Research architecture behind the portfolio
Research continues around a loop. Questions wait in a queue; one is selected and scoped; evidence, analysis and challenge repeat; a version is published; its repository, deposit and project records are brought into agreement; monitoring or reuse can lead to a revision or a new question. A study can span several stages. A source catalogue, calculation or release record illustrates one step within it.
1. Question queue
Keep candidate questions with the date and reason they were raised. An entry is a lead, not a commitment. No public queue-entry example is linked here.
2. Selection and scope
Choose a question the available evidence can address; define its unit, population, period and comparison, and select a method suited to the question; check existing work and practical purpose. A contribution, useful replication or synthesis can justify proceeding. Otherwise narrow, reframe or archive.
Power demand
Distinguish announced AI power demand from the capacity that can be delivered.
SW
Ask how a monitoring service should expose incomplete work and preserve recovery opportunities.
Forecast scorecard
Compare electricity-demand forecasts within compatible units, scopes and horizons.
Article 50
Keep the historical public-evidence snapshot distinct from a current compliance assessment.
FineWeb-Edu
A question at the mechanism level: how score rounding and a threshold affect which documents enter a training corpus.
The published output illustrates the question; this link is not a completed novelty-search record.
3. Research and challenge
Collect records, preserve provenance and check reporting periods. Retain missing and conflicting evidence. Analyse with inspectable assumptions and quantities, test sensitivity where appropriate, seek counter-evidence, test alternative explanations, examine source support and revise claims that exceed the evidence. Return to scope when the question needs changing.
AI Infrastructure
A source catalogue connects inspected locations and source limits to the paper's commitments, delivery stages and access obligations.
SW
Source capture retains document text and content hashes, while failed and deferred assessments remain visible.
Forecast scorecard
The reproducer recomputes the scorecard table and its transparency tiers from one CSV with the Python standard library, and regenerates the dispersion chart.
Open the release and its reproducer
Reproduction recomputes the scorecard from the recorded inputs; it does not re-verify each forecast at its source.
FineWeb-Edu
The explorer exposes the score threshold and rounding rule so their effects can be inspected.
IPO Ledger
A prediction specified in advance is graded from recorded prices. The first lock-up checkpoint is reported as failed, and the release states that the earlier version specified the prediction defectively before restating it as dated checkpoints.
Open the release and its grading script
Grading re-runs from the prices in the bundle; it does not re-fetch or independently verify the exchange data.
Model Dependency
The corrected release distinguishes model-run provenance from answer checking and states what can be reconstructed from stored outputs.
4. Publish a version
After the author’s release decision, publish a version with its method, evidence limits and useful reproduction materials. Preserve earlier editions and make corrections inspectable.
AI Infrastructure
The versioned bundle includes the paper, structured inputs, source lookup, reproduction commands and a reuse example.
CEO Pay scorecard
The versioned bundle regenerates its issuer scorecard from the included pulled data and checks the security set, using the Python standard library.
A matching rebuild confirms the assembly step, not the underlying filings.
5. Align records
Check that the public repository, deposit and project-status record identify the same released version. Project-status records are distinct from registers of methodological decisions. No public alignment receipt is linked here; the component below illustrates decision recording only.
Divergence Registry
Decision registers provide a separate method for recording choices and comparing research runs.
Inspect the decision-record component
This is a decision-recording component. It is not evidence that a publication, repository and project record were aligned, or that every study ran the full Lifecycle engine.
6. Monitor and reuse
Revisit on stated dates or when new evidence appears; record any actual reuse and its limits. Reuse does not revalidate findings. The AI Infrastructure example shows available reuse materials, not proof of a later run.
AI Infrastructure
The release includes a reuse example alongside its paper, structured inputs, source lookup and reproduction commands.
Materials provided for reuse do not establish that a later reuse or monitoring run occurred, or revalidate the findings.
IPO Ledger
After grading the first lock-up checkpoint, the release lists the remaining dated checkpoints through 12 June 2027, with the observation that would confirm the call at a checkpoint and the condition that would falsify it across the full ladder.
Open the release and its checkpoint schedule
The schedule sets when each checkpoint falls; it does not establish that a later checkpoint has been graded.
7. Continue or revise
Continue within the active project, publish a corrected or extended version when warranted, or open a new queue entry if the question itself changes.
Model Dependency
The corrected release distinguishes model-run provenance from answer checking and states what can be reconstructed from stored outputs.
The corrected release illustrates revision. It does not document every earlier stage in the study.
8. Archive branches
A question or study can pause, stop, be reframed or merge into other work. Retain the recorded decision and any useful materials. An archived item has not necessarily failed novelty. Missing history stays unrecorded.
Embodied Energy
The retained literature-position review recommends against a standalone paper because the proposed contribution was largely covered by existing literature. It recommends retaining the work as a method example.
Read the selected decision record
This records a recommendation; the map has not repeated its literature search.
Critical Minerals
The project record labels the standalone publication folded after its novelty check, while retaining a method demonstration.
Read the selected decision record
This records the decision, not a fresh validation of the underlying mineral-market claims.
DRACO-lite
The retained README calls this a parked evaluation skeleton rather than an active project.
Read the selected decision record
That record does not say it failed a novelty check.
FineWeb-Edu
The handover distinguishes the superseded Dolma auditability design from the later FineWeb-Edu work.
Read the selected decision record
A superseded research arm is not a failed published FineWeb-Edu study.