Scientific Table and Footnote Extractor

Extract evidence-linked tables from explicit, authorized HTML and born-digital PDF sources. Scanned pages return an uncharged ocr_required record; version 0.1 has no OCR.

What this does not do

• No OCR, browser rendering, paywall bypass, or arbitrary article crawling.
• No automatic scientific interpretation.
• No conversion of censored symbols, ranges, missing markers, or footnotes into invented values.
• No rights determination beyond retaining source indicators and requiring the user's declaration.

Sources and authorization

Each of the 1–50 sources must be an explicit HTTPS URL or a user-selected Apify key-value-store record. Every source requires a rightsBasis of user_owned, open_license, public_domain, or permission with evidence. AP-029 indicators may inform that declaration but never replace it.

The Actor validates public URLs against SSRF rules, limits redirects and response bytes, never follows article links, and accepts only detected HTML or PDF matching the declared hint. PDF support is bounded to born-digital pages; browser rendering and OCR are excluded.

Input

{
  "sources": [
    {
      "url": "https://flintglade-tools.github.io/apify-fixtures/tables/scientific-table.html",
      "rightsBasis": {
        "type": "user_owned",
        "evidence": "Flintglade-controlled public fixture"
      },
      "formatHint": "html"
    }
  ],
  "includeFootnotes": true,
  "retainRaw": false,
  "concurrency": 1
}

Optional tableSelectors target at most 20 HTML selectors. Optional pageRanges target at most 50 PDF ranges. They are mutually exclusive by format. maxBytesPerSource defaults to 50 MB and cannot exceed 100 MB; maxPdfPages defaults to 200 and cannot exceed 1,000; concurrency is 1–2.

Output

Each admitted table creates one default-dataset result with a stable table ID, caption, section or page and bounding box, span-preserving cells, header graph, units, raw scientific symbols, parsed numeric candidates, explicit footnote links, unassigned nearby footnotes, source evidence, and canonical hashes.

{
  "recordType": "result",
  "status": "extracted_table",
  "entity": {"key": "sft_35b...", "inputIndex": 0},
  "charge": {"event": "extracted-table", "chargeable": true},
  "data": {
    "caption": "Example values",
    "rowCount": 3,
    "columnCount": 2,
    "units": ["mg"],
    "artifacts": [
      {"key": "tables/sft_35b....json", "contentType": "application/json"},
      {"key": "tables/sft_35b....csv", "contentType": "text/csv"}
    ]
  }
}

JSON is authoritative. CSV is produced only for an admitted rectangular representation and prefixes spreadsheet-formula leaders in the display export while JSON preserves the raw value. KVS contains JSON and CSV per table, optional raw-value JSON, and SOURCE_MANIFEST.json, each referenced with byte length, SHA-256, schema version, and generation rule. Every run ends with one uncharged summary row.

Pricing

The only business event is extracted-table at the documented launch price of $0.05 per admitted table. Maximum business-event cost is admitted tables × $0.05, plus Apify platform usage. Ten admitted tables cost at most $0.50 in business events; fifty cost at most $2.50. Captions, figures, false candidates, source failures, rights rejection, oversized inputs, parse failures, ocr_required, summaries, and charge-cap records are uncharged.

The price is economically sensible for launch because the charged unit is a schema-valid reusable table with JSON/CSV artifacts and provenance, while the 5-cent unit leaves operating headroom for the default 4 GB PDF workload and bounded retries without charging for failed work. Revisit price only after hosted cost-per-table, abort-rate, and customer-value evidence exists; do not lower safety ceilings to defend a price.

The Actor consults Apify's charging manager before reading. It stops scheduling and withholds further extracted value when the event limit is exhausted. The run summary reports the SDK-confirmed event count; apify-default-dataset-item must remain disabled.

The frozen publication gate covers 100 CC BY publisher-JATS tables split evenly between HTML and born-digital PDF. It records 0.991086 cell recall, 1.0 header accuracy, 24/24 explicit footnote links, 70/70 units, and zero invented links. See benchmark/README.md for the review protocol and hashes. PPE_RECONCILIATION.md records every deterministic billing branch and isolates the one post-upload Console check; no hosted result is claimed before upload authorization.

State and failure behavior

Version 0.1 is stateless. Stable table IDs derive from the raw source hash and evidence locator; canonical table hashes exclude run IDs, timestamps, retry metadata, and platform metadata. Source failures are explicit and do not suppress successful sibling sources. Ambiguous footnotes remain unassigned instead of being guessed.

Resources and data handling

The launch profile is 1,024 MB minimum, 4,096 MB default, 8,192 MB maximum, and a 1,800-second hosted timeout. Source bodies are processed in memory under byte/page caps. Raw source bodies are not retained; retainRaw stores only raw extracted cell values. Dataset and KVS retention follow the user's Apify account policy and can be deleted in the Apify Console/API.

URLs are emitted with query values redacted. Secrets, cookies, authorization headers, and private storage tokens are never included in results. See PRIVACY.md, SOURCE_POLICY.md, and THREAT_MODEL.md.

Limitations

HTML span/header extraction is deterministic. PDF extraction uses text objects and geometry and will not recover image pixels, handwriting, or every visually implied layout. PDF footnotes without explicit marker evidence remain unassigned. Parsed numeric candidates do not replace raw text and are not scientific conclusions.

Support: support@flintglade.com. Release history is in CHANGELOG.md.