Built to question the premise, not just write the paper.

The author supplies the question—not a guaranteed verdict.

An author can introduce bias before writing a single sentence: through a loaded question, selective sources, an assumed cause or a preferred explanation. A fluent answer can make those assumptions harder to see. R.A.T. — Retrieval-Augmented Thought is designed to put analytical checks between that initial framing and the published narrative.

That means more than asking an AI to “be objective.” Subject-relevant retrieval, multiple assessments, evidence-led synthesis, tool-based observations and a separate editorial fidelity review create opportunities to challenge a claim at different stages. They do not guarantee that every bias will be detected.

• Question the premise. An allegation is something to examine, not a fact established by its presence in the question.
• Compare explanations. Different assessments can expose hidden assumptions, counterexamples and reasons a conclusion may not generalize.
• Separate support from persuasion. A source's statement, a model's inference and an executed measurement are different kinds of support.
• Retain the qualification. A finding limited to one jurisdiction, edition, sample or set of assumptions must not quietly become a universal claim.
• Keep uncertainty meaningful. Missing evidence does not prove falsity; a lack of detected anomalies does not prove authenticity.

For example: “Does this footage prove the event happened?” contains several questions. What is visible? Has the recording been altered? Does its context support the claimed event? A genuine recording can carry a false caption. Separating those questions helps resist a preferred conclusion in either direction.

Omnibrain: bring the right knowledge to the question

Omnibrain is the coordinating research architecture behind R.A.T. It connects an inquiry to relevant subject knowledge and analytical capabilities instead of relying solely on what a general-purpose model recalls. Supported research paths can select material according to the question's subject rather than treating the website on which it was asked as the final authority.

Domain packs organize specialized knowledge and its context: codes and standards, law, technical references, scholarship and other subject collections. Different material calls for different treatment. A regulatory requirement needs an applicable edition and jurisdiction; a historical interpretation needs attribution and context; a database result is not the same as a passage that sounds quantitative.

Retrieval is a relevance process, not a truth stamp. A passage can be outdated, incomplete or biased. When enabled, web search adds current leads. Optional cross-disciplinary lenses widen an investigation where another field may help, but an analogy remains a hypothesis—not evidence that two situations are equivalent.

The swarm: more than one perspective in the room

R.A.T.'s research workflow can bring multiple model/provider assessments into the same investigation. Different analytical roles can examine factual support, assumptions, implications and objections before a synthesis is prepared. The purpose is to make competing assessments available for comparison, not simply to repeat one answer.

Agreement can be informative, but it is not proof. Models may share training data, sources and similar blind spots. A majority can repeat the same mistake, and a critic can be wrong. Synthesis seeks a usable answer while the publication editor and audit are instructed to preserve substantive qualifications and avoid treating criticism itself as a finding. Original analyst outputs remain available for readers to compare.

The Jurist Protocol: claims must answer to evidence

The Jurist Protocol names R.A.T.'s evidence-first approach to adjudicating claims. In supported reasoning and synthesis paths, its discipline is to distinguish what the material says, what is inferred from it, and what remains unsupported. It is a research method—not a court, a legal opinion or a certification.

A confident author, model or source faces the same question: what supports this conclusion, and how far does that support extend? A useful result may be a narrower claim or an unresolved question rather than a dramatic verdict. The protocol's role varies by workflow; not every publication invokes a separate Jurist process.

The Eye: inspect the material, not only its description

For audio and video investigations, the Eye is the visual-assessment capability. The publishing workflow combines sampled-frame cloud analysis with available transcription and measured media properties. This gives the analysis material to inspect alongside the uploader's account of what a recording supposedly shows.

Recording authenticity and the truth of a statement in a recording are kept separate. Sampling can miss important events; a visual model can misinterpret what it sees. This is not a calibrated deepfake detector and does not perform voiceprint matching, measured lip-sync or comprehensive continuous-motion analysis.

Python checks: calculate instead of merely describing a calculation

Where a task calls for it, Python can execute measurements or computations rather than leaving arithmetic to prose generation. In the publishing media workflow, measurements and verification scripts run in an isolated environment. Document extraction and OCR also run separately from the public website.

Execution adds a checkable computational step, not automatic factual authority. Correct code can use a bad assumption or the wrong input. An executed script does not validate statements in its comments, prove a source authentic or remove selection bias. The method should explain what was actually measured and what was not.

Jev: a decision is not a truth score

Jev, from TypeSafe SystemOne, supports structured decisions in the R.A.T. architecture, including subject selection on supported research paths and automatic editorial placement in the publisher. Selecting a useful area of knowledge is not the same as deciding where a completed paper belongs.

For automatic publication placement, Jev considers the completed article and its editorial audit against publication criteria. It is instructed not to treat an author's allegation or a critic's accusation as proof. It can abstain, leaving the paper private rather than forcing a label. Its decision confidence is not a probability that the paper is true.

A separate check between the report and the reader

The editor turns the supplied research or extracted document into a readable paper. A separate model-assisted audit compares that paper with its input: were claims introduced, qualifications lost, quotations altered or uncertainties made to sound settled? Structural checks also require traceable input references.

This is a second review stage, not necessarily an independent model or an independent human reviewer. It checks editorial fidelity rather than independently verifying world facts. Editorial blockers or an unsuccessful automatic destination decision keep the paper private. An authorized operator may select a destination or request a private preview; public visitors do not control publication.

One engine, different investigation paths

The capability list is not a claim that every tool runs on every paper:

• Question-led research: the publishing desk requests R.A.T. ANALYSE, with retrieval and specialist assessments. Web search is enabled by default in this workflow; cross-disciplinary exploration is optional.
• Uploaded documents: short-document text, including OCR where needed, goes to the editor and audit. The short-document path does not automatically run the research swarm or independent web investigation. Long documents use overlapping evidence passes, source-quotation checks, comparison retrieval and scoped Jurist synthesis before editing. Coverage records distinguish chunks processed from evidence selected; neither guarantees complete semantic recall. Search leads are not verified primary documents. OCR can misread words and numbers; layout, images and table relationships may not be fully captured.
• Audio/video investigations: transcription, sampled visual inspection and isolated Python measurements contribute to the report, which must retain their coverage limits.
• Imported reports: the editor works from an existing report. Importing it does not prove the original investigation was complete or rerun all its methods.

Generated covers are labeled conceptual illustrations, not evidence. Reading maps explain article organization. Hypothetical charts must state their assumptions rather than masquerade as observed data. Forensic media papers do not receive synthetic cover illustrations.

Bias is challenged, not magically eliminated.

We aim to make bias easier to challenge and unsupported certainty harder to publish. We do not claim bias-free models, exhaustive retrieval, guaranteed neutrality or a demonstrated numerical reduction in author bias. The author still controls the initial question and supplied material; sources, selection, models and editorial choices can all introduce bias.

Engineering tests are not a substitute for independent validation of an individual paper. This page explains the purpose and boundaries of the system without publishing proprietary prompts, decision rules or operational configurations.

How to read a R.A.T. paper critically

1. Check the question. Does it assume the claim the paper needs to establish?
2. Read the scope and limitations. Which materials and methods were actually used?
3. Open the original report. R1 and R2 identify input sections, not necessarily primary sources or independent corroboration.
4. Distinguish source statements, interpretation, calculation and speculation. Look for assumptions that could change the result.
5. For consequential legal, medical, financial or safety decisions, verify applicable primary sources and obtain qualified advice.

The product is not a machine for making every author right. It is an engineered way to make a claim face questions, competing interpretations and a record that readers can inspect.