Methodology

Before a thesis exists

Each investigation begins with one registered commission, a source packet, and competing hypotheses. Quantitative work fixes its population, endpoint, denominator, threshold, and time horizon before selecting a preferred explanation. The commission also records a loss story: what the paper becomes if the preferred conclusion fails.

A thesis must be falsifiable in operational terms. “More important,” “better,” and “safer” are not measurements until the article defines what is counted, compared, and excluded.

Evidence labels

Scoped observation

A dated observation supported within the stated scope of a primary source. A filing proves what was filed. A company report proves what the company reported. Neither automatically proves the underlying effect.

Interested-party report

A claim from a party with an institutional or commercial interest in the result. These reports can be useful, but they remain attributed and do not receive independent status by repetition.

Estimate

A derived number. The article must expose its formula, inputs, units, range, assumptions, alternate denominator, and calculation artifact. A point estimate is not used when uncertainty is load-bearing.

Forecast

A future claim with a base rate, range, resolution date, and resolution rule. Forecasts remain in a public ledger until they resolve.

Inference

The proposed explanation of observations. Inferences name plausible alternatives and the evidence that would discriminate among them.

Proxy

A measurable substitute for the actual object of interest. The article states what the proxy misses and does not silently rename it as the target.

Prior

A value judgment or starting belief. Priors can guide recommendations, but they cannot upgrade confidence in an empirical claim.

Recommendation

A normative action. Recommendations identify beneficiaries, losers, coercive elements, externalities, opportunity costs, reversibility, and at least one non-house regime consistent with the same evidence.

Sources and calculations

Load-bearing claims use the closest available primary source: papers and supplements, standards, dockets, filings, raw public datasets, or official incident records. Press releases establish what an issuer said, not whether the claimed effect occurred. Sources record publication and access dates, version, jurisdiction, scope, and reuse limits.

Public access is not permission to republish a chart or database. We redraw only from data whose reuse permits it and publish calculation inputs when rights allow. Derived results are reproducible from checked-in scripts or carry a visible manual attestation when public redistribution is not lawful.

When primary sources conflict, the article compares definitions and methods before selecting a figure. Missing, confidential, unreported, zero, and not applicable are distinct states.

Adversarial review

Every publishable investigation is reviewed with Grok 4.5 under an anti-prior brief. The reviewer can block publication for metric switching, a headline stronger than the body, missing counterevidence, an unlabeled prior, an institutional-object mismatch, or a number that cannot be reproduced.

Material objections appear at the point where they challenge the argument. Each records the evidence considered, whether the thesis stood, narrowed, reversed, or was killed, what changed in the paper, and what remains unresolved. A detached review transcript is not a substitute for updating the argument.

The same evidentiary rule must be applied to institutions favored by the publication. Speed, abundance, and permissionless operation are values, not proof.

AI assistance

AI systems assist with source discovery, claim indexing, calculation checks, drafting, editing, and adversarial review. They do not make a source true. Load-bearing links are opened, dates and scopes are checked, derived numbers are rerun, and citations are compared with the final text before release.

Style tools are used first in detect-only mode. Any later edit is narrow. A prose pass may not silently change a number, quotation, citation, confidence level, or modal verb. The final artifact and its source history, not a model's self-assessment, determine release.

Corrections

Material corrections are never silent. The affected article receives a dated notice tied to claim identifiers, and the same entry appears in the corrections ledger. The notice states what was wrong, why it was wrong, what changed, and which conclusions moved. Typography-only fixes that do not change meaning may be made without a ledger entry.

Conflicts and accountability

Each article discloses material funding, holdings, consulting, access arrangements, embargoes, and issuer review. Serious allegations require source-level verification and a fair opportunity to respond. Medical, investment, export-control, security, privacy, and dual-use claims receive additional review appropriate to their risk.

Corrections may be submitted through the Quirky Ideas contact page. Include the article URL, claim identifier, source, and the smallest precise statement of the problem.

Current limits

The method cannot make closed data public, produce a clean counterfactual where none exists, or prove that the selected objection was the strongest possible one. Public-only evidence often forces narrower conclusions. Formula checks establish reproducibility, not validity. Confidence remains provisional until the evidence survives use and criticism.