already exists are not commitments;
• funding for research into computing is not funding for computing capacity;
• regulatory text governing compute is not a commitment of compute;
• electrical megawatts are not computing capacity; and
• a sector named only as part of the actor's market or customer base is not a purpose.

III.3 Match test

A passage in an instrument passes the match test where all three hold:

• a named body makes the commitment;
• the object committed is computing capacity, measured as the source states it: hours on a system, accelerator counts, dedicated capacity in petaflops or exaflops, cloud credits, a quota on a named system, or money earmarked for compute (§III.5); and
• the stated purpose maps to a candidate in the run table.

A purpose that maps to no candidate in the run table is logged as matching no candidate and held for the refresh check (§V.6).

III.4 Instrument-class gate

In Version 1.0 of the Register Standard, ten classes of document pass the gate:

1. budget line or appropriation;
2. allocation call with published awards;
3. allocation call with awards pending;
4. allocation or utilisation report;
5. regulatory, exchange, or statutory filing;
6. assured or audited sustainability or impact report;
7. parliamentary or legislative record;
8. grant or credit programme with a published list of recipients or of contributors;
9. executive order or equivalent; and
10. public procurement notice.

Class 5 admits documents a company files with a securities regulator. It also admits documents a company furnishes to the regulator, which, under United States securities law, carry a lighter legal liability for misstatement than filed documents. A furnished document is admitted because the company wrote it, it is dated, and it sits on the regulator's own repository. It is recorded with the sub-label "furnished exhibit". The register's one such record (§IX.2) is a press release announcing quarterly financial results, which the company furnished to the securities regulator as an exhibit to a Form 8-K. The same text published only as a press release does not qualify.

Class 6 admits only the audited part of an assured report.

The following do not qualify: press releases, blogs, keynotes, interviews, news articles (including state media reprinted on a government portal), institutional web pages, strategies and declarations without a funded instrument, unfunded memoranda, supplier marketing, and third-party monitors.

The class gate applies to commitment evidence only. Delivery evidence may come from any document that meets the evidence standard in §III.1.

III.5 Quantified-in-instrument gate

The instrument must state, in its own text, how much is committed. The object committed is computing capacity. The figure that measures it may be expressed as compute, money, capacity, or duration. A figure of any size qualifies. The register stores the figure exactly as declared, with its unit. It is never converted into another unit (dollars into accelerator-hours), scaled to a common measure (exaflop-equivalents), or summed with other records into a total.

The figure must measure compute. A figure already stated in compute terms, such as cloud credits, accelerator hours, or an accelerator count, qualifies on its own. A sum covering several things, only part of which is compute, does not qualify on its own: a national AI mission's appropriation covering seven components with no separate compute line is not a record.

A mixed sum qualifies under the compute sub-line rule where the actor has stated that the sum, or a named part of it, is compute. The statement must appear in another document on the actor's own domain that meets the evidence standard. That second document identifies the compute component and is attached to the record as corroborating evidence. The register's one such record (§IX.2) is a commitment of sixty million US dollars stated in a Form 8-K, whose compute form — cloud credits — is stated in the actor's sustainability report.

Corroborating evidence is a document that identifies the compute component, as above, or restates the commitment (§IV). A corroborating document repeats or details a commitment already recorded. A corroborating document is not independent confirmation of the commitment.

IV. RECORD RULES

One record per instrument. A programme with several instruments has several records. A later instrument that restates an existing commitment with no new quantity, scope, or period is not a new record. It attaches to the existing record as corroborating evidence. A later instrument that changes the quantity, scope, or period is a new record. Whether a document restates or changes a commitment is decided by comparing the two texts, not by their dates. An undated restating document is recorded as undated, and the record's own instrument date is the only date the commitment carries.

Actor of record. The actor of record is the body that makes the commitment, normally the instrument's author. For an allocation record, in which one body distributes resources across many awards, this is the allocating body. The resource named on each award is carried on the line item exactly as printed. Contributors are not inferred from a resource's name.

