GET /extractions/{id}: the result

Every field this endpoint returns, in every state · processing version 9

The five possible responses

Job state	Status	Body	Headers
queued	200	`{"id", "status": "queued", "version", "created", "queue_position"}`	`Retry-After: 30`
running	200	`{"id", "status": "running", "version", "created", "started"}`	`Retry-After: 30`
done	200	The result, explained below.	`Expires-At`
failed	500	problem+json: `{"type", "title": "Extraction Failed", "status": 500, "detail", "instance", "id", "version", "created", "started", "finished", "expires_at"}`. `detail` says what broke.	`Expires-At`
unknown or expired	404	problem+json, `"title": "Not Found"`

queue_position counts from 1 and excludes the job that is running: 1 means next. Poll every Retry-After seconds.

A complete done result

A shortened example: one performer with one ID copy and one document. The people and numbers are invented. Real results list every performer, copy, document and page.

{
  "status": "done",
  "version": 9,
  "created": "2026-10-18T16:28:39+00:00",
  "started": "2026-10-18T16:37:55+00:00",
  "finished": "2026-10-18T16:44:12+00:00",
  "expires_at": "2026-10-18T19:44:12+00:00",
  "packet_sha256": "ec3ff078b2bdf0d7a76adfd38f2e0ea9326cd062b2e4cd35f0ee3dd95babb3ef",
  "pages": 6,
  "page_map": [
    {"file": "Scene 1/release.pdf", "pages": [1, 2, 3, 4]},
    {"file": "Scene 1/Jane ID.heic", "pages": [5]},
    {"file": "Scene 2/Jane ID.heic", "pages": [5], "duplicate_of": "Scene 1/Jane ID.heic"},
    {"file": "cast list.xlsx", "pages": [6]}
  ],
  "metadata": {"title": "Example Movie 5", "scene": ["1", "2"]},
  "answer": {
    "pages": [
      {"page": 1, "what": "model release, page 1"},
      {"page": 5, "what": "passport photo page"},
      {"page": 6, "what": "cast list"}
    ],
    "performers": [
      {
        "stage_names": ["Jane Dynamite"],
        "legal_name": "DOE, JANE A",
        "date_of_birth": "15 MAR /MARS 99",
        "date_of_birth_parsed": "1999-03-15",
        "id_type": "passport",
        "id_number": "AB123456",
        "id_expiration": "2027 08",
        "id_expiration_parsed": "2027-08-01",
        "production_date": "May 4, 2021",
        "production_date_parsed": "2021-05-04",
        "id_copies": [
          {
            "page": 5, "box_2d": [112, 86, 878, 561],
            "id_type": "passport", "legal_name": "DOE JANE A", "id_number": "AB123456",
            "date_of_birth": "15 MAR /MARS 99", "date_of_birth_parsed": "1999-03-15",
            "expiration_date": "04 AUG /AOÛT 27", "expiration_date_parsed": "2027-08-04",
            "unreadable_fields": []
          }
        ],
        "documents": [
          {
            "pages": [1, 2, 3, 4], "kind": "model release",
            "stage_name": "Jane Dynamite", "legal_name": "Jane A. Doe", "id_number": "AB123456",
            "date_of_birth": "03/15/1999", "date_of_birth_parsed": "1999-03-15",
            "production_date": "May 4, 2021", "production_date_parsed": "2021-05-04",
            "form_date": "5/4/21", "form_date_parsed": "2021-05-04"
          }
        ],
        "notes": null
      }
    ],
    "unassigned": [],
    "production_date": null,
    "production_date_parsed": null
  },
  "id_crops": [
    {"performer": 1, "copy": 1, "page": 5, "status": "verified",
     "note": "passport; DOE JANE A", "jpeg_base64": "/9j/4AAQSkZJRgABAQAA…"}
  ],
  "pdf": "/extractions/ec3ff078b2bdf0d7a76adfd38f2e0ea9326cd062b2e4cd35f0ee3dd95babb3ef/pdf"
}

