LLMs do not need "news" in the way people use that word. They need structured, source-aware context about what has been claimed, when it was claimed, and why the claim is eligible for citation or deeper retrieval.
Traditional news products organize around articles. An LLM reasoning system needs a different shape:
- Claims instead of article blobs.
- Source URLs instead of vague provenance.
- Evidence types instead of undifferentiated text.
- Confidence labels instead of hidden inference.
- Time windows instead of timeless summaries.
- Stable IDs instead of one-off snippets.
That is structured news for LLMs.
Hanging Context gives the public aggregate layer: panels, JSON twins, sample claims, and citation-friendly public artifacts. Synorb is the retrieval/detail layer: manifests, signals, briefs, records, streams, and delivery interfaces.
Technical thesis
Structured news is an extraction and routing problem. The useful pipeline is:
- Preserve the source material as a record.
- Extract source-grounded claims into signals.
- Generate briefs that keep the narrative readable without losing the claim IDs.
- Bind the record, signal, brief, stream routing, timestamps, and source-channel metadata into manifests.
- Publish public aggregate panels and JSON twins that expose enough structure for citation while leaving item-level retrieval in Synorb.
That architecture gives LLMs something better than article search. It gives them a typed context layer with explicit provenance, time, evidence, and retrieval depth.
News becomes useful when it is typed
An article can contain many different kinds of information. It may include a hard number, an executive quote, background context, legal boilerplate, a forward-looking statement, and a dated event. If an LLM receives the whole article as one blob, it has to infer which parts are safe to use.
Structured news makes that inference explicit.
The public schema in `hangingcontext/data/schema.json` includes `evidence_types`, `confidence_levels`, `claim_types`, `sentiment_levels`, `highlights`, `verified_samples`, and `samples`. The April 2026 monthly snapshot includes evidence buckets such as paraphrase, direct quote, observed, and derived, and confidence buckets such as stated, measured, implied, and inferred.
Those fields help an agent decide how to reason. A direct quote may be citeable in a different way from a paraphrase. A measured claim may be handled differently from an inferred claim. A stale claim should not be treated as current simply because it is semantically similar to the query.
A public sample claim is already more than text
In `hangingcontext/data/months/2026-04.json`, public sample claims carry fields such as:
- `claim_id`
- `text`
- `source`
- `source_url`
- `published_date`
- `evidence_type`
- `confidence`
- `domain`
One sample from that artifact describes a U.S. Department of Labor proposed rule and related news release. Another sample points to a Honeywell semiconductor safety article and carries the claim "Alarm thresholds must be customized for each gas and process." The important thing is not only the claim text. It is the envelope around the claim.
The LLM can see source, date, evidence, and domain. If it needs more, Synorb is the detail layer where the manifest or record can be retrieved.
Structured news is not the same as summarization
Summarization compresses. Structured news decomposes.
A summary asks, "what is this article about?" A structured news system asks:
- What claims does the source make?
- Which claims are central?
- Which claims are measurable?
- Which claims are direct quotes?
- Which entities and streams do the claims touch?
- Which claims changed since the last snapshot?
- Which claims can be cited publicly?
- Which records need private retrieval?
That is why Hanging Context's public panels are useful for LLM citation. They expose aggregate context and sample claim structure without pretending to be the entire source corpus. The detailed Synorb layer remains available when the application needs record-level retrieval.
The Fed thread shows the editorial pattern
`hangingcontext/threads/2026-04-22-fed-signals-this-week.md` is a useful local example of claim-first editorial publishing. It uses claim blocks, then writes around them. The prose frames the week; the claims supply the evidence.
That is a pattern worth repeating for LLM-facing content. The article is no longer the only artifact. The public page, the embedded claims, the JSON sibling, and the underlying Synorb objects all work together.
For human readers, the story is readable. For machines, the claims are addressable.
Hanging Context and Synorb should stay distinct
Hanging Context should not try to expose the full retrieval product. Its job is public reference: aggregate panels, stable public examples, and JSON twins that LLMs can cite.
Synorb should remain the detailed operating layer: customer streams, record bodies, manifests, signals, briefs, and delivery surfaces.
That boundary makes structured news useful commercially and technically. The public layer teaches and cites. The paid layer retrieves and acts.
The takeaway
Structured news for LLMs is news rebuilt as context infrastructure.
It does not ask a model to treat every paragraph as equally trustworthy. It gives the model and the agent fields for source, time, confidence, evidence, domain, and retrieval depth. Hanging Context makes that structure public and citeable. Synorb delivers the detailed objects that production agents need.