A contributor appears as an actor of record only where an instrument records that contributor's own commitment — for example, a programme's published list of contributor commitments, a filing, or an assured report. Each contributor's commitment in such an instrument is its own record, even where the programme authored the instrument. The WEO Methodology Manual's Distinct Actor Requirement (§9.1.6) applies: a corporation and its foundation, or a ministry and its agency, count as one actor unless separate governance is documented.

Participants. Bodies other than the actor of record that the instrument itself names as author, funder, operator, contributor, host, or co-signatory are recorded as participants. They are transcribed from the instrument and never inferred.

Dates. The instrument's own time-frame wording is stored as printed. A start or end date is recorded beside it only where the instrument states one. A phrase such as "by 2030" is kept as wording and is not converted to a date.

Line items. A line item is one purpose-bearing entry within a record. Where an instrument enumerates awards with individual purposes, each award is a line item, carrying the quantity field the instrument prints for it in whatever form. Where an instrument commits one system to several named challenges, there is one line item per challenge, each showing the same declared quantity. Line items are never counted in the document figure (§VIII).

Bands. Each record carries one of four bands describing the form of the commitment. The bands are never weighted.

• A — a grant or credit programme the actor awards to applicants it selects;
• B — a dedicated allocation or capacity reservation on a named system, including credits or capacity reserved for allocation by a public programme;
• C — sovereign or multilateral infrastructure; and
• D — commercial or philanthropic supply outside a public programme.

A credit commitment is A where the actor awards the credits to applicants it selects, and B where the actor reserves them for allocation by a named system or public programme. This rule applies to credits only: an allocation of compute time on a named system is B whoever selects the awardees.

Removal. A record is removed only when its instrument is shown to be withdrawn, misattributed, or to fail a gate on re-examination. Removed records persist archivally. A removed instrument re-enters only as a fresh record.

V. THE CHALLENGE LIST REFERENCE

V.1 Principle

The Observatory records where referenced institutions agree on the world's critical challenges, under a qualifying definition and a merge protocol.

V.2 Qualifying definition

An institution's challenge list qualifies when all six hold:

• recurring, with at least two editions on a stated cycle;
• issued by an institutional author — intergovernmental, governmental, scientific-assessment, or an established research institution with published governance; for-profit consultancies are excluded;
• publicly dated and openly accessible;
• cited by at least two independent intergovernmental or national-government institutions in authored documents;
• global and cross-domain in scope; and
• presented as an enumerated list rather than narrative.

Catalogue records, repositories, and hosted third-party documents do not count as citations. Funding or commissioning by a qualifying body is not citation.

V.3 The search and the qualifying set

The inventory of challenge lists was run across three research models in seven languages (English, Chinese, Japanese, Korean, French, Arabic, and Russian). Forty lists were inventoried.

Russian- and Arabic-language institutional outlooks were sought, and none was confirmed in the searches run.

The non-English lists the search found are national science-and-technology priority lists:

• the China Association for Science and Technology's annual frontier questions;
• the Chinese Academy of Sciences' long-horizon roadmaps;
• the National Institute of Science and Technology Policy's Foresight Survey for Japan; and
• the Korea Institute of Science and Technology Evaluation and Planning's Ten Future Promising Technologies.

This is a different genre from the global cross-domain risk list, and each failed on scope. The Chinese Academy of Social Sciences' International Situation Yellow Book is global in subject and failed on enumerated form: its content is organised as thematic and regional chapters.

Four institutions qualified:

• the World Economic Forum, for the Global Risks Report 2026 (33 risks);
• the Munich Security Conference, for the Munich Security Index 2026 (32 risks);
• the United Nations, for the 17 Sustainable Development Goals as tracked by the annual Sustainable Development Goals Report; and
• the European Commission Joint Research Centre, for the Megatrends Hub (14 megatrends).

The Global Risks Report 2026 was re-issued between late January and early March 2026 with typographical corrections, not as a new edition. The report's Appendix A, the source of the 33 risks, is identical across both issues. The Munich Security Index list is taken from the G7 aggregate chart. The chart for the other surveyed countries names the same risks, and the merge protocol counts naming, not rank, so the choice of chart does not change the result.

