How Research Works
Our automated analysis follows a source-bound review process. Passing it is evidence of support within the retained source; it is not independent replication or a guarantee that a paper is correct.
From paper to publication
Discover
Identify a paper and its exact source version.
Acquire
Check source-use permission and retain the source.
Normalize
Extract source structure and track missing or damaged content.
Classify
Choose the research domain, kind and appropriate lens.
Extract
Associate claims, conditions and limitations with admissible evidence.
Synthesize
Generate a reader draft from the admitted claims.
Validate
Check identity, references, coverage and quantitative source bindings.
Evaluate
Assess whether the admitted evidence supports each material statement.
Publication decision: publish an accepted analysis, quarantine a failed draft, or abstain when a supported analysis cannot be produced. Source, budget and infrastructure holds remain separate outcomes.
What code decides
Deterministic checks own exact source and version identities, evidence locations, relationships, budgets, publication state and canonical URLs. A valid citation reference establishes where evidence is located. It does not prove that the evidence supports the statement.
What models assess
Models classify, extract and explain research, then assess semantic support within each claim’s admitted evidence. A model judgment cannot override a failed identity, provenance or other deterministic publication check.
Evidence and limitations stay attached
Reader citations are a compact view of the evidence. Evaluation uses the complete admitted evidence associated with the relevant claim or support unit; material elsewhere in the paper is not automatically authority for that statement.
Numbers, conditions, uncertainty and limitations must remain grounded. Normalization defects are assessed where they affect claims. A damaged symbol unrelated to a claim need not block the whole paper; an unresolved material dependency cannot be accepted as supported.
Why an analysis may not be published
Quarantine records a generated draft that fails the publication contract. Abstention records an intentional decision not to produce a supported analysis. Neither is evidence that the source paper itself is poor research.
Capacity, source access, cost limits and infrastructure failures are reported separately from scientific failures. We do not regenerate repeatedly just to improve a publication rate.
Revisions and withdrawals
Each source version has its own immutable processing record. A newer version may replace a published analysis only after independent acceptance. The stable Research URL remains the same, and the previously accepted version remains public while the update is evaluated.
A detected withdrawal enters review with its source event and provenance. It does not silently delete or rewrite a published article.
Read the results with their boundaries
Automated analysis and automated judges can miss subtle qualifications or misread evidence. Human scientific review is recorded separately; a model verdict does not stand in for it. Audio narrates only the accepted public reader and does not change publication status.
Research Reliability separates methodology cohorts, shows available denominators, and labels missing measurements as unavailable. Model and evaluation versions are reported with the cohort that actually used them.
Method documentation: research-method:v1. Operational activation and measured outcomes are reported on Reliability; this page does not itself assert that a scheduled pipeline is live.