# NRC Daily Drop data guide

The [NRC Daily Drop](https://coriumrecords.com/records) publishes a locally ranked
shortlist of public NRC ADAMS metadata for research and project discovery.
It is not a complete NRC archive, a document-content summary, or an NRC alert service.

## Choose the right file

- [Latest shortlist](https://coriumrecords.com/data/adams/latest.json): start here for the most recently published intake. The current pipeline retains up to five ranked topics.
- [Manifest](https://coriumrecords.com/data/adams/manifest.json): publication metadata, dates, byte lengths and SHA-256 fingerprints for the latest feed and history index. Fingerprints verify those published files, not the NRC originals.
- [History index](https://coriumrecords.com/data/adams/history_index.json): retained shortlists reassessed using the current reading classifier, with coverage in `date_range`. It is historical context, not a second current feed or a complete document archive.
- [OpenAPI schemas](https://coriumrecords.com/openapi.json): field types and meanings for these public, read-only files.

## Read the dates before describing freshness

`requested_date` is the posting day initially requested from ADAMS. `used_date`
is the posting day actually selected after any fallback. The upstream query uses
`DateAddedTimestamp`: this is when a record was added to ADAMS, not the document's
authorship date or the date an event occurred.

`generated_at` is when the collector created the exported shortlist. The manifest
and history index use `generated_at_utc` for their own generation times.
`review_metadata.assessed_at_utc` describes a later assessment of retained titles
where present; it does not update the original collection time or posting day.
A recent export or assessment does not make an older posting day current.

If a requested day has no retrieved records, the tool can look back up to
`source.fallback_days` days. Consult both posting dates rather than calling every
successful fetch today's news. The published snapshot alone does not prove the
refresh job is currently healthy.

## Use the current reading assessment

Prefer `topic.review.priority` and `topic.review.reasons` when deciding what to
read. Classifier version `2.0.0` uses document role and context in the title, then
bounded topical signals. It does not read document bodies. Related synonyms do
not each add independent priority points. `document_type` and `themes` are
separate from priority; an editorial subject is not itself a finding of harm.

| Priority | Meaning |
| --- | --- |
| `HIGH` | Read first: a title indicates a substantive reported condition, enforcement action, decision, or consequential change. |
| `MEDIUM` | Worth reviewing: substantive applications, reviews, updates or supporting developments. |
| `BACKGROUND` | Lower default reading priority: routine process, reference material, or insufficient substantive context in the title. |

These are local reading priorities, not NRC safety ratings, accident
classifications, urgency guarantees or verified document conclusions. There is
no automatically assigned Critical priority. A title-based explanation tells you
why the record was selected; open the original before asserting what it establishes.
`review.rank_score` only orders records within this classifier version and is not
a hazard measure or a value comparable with the old score. A reassessed legacy
shortlist keeps its original `rank` fields; use `review.rank_score` for current
ordering. Fresh `review_priority` selections use the new classifier order.

`score`, `severity`, `hits` and `angle` preserve the earlier keyword heuristic for
historical comparison. They no longer determine current rank. In particular, an
old `severity: "CRITICAL"` is not a current review priority. `angle` is a legacy
editorial prompt, not a quotation or finding from NRC. Original dated snapshots
may lack `review`; absence means they have no assessment stored in that file.
Do not substitute their legacy severity for the current reading assessment.

`review_metadata` records the assessment version, basis, time and scope.
`selection_method: "legacy_keyword_shortlist"` means the current assessment covers
an already selected old shortlist. `selection_method: "review_priority"` means the
new classifier selected the retained shortlist. Neither establishes complete NRC
coverage. Assessment provenance is separate from collection provenance; unknown
historical collector revisions or other missing provenance remain unknown.

## Understand selection limits

The tool retrieves bounded pages of public metadata, groups related records,
classifies titles and selects up to five topics. The publisher retains that
shortlist, not all retrieved documents or their PDFs. The value
`summary.docs_total_for_used_date` is the count of metadata records retrieved for
the chosen day before selection. It is not a certified count of all NRC documents
for that day, nor evidence that their contents were read.

`source.page_size` and `source.max_pages` describe retrieval limits. Where present,
`coverage.completeness` reports that pagination completeness is unverified.
Reassessing saved titles cannot recover records excluded by old filters or
selection. A change in classification can change which saved records are
highlighted without revealing new underlying documents.

## Use history as history

Preserved `YYYY-MM-DD.json` files retain their original fields. Later
exports for the same posting day are kept under
`revisions/YYYY-MM-DD/<timestamp>-<hash>.json`; existing snapshots are not rewritten
to insert new assessments. This preservation scheme begins with the revised
publisher. Earlier exports that were overwritten before this scheme are not
reconstructed by the history index.

The index uses the most recently generated retained snapshot for each posting day.
`snapshot_count` counts distinct posting days; `preserved_snapshot_count` counts
original and revision files. `topic_count` and `review_priority_counts` count the
selected topic appearances across those days, not unique incidents, all refresh
runs or all retrieved documents. `severity_counts` retain legacy-band comparisons.

`high_signal_topics` contains at most 60 historical appearances currently assessed
`HIGH`; `high_signal_count` counts all qualifying appearances before that limit.
`recent_pulls` contains at most 14 posting-day summaries. Its `top_title` and
`top_review` describe the first retained topic under the current assessment;
`top_score` and `top_severity` are that same topic's legacy comparison fields.
`retention` reports the actual returned sizes. Old records with High reading
priority must not be presented as current alerts.

## Follow and cite the original record

Use each topic's supplied `urls` and accession identifiers. A URL can point to a
PDF or an HTML package page; preserve that distinction. Open the original record
before making claims about its content, authorship date, event date or conclusions.
A title classification is not independent confirmation of those claims. If
retrieval is blocked, report it as unverified rather than calling the source
missing or substituting an unrelated record. The [NRC ADAMS search](https://adams-search.nrc.gov/)
can be searched by accession for manual review.

These endpoints require no login or Trip Ticket. The music catalog and its build
reports are separate from this feed and do not establish ADAMS freshness or source health.