Two lists failed only the citation criterion: the Global Challenges Foundation's Global Catastrophic Risks report, at one qualifying citation of two, and The Millennium Project's 15 Global Challenges, at none. Both are re-tested at the first refresh. A list this search did not find is tested against the same definition at the next refresh after it is identified.

V.4 Aggregation rule

A challenge enters the Reference when at least three of the four institutions name it (k = 3 of N = 4). At k = 2, two institutions whose lists each run to more than thirty items would satisfy the rule on their own. The same lists at k = 2 yield 21 challenges against 9.

Whether an institution names a challenge is decided under the merge protocol:

• an item on an institution's list matches a candidate when its headline noun names the same hazard;
• goals and trends are read as the challenge they answer or describe, and this inversion is recorded per item;
• one item matches one candidate, by its first noun;
• items that are instruments or mechanisms rather than challenges are recorded as non-matching; and
• institutions are counted, not items.

Seven of the nine published challenges take their names from a single item on the World Economic Forum's list. In three of the seven, the challenge holds two items on the Forum's list, and the name comes from the item that the other institutions' items most closely match. One name joins two items on the Forum's list with "and": "State-based armed conflict and geoeconomic confrontation". One name is not an item on the Forum's list: "Climate change" is the Forum's defined structural-force term, and two other institutions use it in their items. Names are taken from the institutions so that the Observatory authors none: every word in every challenge's name appears in an institution's item. Only the World Economic Forum's list has an item under every published challenge (Appendix A).

V.5 Challenge List Reference Version 1.0, Epoch E1

Nine challenges, in fixed order: descending institution count; within a count, the Forum's two-year severity rank where published, then the Forum's own table order for unranked items. The right-hand column carries the register's first observation (§IX.2).

All 96 items across the four institutions are accounted for in the run table. Twelve candidates fell at exactly two institutions: misinformation and disinformation, societal polarisation, cyber insecurity, weapons of mass destruction, economic downturn, crime, erosion of human rights, energy supply, discrimination, education, urbanisation, and consumption. Nine fell at one. Four of the nine challenges are named by exactly three institutions. Three of those four reach the threshold through a Joint Research Centre megatrend read as a challenge under the inversion rule, and a change to that institution's edition can take them below the threshold.

The run table, listing every candidate, every source item, and every inversion, is given in Appendix A. Every candidate below the threshold — a sub-threshold candidate — remains in the mapping vocabulary (§VI). For example, a commitment to energy supply or cyber security is a full record: it is assigned to its candidate and held out of the published table until that candidate reaches the threshold.

V.6 Epoch discipline

An epoch is the period during which the Reference and the editions it rests on are unchanged. The Reference is epoch-stamped.

A refresh check runs each 1 October. It re-applies the qualifying definition and the threshold to the current editions of every list.

Any change to the definition, the institution set, the threshold, the merge protocol, or an institution's edition opens a new epoch. Before a new epoch's first observation, one record set is observed under both the old and the new Reference. No trend is drawn across an epoch boundary. The Sustainable Development Goals expire in 2030, a foreseeable epoch break.

VI. ASSIGNMENT

The mapping vocabulary is the full run table. A purpose becomes a line item under a candidate only where its own wording names that candidate's hazard as the problem the compute is directed at. A sector named as the destination of the compute is read as naming that sector's hazard, however generally it is stated: "to advance global health" is assigned to decline in health and well-being. Where the purpose names an AI capability, method, benchmark, or course, and a domain only as an application, no candidate is assigned. An empty purpose field is recorded separately as stating no purpose.

Privacy work is assigned to erosion of human rights where protecting individuals' privacy, or preventing surveillance or rights harm, is the stated purpose. It is not assigned where privacy is a constraint on a method whose purpose is something else. Classroom allocations are assigned to education only where an access or skills deficit is named. Where a purpose names several hazards, the line-item rule in §IV applies.

VII. THE CHALLENGE ATTRIBUTION RELATION

A Challenge Attribution links a corpus event to a register record. The test is document identity: the event's own source must be the record's instrument. A related document from the same programme (for example, a guidance page where the record's instrument is a call) does not qualify.

