How the research is done
Methodology
A reader who disagrees with a Frontier Labs conclusion should be able to locate precisely where they disagree. This page describes how we get from a question to a published finding, and what we commit to at every stage.
The research process
Six stages, applied without exception.
- 01
Research
A question is specified precisely enough that it could be answered wrongly.
Work begins with a written question, a statement of what evidence would answer it, and an explicit account of what the answer would change. Questions that cannot be stated in a form that admits of being wrong do not enter the programme. At this stage we also record the prior view of the research group, so that the distance travelled by the evidence is visible in the final publication.
Outputs- Question specification
- Evidence plan
- Recorded prior
- 02
Verification
Every quantitative claim is traced to a primary source and the source is assessed.
We work from primary material: regulatory filings, statutory instruments, interconnection registers, customs records, planning submissions, published microdata and disclosed technical documentation. Secondary sources are used to locate primary material, never as evidence in themselves. Each source is assessed for provenance, collection method, known bias and revision history, and that assessment is recorded alongside the datum.
Outputs- Source register
- Provenance assessment
- Revision history
- 03
Cross validation
Independent methods are used to check whether they reach the same answer.
Wherever a quantity can be estimated by more than one route, we estimate it by more than one route. Compute capacity is derived independently from electrical records, from construction footprint and from hardware trade flows. Where methods disagree beyond their stated intervals, the disagreement is investigated and, if it cannot be resolved, published as an open discrepancy rather than averaged away.
Outputs- Independent estimates
- Reconciliation record
- Open discrepancies
- 04
Analysis
Assumptions are made explicit and their influence on the conclusion is quantified.
Every model we publish carries a sensitivity analysis identifying which assumptions the conclusion actually depends on. Where a conclusion is fragile to a single assumption, we say so in the body of the text rather than in an appendix. Analysis is conducted by a team that did not assemble the data, and a designated reader argues the opposing case before any finding is accepted.
Outputs- Sensitivity analysis
- Assumption register
- Adversarial review note
- 05
Visualisation
Charts are constructed to be checkable, not to be persuasive.
Every chart is generated directly from the underlying series, never redrawn by hand. Axes begin at zero unless there is a stated reason otherwise, in which case the reason appears in the caption. Intervals are shown where they exist. Any chart that could not be reproduced by a reader holding the published data is not published.
Outputs- Reproducible chart specification
- Underlying series
- 06
Publication
The finding, its uncertainty and the conditions that would overturn it are published together.
Publication includes the finding, the interval around it, the methodology in full, the source register, and an explicit statement of what evidence would change the assessment. Corrections are published with the same prominence as the original claim and are recorded permanently in the change log. We do not remove or silently revise published work.
Outputs- Publication
- Methodology appendix
- Change log
- Falsification conditions
Evidence hierarchy
Not all evidence is treated equally.
Every datum entering a Frontier Labs publication is classified into one of five tiers. The tier determines what weight it can carry: what a finding may rest on, what requires corroboration, and what is permitted only as supporting context. Tiers are recorded in the source register and are visible to anyone auditing a publication.
Direct primary record
Regulatory filings, statutory instruments, interconnection registers, customs declarations, planning submissions, court records and published microdata. Material created for a purpose other than persuasion, held by a party with an obligation to accuracy.
Sufficient to establish a finding on its own.
Disclosed operational data
Company disclosures, technical documentation, systems papers and telemetry released by an operator. Reliable in most respects but subject to selection: what is released is chosen by the releaser.
Sufficient with corroboration or with the selection risk stated.
Reconstructed measurement
Quantities we derive ourselves from physical or documentary evidence — capacity from electrical records, deployment from construction sequencing, compute from hardware and elapsed time.
Published as an interval, never as a point.
Structured expert judgement
Assessments elicited from practitioners under a documented protocol, with the question, the participants' domain and the dispersion of responses recorded.
Supporting only. Never the sole basis for a finding.
Secondary reporting
Journalism, analyst notes, conference remarks and industry commentary.
Used to locate primary material. Never cited as evidence in itself.
Estimation
How we handle what cannot be observed directly.
A large share of what matters in this field is not disclosed. Training compute, installed capacity, unit costs and deployment scale must generally be reconstructed rather than read. Reconstruction is legitimate research when its assumptions are visible and illegitimate when they are not.
Our rules are as follows.
- 01
Estimates are published as intervals
A reconstructed quantity is never published as a single number. Where the interval is wide enough that the quantity cannot be usefully classified, we mark it indeterminate rather than assigning a midpoint.
- 02
Bounds come before the estimate
We establish what is physically possible first — an upper bound from delivered electrical supply, from wafer capacity, from elapsed time — and narrow from there. Bounds are more defensible than models and are stated separately.
- 03
Independent methods must agree
Wherever a quantity can be estimated by two unrelated routes, it is. Where the routes disagree beyond their stated intervals, the disagreement is investigated and, if unresolved, published as an open discrepancy.
- 04
Definitions are stated before figures
Most apparent disagreement about quantities in this field is definitional. Every estimate is preceded by the definition it uses, including the choices that move the number by a factor of two or more.
- 05
Sensitivity is reported in the text
If a conclusion depends on one assumption, that dependency appears in the body of the publication, not in an appendix.
Charts
Constructed to be checked, not to persuade.
- Every chart is generated directly from the underlying series. No chart is drawn by hand or reconstructed from an image.
- Axes begin at zero unless there is a stated reason otherwise, and the reason appears in the caption rather than in a note.
- Values are labelled on the chart itself. A chart whose numbers require a mouse is not a chart a reader can check.
- Intervals are drawn where they exist. A reconstructed quantity shown without its interval is a misrepresentation regardless of intent.
- Colour carries no meaning anywhere in our work. Every chart is monochrome and remains legible in print and to readers with any form of colour vision deficiency.
- Any chart that could not be reproduced by a reader holding the published data is not published.
Transparency
What we commit to, permanently.
The methodology travels with the finding
Every publication carries its methodology in full. There is no separate tier of work published without it.
Sources are registered
Each publication has a source register listing every primary source, its tier, its provenance assessment and its revision history.
Prior views are recorded
The research group's prior view is recorded before the evidence is gathered, so that the distance travelled is visible in the published work.
Falsification conditions are stated
Every substantive assessment is published with an explicit account of what evidence would change it.
Corrections are permanent and prominent
Corrections are published with the same prominence as the original claim and recorded in the change log. Nothing is silently revised or removed.
Data is available on request
Underlying datasets are shared with any researcher who asks, subject only to withholding site identities where publication would create a security concern.
If you believe a Frontier Labs publication contains an error, please tell us. Factual corrections are the most important category of message we receive and are answered before anything else. Write to us through the contact page and mark the subject as a correction.