News features are dangerous when publication time, source identity, or later edits are unclear. Treat each article as a versioned document observed at a known cutoff. A headline alone is not evidence, and a search-engine snippet is not the source.
Create a source policy prioritizing official resolution sources, government or institutional releases, and named primary documents. Secondary reporting can point to evidence but should not replace the underlying document when it is available. Record URL, title, publisher, observed time, stated publication/update time, content hash, language, and retrieval result.
Prevent lookahead with an availability timestamp: the earliest time your system actually captured the document. A page claiming an earlier publication time but first observed after the paper decision is not available to that decision. For edits, store a new hash and version; never overwrite.
Features should be modest: source count by approved category, contradiction flag, hours since first verified document, rule-term match, and missing-primary-source flag. Avoid sentiment as a magical prediction feature. If used experimentally, validate it on labelled samples in the relevant languages and report error groups.
An LLM may summarize the locked packet, but must cite source IDs and flag uncertainty. The deterministic pipeline checks IDs and dates. Prompt-injection text inside a page is treated as quoted data and cannot alter instructions.
Define document deduplication carefully. Syndicated copies do not become independent confirmation merely because they appear on several domains. Cluster exact and near-duplicate text, preserve each provenance record, and count the originating primary document once in evidence-strength summaries. When the original cannot be located, label the chain secondary and lower the case to human review rather than filling the gap with repetition.
Audit extraction by sampling documents with and without relevant claims. Record missed dates, wrong entities, broken quotations, and false duplicates by rule version. A news feature is promoted only after its time cutoff, provenance, and transformation can be replayed from the sealed packet. Otherwise the case stays descriptive or abstains.
🇵🇰 Pakistan Angle
Official Pakistani releases may be PDFs, bilingual pages, or social posts. Preserve the original document and verify the issuing domain/account. Screenshots without a stable source are weak evidence. Translation must be labelled, and sensitive personal information should not enter the dataset.
Hands-On Exercise
Build a five-document event packet containing one late document and one contradiction. Generate availability-safe features at two cutoffs. Prove the late document is excluded from the earlier decision. Produce a provenance table and a human-readable contradiction note.
Completion Rubric
- Every feature traces to a hashed source observed before cutoff.
- Updates create versions instead of overwrites.
- Primary sources and contradictions are visible.
- Model summaries cite only allowed source IDs.
- Sentiment or translation limitations are measured, not assumed.
Sources
Key takeaway: News becomes a defensible feature only when source, version, and availability time are provable.