Nothing enters the corpus through the register. Where a record's instrument later enters the corpus as an event in its own right through the Observatory's rolling-update protocol, the attribution follows. Where an event's source has been superseded, the earlier source can still ground a record provided the superseding source does not replace the commitment it states.

The relation is a new class, distinct in vocabulary from the Observatory's other relations, including Coordination Connections (evidence-based, typed relationships between verified events) and Infrastructure Threads (factual links tracing named assets to origin events). On the event it is metadata — a note that a register record exists — and changes no classification. On the record it marks the record as corpus-grounded. The set of all attributions is the attribution edge.

VIII. MEASURE

VIII.1 The document-basis challenge table

Three figures are computed for each published challenge:

• the number of documents touching it, each record counted once under every challenge it touches;
• the number of line items; and
• the Committed and Allocated document counts.

While the register is pre-operational (§VIII.3) only counts are published. When that status changes, the document count is published as a share of all document–challenge pairs.

Each observation also names the record supplying the most document–challenge pairs, with its count, and reports four further counts:

• records assigned to more than one candidate, published or sub-threshold;
• records assigned only to sub-threshold candidates;
• passages matching no candidate; and
• passages stating no purpose.

The document figure is the primary measure, so that a single allocation record with 847 awards does not dominate the observation.

VIII.2 Record status

No ratio of Allocated to Committed records is published. In allocation records the pledge and the delivery are one act, so those records enter as Allocated. At publication four records are Allocated: in three the instrument is itself the delivery evidence, and one carries separate delivery evidence (§IX.2). A ratio across the register would therefore sort records by instrument class and measure the presence of delivery reporting rather than delivery.

VIII.3 Actor table, floors, and register status

The three figures of §VIII.1 are computed per actor of record once the actor holds five or more records. Below that floor only the actor's record count is published, with the status insufficient_corpus.

The register as a whole carries the status pre_operational until the allocation records of the following programmes have been assessed and the first Challenge List refresh check (§V.6) has run. The programmes are the full EuroHPC awarded-projects list, the Department of Energy's ALCC awards, and the allocation reports of national computing centres (France's Grand équipement national de calcul intensif, Australia's National Computational Infrastructure, and Italy's CINECA consortium). The register's status does not affect actor tables.

VIII.4 Observations

Every published table is a dated observation, epoch-stamped to the Challenge List Reference and comparable only within its epoch.

VIII.5 Coverage acknowledgement

The register's figures are published under the following coverage acknowledgement. "The figures describe compute commitments the Observatory has located and verified against this standard, not the population of commitments that exist. Counts are floors. Coverage expands as instruments are found. A record's presence carries no information about its scale beyond its band; declared quantities are shown as published and are never compared or summed. The challenge list is a dated observation of four institutional sources and changes only at an epoch boundary."

IX. PILOT — DATED OBSERVATIONS

IX.1 Phase A: the corpus (7 September 2026)

The sources of all 214 corpus events were swept: 1,468 source objects, of which 44 were unreadable. The sweep found 54 passages in which a named body commits compute, on 40 events. By purpose, 22 mapped to a published challenge, four to a sub-threshold candidate, and 28 to no candidate. Of the 54, seven passed the class gate under the initial class list and 42 the quantified-in-instrument gate. Three passed both, and none of those three named a challenge.

Across the corpus, instruments that quantify (such as Command Papers, Council Regulations, and budget lines) state purposes such as "the resolution of science problems", which map to no candidate. The register's population therefore comes from allocation-record instruments, which the corpus rarely cites. The corpus contributes the attribution edge and calibrates the gates.

Four instrument classes were added from this phase. Each covers a document type that the corpus produced and for which the initial list had no class:

• an executive order or equivalent;
• an allocation call with awards pending;
• a public procurement notice; and
• an institution's published statement of its standing resources, since withdrawn (§IX.3).

The line-item rule for multi-purpose instruments (§IV), the allocator/contributor rule (§IV), the five recording rules (§III.2), and one band were fixed from the same phase.

