Research Watch Feed

Run a compact, schedulable watch over scholarly DOIs, model cards, clinical studies, and software repositories, producing evidence-linked JSON, Atom, or RSS feed artifacts without another Flintglade service.

What this does not do

• No broad web monitoring.
• No hosted dashboard dependency.
• No nested calls to paid Flintglade Actors.
• No generic RSS reader.
• No LLM summaries.
• No source type outside the four launch profiles.

Source classes and authorization

scholarlyDoi checks both Crossref and NCBI. modelCard resolves an immutable Hugging Face revision and reads its README.md. clinicalStudy reads ClinicalTrials.gov API v2. softwareRepository reads bounded GitHub repository, release, tag, and requested repository-advisory metadata. These are official APIs or official repository content. Optional hfToken and githubToken values go only to their matching adapter and never appear in output.

Inputs

Provide 1-500 discriminated entities, a stateless preview or persistent baseline mode, a deterministic changeThreshold, one or more feed formats, and concurrency from 1 through 8. Persistent mode requires baselineStoreId; stateless mode forbids it.

{
  "entities": [{"profile": "scholarlyDoi", "doi": "10.1038/s41586-021-03819-2"}],
  "mode": "stateless",
  "feedFormats": ["json"],
  "includeUnchanged": true,
  "concurrency": 1
}

Outputs

Each admitted entity produces one strict dataset result with its normalized identity, terminal status, source status, source evidence, baseline transaction, deterministic summary, stable event ID, and artifact link. A final uncharged run-summary row reports profile/status counts, paid events, baseline commits, failures, and artifacts. Requested KVS files are RESEARCH_WATCH.json, RESEARCH_WATCH.atom, and RESEARCH_WATCH.rss; RESEARCH_WATCH_MANIFEST.json records content type, byte length, SHA-256, schema version, and generation rule.

{
  "recordType": "result",
  "profile": "scholarlyDoi",
  "status": "baseline_created",
  "sourceStatus": "complete",
  "charge": {"event": "entity-check", "chargeable": true, "reason": "baseline_created"},
  "event": {"changeType": "baseline_created", "severity": "info", "summary": "Baseline created for 10.1038/s41586-021-03819-2."},
  "artifactLink": "RESEARCH_WATCH.json"
}

Pricing

The primary pay-per-event business event is entity-check at $0.02. The maximum estimate is the number of terminal entity checks multiplied by $0.02: one check is $0.02, 100 are at most $2.00, and 500 are at most $10.00. Baseline creation, unchanged, changed, and profile-confirmed terminal negatives are chargeable. Source outage, auth failure, unsafe input, malformed responses, duplicate entries, and charge-limit rows are not. Feed generation and the run summary are never charged.

The runtime asks Apify for the preflight event allowance, never schedules beyond that allowance, stops new scheduling when event_charge_limit_reached is returned, uses the exact entity-check spelling only on chargeable push_data calls, and commits baseline state or includes a paid feed event only after charged_count == 1.

State and baseline behavior

Stateless mode creates preview baseline events without writing monitoring state. Persistent records are namespaced by profile and hashed entity identity. A prepared journal is written before billing; the current baseline and history are committed after billing succeeds. Recovery completes charged pending journals. Expected-prior-hash checks refuse concurrent overwrite. Clinical-study absence requires two persistent observations at least six hours apart.

Partial failures and negative results

Mixed batches continue when one profile fails. A failed entity has a stable status, evidence locator, safe error code, and no charge. Scholarly absence requires terminal Crossref and NCBI responses. Hugging Face and GitHub official 404 responses are definitive. ClinicalTrials.gov 404 is uncharged until persistent confirmation. Duplicate normalized identities are not fetched twice.

Resource limits

The Actor accepts at most 500 entities, runs at concurrency 1-8, limits each response to 5 MB and JSON depth to 50, caps GitHub releases/tags/advisories at 20 each, and uses at most 256 MB minimum, 1024 MB default, and 2048 MB maximum memory. Hosted timeout is 900 seconds.

Data handling, privacy, and deletion

Source records are canonicalized in memory. Result rows contain bounded evidence and semantic changes, not full raw responses. Feed artifacts and persistent baselines live in Apify storage selected by the user and follow that storage's retention/deletion controls. See PRIVACY.md.

Source rights and attribution

Every source record names the provider, official request locator, raw hash, attribution, and terms URL. This Actor does not bypass access controls or retrieve paywalled full text. See SOURCE_POLICY.md.

Limitations

Deterministic templates are not expert interpretation. Provider metadata can be incomplete or delayed. materialOnly is a versioned operational filter, not a legal, medical, security, or scientific significance judgment. A moved Git tag is high-severity evidence but is not automatically malicious.

Support and changelog

See SUPPORT.md for support and CHANGELOG.md for release history.