Permit
Check source, purpose, field, geography, retention, and training rights.
Evidence has a lifecycle.
The system is designed so an analyst can reconstruct how an item was acquired, changed, connected, assessed, reviewed, and—if allowed— published or used for model training.
What the system does—and refuses to pretend it knows.
The separately deployed private system is designed to identify claims, evidence relationships, topic and narrative patterns, changes in attention, coordinated-behavior indicators, and uncertainty. It does not reduce truth, source credibility, ideology, or people to a single score. A model output is a structured research artifact, never self-authenticating evidence.
Normalize
Preserve the original, canonical text, timestamps, language, and provenance.
Connect
Resolve duplicates, entities, claims, evidence, and evolving story clusters.
Analyze
Run typed tasks with citations, uncertainty, and provider-neutral models.
Review
Route consequential, disputed, uncertain, or policy-sensitive output to people.
Collection and rights
Approval comes first. Each connector is disabled until its legal basis, credentials, permitted fields, rate limits, retention, deletion behavior, training use, and publication use are recorded.
Source classes stay distinct. Licensed content, publisher feeds, official records, public social posts, and analyst-provided material keep separate policies.
Deletion propagates. Tombstones flow through raw storage, indexes, derived artifacts, caches, datasets, and future model-release eligibility.
Public is not unrestricted. Technical access does not itself authorize collection, redistribution, or training.
Hosted deployment scope
This site is a curated public research snapshot. Its world-event dossiers link to current representative sources and expose original summaries, evidence notes, short permitted extracts, dates, source classes, and scoring metadata.
The analyst data plane remains private. Source ingestion, model execution, evidence review, and operational records stay in the separately controlled pilot environment until its API, database, identity, and governance controls are deliberately promoted for hosted use.
The snapshot is not continuous monitoring. Each dossier carries an update and access time. Events can change after publication, and readers should follow the source links for later revisions.
Representative is not exhaustive. The source ledger is selected to show the strongest direct evidence and meaningful disagreement, not every article or public post about an event.
Signal and claim scoring
A signal score is an evidence-trail indicator. It is not a probability that an event is true, a rating of a person or publisher, or a prediction. Scores can fall when newer evidence introduces disagreement or exposes a verification gap.
Evidence quality contributes 35%. Direct official records, original data, field reporting, and clearly attributed documentation score above circular commentary or anonymous restatements.
Independent corroboration contributes 30%. The method rewards agreement across institutions that do not merely cite one another and preserves explicit contradictions.
Recency contributes 20%. A source must be current for the material claim. Older baselines remain useful only when labeled as historical and not presented as a current measurement.
Source diversity contributes 15%. Different source classes, jurisdictions, and geographic perspectives reduce the risk that one institution defines the entire evidence picture.
Claim scores are separate. Each consequential claim receives its own evidence-fit score, status, rationale, linked source IDs, and caveat. The dossier score cannot override a weak or disputed individual claim.
Analysis and uncertainty
Tasks are typed. Claim extraction, stance, evidence relation, clustering, summarization, and narrative-frame detection each have distinct schemas and release gates.
Citations are mandatory. Consequential assertions must point to permitted evidence spans. Unsupported generations fail validation.
Confidence is calibrated. Probabilities are assessed by task, source class, language tier, and time period—not treated as universal certainty.
Unknown is a result. The system can abstain, report disagreement, or request review instead of forcing a conclusion.
Models and evaluation
Fine-tuning is earned. Retrieval, prompting, rules, and calibration establish a baseline first. Training begins only when evaluation shows a durable gap that labels can address.
Splits prevent leakage. Evaluation separates time periods, source families, story clusters, and near-duplicates.
Languages have release tiers. A model may assist analysts in a language before it is eligible for automation or public output in that language.
Models remain replaceable. Open-weight, self-hosted, and managed providers use the same task contracts, evidence rules, and release evaluation.
Publication and human review
Default review is risk-based. Consequential, disputed, uncertain, named-person, or policy-sensitive output enters a review queue.
Automation is scoped. Approval is specific to a task, source class, language, audience, model version, and policy version; it is never a blanket switch.
Public output is source-conscious. The public surface exposes dossiers, metadata, scoring, and a research API while linking out for full copyrighted material and honoring quotation and redistribution restrictions.
Corrections remain visible. Material changes, reversals, appeals, and provenance gaps are recorded rather than silently overwritten.