IX.2 Phases B and C: external instruments (8–9 September 2026)

Phase B assessed three sources:

• a three-model disclosure inventory (95 entries merged to 88);
• the 54 corpus passages, re-assessed under the Register Standard; and
• the United States National AI Research Resource pilot's allocation export (1,163 award-resource rows over 847 awards) with its published contributor-commitments list (28 lines).

Phase C examined eleven further instruments, of which one entered as a record.

Register at publication. Eleven records from nine actors, in four groups:

• five public-programme records: two United Kingdom compute instruments, the United States Department of Energy's leadership-computing awards, the National Science Foundation's allocation export, and a EuroHPC AI Factory;
• one corporate open-call record: Microsoft's AI for Good compute call;
• four corporate-contributor records (Anthropic, Microsoft, OpenAI, and Vocareum), evidenced by the National Science Foundation's published contributor list; and
• one corporate exchange-filing record (class 5, furnished exhibit): Amazon's Health Equity Initiative, with the commitment stated in a Form 8-K exhibit as sixty million US dollars.

Amazon's exhibit states an increase of twenty million US dollars, bringing the commitment since 2021 to sixty million. The register carries the total the instrument states. The Amazon record entered under the compute sub-line rule (§III.5), which was fixed in Phase C from this record. The actor's sustainability report states the compute form of the sum, cloud credits, and attaches as corroborating evidence. A separate passage of the same report attaches as delivery evidence. A filed proxy statement restates the commitment and attaches as corroborating evidence.

The register's figures at publication:

The National Science Foundation's allocation export supplies eight of the nineteen document–challenge pairs and 303 of the 321 line items, and touches eight of the nine published challenges. Without it the table holds three challenges: health with seven documents, adverse outcomes of AI with two, and climate change with two.

Further observations from the external instruments:

• Seven contributors to the National AI Research Resource pilot committed quantified compute (up to four exaflops, twenty-four million US dollars of cloud capacity, a million accelerator-hours) and named no purpose that maps to any candidate.
• At the date of search the United States' national AI compute programme had published 847 named awards. For the United Kingdom's programme the search found no published award list.

IX.3 Verification

The assignment rule (§VI) was first applied, with model assistance, to 673 free-text award abstracts. The Observatory audited it on a seeded random sample of thirty awards: ten assigned to a candidate, ten logged as matching no candidate, and ten logged as stating no purpose. The audit agreed with the assignment in 28 of the thirty. Both disagreements were assignments under the privacy convention, which was corrected, and the rule was then fixed in the Register Standard. Three successive wordings of the privacy convention moved 23 awards out and 13 back. No published figure changed, because the candidate is sub-threshold.

Every record's commitment passage, every delivery passage, and every attribution was verified against its source on 8–9 September 2026. Verification here means that the passage as stored matches the source as archived.

A second seeded sample of twenty line-item assignments was checked during verification. All twenty agreed. The second sample is drawn from assignments made: it tests whether an assignment was correct. Whether a purpose logged as matching no candidate should have mapped to one was tested only on the ten such awards in the first sample.

One attribution was removed at verification: the event's source and the record's instrument were related documents of the same programme, not the same document. The record remains.

One record was removed on re-examination on 16 September 2026 under the removal rule in §IV. Its instrument was the recipient institution's own description of the resources it had been given. No named body makes a commitment in that document, so it fails the first leg of the match test (§III.3), and the first recording rule sets it aside as a statement of existing capacity (§III.2). It had entered under a class that was drafted during Phase A from this instrument and has since been withdrawn. Its attribution to a corpus event was removed with it.

Two other records entered under rules generalised during the pilot. In the first, the instrument is a dated gov.uk publication by the committing department that sets out the resource and the route to apply, admitted as an allocation call with awards pending. It came from the corpus sweep and is the register's one corpus-grounded record. In the second, it is the exchange filing, admitted under the compute sub-line rule. Both pass those rules as written and remain.

The eleven records hold twenty-one evidence attachments (eleven as commitment evidence, four as delivery evidence, and six as corroborating evidence) at eleven distinct addresses. One address is held as a copy of April 2026, and three are undated programme pages, fixed as of the archive date.

The eight documents behind the four institutions of the Challenge List Reference are archived on the same terms.

