News desk, source registry, and developing-cycle model

Every meaningful development is one atomic Markdown story with a stable news_item_id, canonical page, source and update timestamps, discovery time, authors, classifications, evidence relations, review history, and correction log. A story may join multiple developing cycles, but shared vendor identity alone is not a relationship.

Reviewed source registry

Each version-controlled source entry declares its stable ID, display name, canonical site and feed, source type, adapter, allowed and excluded paths, reuse and attribution notes, terms and robots review, language and region, expected cadence, rate limit, timeout, retry policy, owner, review date, and active/paused/retired reason. Material registry changes require human review.

Collection stores only fields needed for discovery and review: canonical URL, redirects, title, author, timestamps with zone, permissible excerpt, retrieval metadata, response validators, and hashes. Full pages and source media are not committed unless license and research need explicitly permit it; source imagery is never hotlinked as decoration.

Candidate record and deduplication

A normalized candidate includes permanent ID; optional cycle ID and title; original headline and authors; language; source, publication, update, and discovery times; short permissible excerpt; hash and story key; affected entities and observability classes; proposed story type and relevance; parser/retrieval versions; links to research; and typed relations: extends, confirms, contradicts, corrects, supersedes, implements, deprecates, or releases.

The pipeline compares canonical URL and redirects, normalized headline, content fingerprint, DOI, release identifier, and entity-event similarity. Reports of the same event become one cluster with source composition, not artificial duplicate stories.

Coverage composition

Each card and story distinguishes vendor-originated from independent reporting, primary from secondary evidence, declared commercial relationships, source ownership context, agreement and contradiction, and unknowns. These are provenance dimensions—not political-bias labels. Compact cards expose headline, source mix, published and updated time, affected classes, and evidence state.

Topic and entity feeds, filters, RSS/Atom, related profiles and comparisons, and stable archive routes make stories discoverable. Developing-cycle pages add a dated event timeline and a graph whose edges carry relation type and evidence. Accessible tables duplicate every graph and show denominator, time range, aggregation, update time, and exclusions.

Human decision and corrections

The accountable editor confirms scope, classification, title fidelity, claim support, vendor-claim labeling, version and freshness, reuse rights, byline and automation disclosure, conflicts, duplication, accessibility in both themes and supported viewports, and like-for-like comparison framing. Green checks or elapsed time never equal approval.

Published pages show original publication and update times, sources, evidence table, author profile, automation disclosure, correction history, and material revisions. Corrections retain the stable URL, explain what changed and why, timestamp the change, and link the reviewed pull request. A retraction preserves a clear record unless legal or safety needs require restricted handling.