IX.4 What did not enter

At close, eleven instruments from all phases of the pilot were held back for possible future records, two of them because they state a mixed sum and no statement identifying its compute component has been found (§III.5).

X. RELATED WORK

A prior-art search was run across four research models, followed by an examination of the closest comparators, those named in the table below. It located no instrument that records verified, actor-level commitments of AI compute against a challenge taxonomy, and no method that derives a challenge taxonomy by a threshold over the lists of qualifying institutions. Each comparator is described from its own pages. The instruments located fall into six families, each lacking an element the register has.

Opportunity directories. ImpactOpen, built and maintained by TechChange, indexes open AI-for-good funding, accelerators, fellowships, credits, and challenges for applicants, sorted by closing date, with focus areas and hand-verified deadlines. Each entry links to the official source. Its unit of record is an opportunity open to applicants, not a commitment. It is the nearest instrument by subject.

Programme self-reporting. Each programme publishes only its own calls or awards. Four of the six supply nine of the register's eleven records (§IX.2).

Compute-supply measurement. Chinese-language policy instruments fall in this family. The National Development and Reform Commission's implementation opinions on the East Data West Compute project set a compute target and a green-electricity target for the national hub nodes. The research report on the green computing power index of the hub nodes, issued by the Tsinghua University Global Industry Research Institute with the Horinger Green Computing Power Development Research Center, assesses regions by index score. In both, compute is recorded and no purpose is stated as a challenge. The comparators are therefore not exclusively Western, though the two nearest instruments are.

Promise trackers. AI Canada Pulse tracks seventy commitments in one national strategy with an evidence model separating official promises from independent progress evidence. Its own description states that strategy promises are commitments, not proof of delivery. It is the nearest instrument by method: it uses the same two-slot logic as the register's status. It covers one country, and its unit of record is a promise in that strategy.

Aggregated taxonomies. Rakowski et al. (2025) cluster crises from a systematic review of 59 publications; Rees and Wilhelm (2025) merge the Sustainable Development Goals and two other frameworks into a taxonomy of planetary health knowledge in seven domains; Avin et al. (2018) classify global catastrophic risks; Oppenlaender and Hämäläinen (2023) extract research challenges from the papers of one conference with model assistance. All aggregate from literature or expert judgement. None applies a threshold to the lists of institutions passing a qualifying definition.

As of 8 September 2026, no instrument of the register's form was located.

XI. LIMITS

1. Disclosure base. At publication, nine of the register's eleven records rest on documents published by public allocation programmes (§IX.2). In the documents the pilot examined, governments state quantities in money or aggregate exaflops, and corporations state them in press releases, which are not among the classes that pass the gate (§III.4).
2. Regional concentration. Three European-authored institutions and the United Nations. No Asian, African, or Latin American institutional list qualified: the non-English lists found are national research-priority lists, a different genre, and failed on scope. The Reference measures agreement among these four institutions, and two of them survey perceived risk. Adverse outcomes of AI is a published challenge on that basis, and AI-safety compute is counted under it. The definition is published so that the regional concentration can be tested at each refresh. The programmes named in §VIII.3 are European, United States, and Australian public allocation programmes and their regional distribution does not offset this concentration.
3. Energy supply. Two of the four institutions name it. The threshold excludes it. Energy commitments are recorded, assigned, and held. They enter the table at the first refresh at which a third institution names energy.
4. Privacy and data protection. The run table has no data-protection candidate. The nearest is erosion of human rights, a sub-threshold candidate. A block of research in the National Science Foundation's allocation export (differential privacy, federated learning, unlearning, de-identification) is therefore assigned but unpublished, or logged as matching no candidate.
5. Programme purpose. 441 of the pilot's passages mapped to no candidate and 174 stated no purpose, against eleven records. Under the assignment rule, a purpose that names an AI method and a domain only as its application is logged as matching no candidate, so part of the 441 reflects the rule's strictness.
6. Asymmetry of publication. One actor's two records represent a programme for which the search found no published award list. Another actor's one record represents a programme that has published 847 awards, and that record supplies eight of the nineteen document–challenge pairs. Counts are floors of disclosure, not of activity.
7. Non-quantification. A record's presence carries no scale information beyond its band.
8. Attribution edge. At publication the edge holds one record and one event.
9. Maintenance. Open-ended intake, the epoch discipline of §V.6, and evidence archiving are recurring work the Observatory undertakes for the register.
10. Band D. The class gate excludes press releases, blogs, and keynotes (§III.4). A corporate or philanthropic pledge made only in such a document cannot enter, and at publication the register holds no Band D record.
11. Discovery circularity. The external intake was model-assisted: research models found the instruments. What the models did not find is not in the register, so its coverage reflects their reach.

XII. PROVENANCE

The figures in this note describe the register at publication. The Observatory maintains the register as a dated log, and every published figure is a dated observation reproduced from it.

The Register Standard (Version 1.0, 16 September 2026), the Challenge List Reference (Version 1.0, Epoch E1, 7 September 2026), the search records, and the verification records are held by the Observatory.

APPENDIX A. RUN TABLE AND EPOCH RECORD OF THE CHALLENGE LIST REFERENCE

A.1 Run table

The run table assigns every item on the four institutions' lists to a candidate under the merge protocol (§V.4). The nine published challenges come first, in the fixed order of §V.5; the sub-threshold candidates follow in descending institution count. Items are given by their headline wording, in the institution's spelling. Each Sustainable Development Goal is preceded by its official number. The Sustainable Development Goals and the Joint Research Centre megatrends are read under the inversion rule; where the reading differs from the printed wording, the reading is given after an arrow (→). One Munich Security Index cell combines three items: the use of nuclear, of biological, and of chemical weapons by an aggressor. The last row holds the items recorded as non-matching. Six of them are Munich Security Index items that name an actor (five states and the European Union) and no hazard.

Item accounting: World Economic Forum, 33 of 33 assigned; Munich Security Index, 32 of 32 (26 matched, 6 non-matching); Sustainable Development Goals, 17 of 17 (15 matched, 2 non-matching); Joint Research Centre megatrends, 14 of 14 (12 matched, 2 non-matching).

A.2 Epoch record

REFERENCES

AI Canada Pulse. Canada AI Strategy Promise Tracker. https://www.aicanadapulse.ca/strategy

AidData. Financing the 2030 Agenda for Sustainable Development, Version 1.0. https://www.aiddata.org/data/financing-the-2030-agenda-for-sustainable-development-version-1-0

Avin, S., Wintle, B. C., Weitzdörfer, J., Ó hÉigeartaigh, S. S., Sutherland, W. J. and Rees, M. J. (2018). Classifying global catastrophic risks. Futures. https://doi.org/10.1016/j.futures.2018.02.001

Bunde, T. and Eisentraut, S. (eds.) (2026). Munich Security Report 2026: Under Destruction. Munich Security Conference. https://doi.org/10.47342/JWIE5806

China Association for Science and Technology (2025). 十大前沿科学问题 (Ten Frontier Scientific Questions). https://www.cast.org.cn/xw/BWTJ/art/2025/art_0f20c8a62cfe4584b13a53271ae73837.html

Chinese Academy of Social Sciences, Institute of World Economics and Politics (2025). 国际形势黄皮书：全球政治与安全报告 (International Situation Yellow Book: Report on Global Politics and Security). http://iwep.cssn.cn/xscg/xscg_zzjyz/202501/t20250124_5842345.shtml

Climate Policy Initiative. Global Landscape of Climate Finance. https://www.climatepolicyinitiative.org/resources/data-visualizations/global-landscape-of-climate-finance-data-dashboard/

Department for Business, Innovation, Science and Trade, Department for Business and Trade, and UK Research and Innovation. AI Research Resource. https://www.gov.uk/government/publications/ai-research-resource

Epoch AI. Data on AI Models. https://epoch.ai/data/ai-models

European Commission Joint Research Centre. Megatrends Hub. https://knowledge4policy.ec.europa.eu/foresight/tool/megatrends-hub_en

European High Performance Computing Joint Undertaking. Awarded Projects. https://www.eurohpc-ju.europa.eu/supercomputers/awarded-projects_en

Global Challenges Foundation (2025). Global Catastrophic Risks 2026. https://globalchallenges.org/gcr-2026/

ImpactOpen (TechChange). The AI-for-good opportunity index. https://www.impactopen.org

Korea Institute of Science and Technology Evaluation and Planning. 10대 미래유망기술 (Ten Future Promising Technologies). https://www.kistep.re.kr/board.es?bid=0031&cg_code=C04&mid=a10306050000

Lehdonvirta, V. et al. (2025). Measuring domestic public cloud compute availability for artificial intelligence. OECD Artificial Intelligence Papers, No. 49. https://www.oecd.org/en/publications/measuring-domestic-public-cloud-compute-availability-for-artificial-intelligence_8602a322-en.html

National Artificial Intelligence Research Resource Pilot. Resource Allocations. https://nairrpilot.org/awarded-projects

National Development and Reform Commission and other departments, People's Republic of China (2023). 关于深入实施"东数西算"工程加快构建全国一体化算力网的实施意见 (Implementation Opinions on Deepening the "East Data West Compute" Project and Accelerating Construction of a National Integrated Computing Power Network). https://www.gov.cn/zhengce/zhengceku/202401/content_6924596.htm

National Institute of Advanced Industrial Science and Technology and AIST Solutions, Japan. ABCI 3.0 開発加速利用 (Development Acceleration Use). https://abci.ai/ja/acceleration_program/

National Institute of Science and Technology Policy, Japan. 第12回科学技術予測調査 (12th Science and Technology Foresight Survey). https://www.nistep.go.jp/archives/60706/

Oppenlaender, J. and Hämäläinen, J. (2023). Mapping the challenges of HCI: an application and evaluation of ChatGPT for mining insights at scale. arXiv:2306.05036. https://arxiv.org/abs/2306.05036

Rakowski, J. J. et al. (2025). Characterizing the global polycrisis: a systematic review of recent literature. Annual Review of Environment and Resources. https://www.annualreviews.org/content/journals/10.1146/annurev-environ-111523-102238

Rees, F. and Wilhelm, O. (2025). What you need to know about the world. Toward a taxonomy of planetary health knowledge. Frontiers in Public Health, 13. https://pmc.ncbi.nlm.nih.gov/articles/PMC12098363/

The Millennium Project. 15 Global Challenges. https://www.millennium-project.org/challenges-overview/

TOP500. The List. https://www.top500.org/

Tsinghua University Global Industry Research Institute and Horinger Green Computing Power Development Research Center (2024). 国家"东数西算"枢纽节点绿色算力指数研究报告 (Research Report on the Green Computing Power Index of National "East Data West Compute" Hub Nodes). https://www.igi.tsinghua.edu.cn/__local/7/04/E0/C396CD8BE0BC411C12959A0A8F7_9964D5FC_DB7F1.pdf

United Nations. The 17 Sustainable Development Goals. https://sdgs.un.org/goals

United Nations Statistics Division. The Sustainable Development Goals Report 2026. https://unstats.un.org/sdgs/report/2026/

United States Department of Energy, Office of Science. INCITE Program. https://doeleadershipcomputing.org

United States Department of Energy, Office of Science. ASCR Leadership Computing Challenge. https://science.osti.gov/ascr/Facilities/Accessing-ASCR-Facilities/ALCC

World Economic Forum (2026). Global Risks Report 2026, 21st edition. https://www.weforum.org/publications/global-risks-report-2026/

Warmth Engine Observatory documentation

Warmth Engine (2026). WEO Methodology Manual, Version 1.5. Zenodo. https://doi.org/10.5281/zenodo.18427565

Warmth Engine (2026). WEO Methodology Rationale, Version 1.5. Zenodo. https://doi.org/10.5281/zenodo.18427582

Warmth Engine (2026). Towards Coordination Science. Zenodo. https://doi.org/10.5281/zenodo.18427584

Warmth Engine (2026). Tier-Crossing Dynamics in Coordination Networks. Zenodo. https://doi.org/10.5281/zenodo.18427586

Warmth Engine (2026). WEO Contribution Architecture. Zenodo. https://doi.org/10.5281/zenodo.21495917