Top level

Field	Type	Meaning
`status`	string	Always `"done"` in a result.
`version`	integer	The processing version that produced this result: one number for the whole of the processing (assembly, model, prompt, crop checks, date parsing, response shape). `null` for jobs from before versioning.
`created`	string	When the job was submitted. All timestamps are ISO 8601 UTC.
`started`	string	When the extraction began (after waiting in the queue).
`finished`	string	When the extraction ended.
`expires_at`	string	When the job, this result and its PDF are deleted (3 hours after `finished`); same as the `Expires-At` header.
`packet_sha256`	string	SHA-256 of the combined PDF. Equals the job id when no metadata was sent and no file was excluded.
`pages`	integer	Pages in the combined PDF.
`page_map`	array	One entry per uploaded file, in upload order: where its pages went in the combined PDF.
`page_map[].file`	string	The uploaded filename, as sent.
`page_map[].pages`	array of integers	Its page numbers in the combined PDF; empty when the file was excluded.
`page_map[].duplicate_of` sometimes	string	Present when this file has the same bytes as an earlier upload: it was included once, and `pages` points at that copy.
`page_map[].render_notes` sometimes	array of strings	Present when the file wasn't read or rendered the way its name suggests: its content is another type than its extension says (it is read as what it really is), an image had several frames and only the first was included, or an office file could not be rendered as intended (for example, a spreadsheet that could not be set to landscape, so wide rows may split).
`page_map[].excluded` sometimes	string	Present when the file was left out of the combined PDF, with the reason: its content is no accepted type, or it couldn't be read in full (for example `unreadable image (JPEG): Premature end of JPEG file`). The model never saw it; `pages` is empty. On a duplicate of an excluded file it is repeated alongside `duplicate_of`.
`metadata`	object	The metadata fields you sent, unchanged; `{}` if none. A key sent more than once has a list of values.
`answer`	object	What the model found in the packet: below.
`id_crops`	array	Every ID copy cut out of the PDF as a JPEG, with its check: below.
`pdf`	string	Where to fetch the combined PDF: `/extractions/{id}/pdf`.

answer

The model's answer, plus a …_parsed field beside every date (how). Names, numbers and dates are copied as printed on the documents; null means not shown or not legible. An answer with no performers is a valid result, not an error. Every page number refers to the combined PDF.

Field	Type	Meaning
`pages`	array	The model's inventory of the packet, written before anything else: one entry per page.
`pages[].page`	integer	The page.
`pages[].what`	string	What the page shows, in a few words (`"driver's license, front"`, `"model release, page 2"`).
`performers`	array	One entry per real person, with everything about them from every page: below.
`unassigned`	array	IDs, releases or forms the model could not tie to a performer.
`unassigned[].page`	integer	Where it is.
`unassigned[].what`	string	What it is.
`production_date`, `production_date_parsed`	string or null	A production or shoot date the paperwork states for the movie as a whole.

answer.performers[]

Field	Type	Meaning
`stage_names`	array of strings	Every stage name or alias given for this person anywhere, each once.
`legal_name`	string or null	As printed on their government photo ID; from their documents if no ID was found.
`date_of_birth`, `date_of_birth_parsed`	string or null	From the ID (else their documents).
`id_type`	string or null	Driver's license, passport, national ID card, …
`id_number`	string or null	The ID document's own number, character for character.
`id_expiration`, `id_expiration_parsed`	string or null	The ID's expiration date (not its issue date).
`production_date`, `production_date_parsed`	string or null	When this person's content was produced: the production or shoot date on their own release or form, else the date the form is dated.
`id_copies`	array	Every copy of their government photo ID in the packet: below.
`documents`	array	Every release, form, statement or cast-list entry about them: below.
`notes`	string or null	Anything a reviewer must know (documents that disagree, an ID that looks expired).

performers[].id_copies[]

One entry per copy of a government photo ID: the front (photo and name), plus the back when the front shows no date of birth.

Field	Type	Meaning
`page`	integer	Where the model says the copy is. See which page number to trust.
`box_2d`	array of 4 integers	The ID's outline on that page, `[x1, y1, x2, y2]` on a 0–1000 scale from the top-left corner.
`id_type`	string	The document type on this copy.
`legal_name`, `id_number`	string or null	As printed on this copy.
`date_of_birth`, `date_of_birth_parsed`	string or null	As printed on this copy.
`expiration_date`, `expiration_date_parsed`	string or null	As printed on this copy.
`unreadable_fields`	array of strings	Which of `legal_name`, `id_number`, `date_of_birth`, `expiration_date` are on the copy but not fully legible.

performers[].documents[]

Field	Type	Meaning
`pages`	array of integers	Where the document is.
`kind`	string	`"model release"`, `"2257 form"`, `"cast list"`, …
`stage_name`, `legal_name`, `id_number`	string or null	As the document gives them for this person.
`date_of_birth`, `date_of_birth_parsed`	string or null	As the document gives it.
`production_date`, `production_date_parsed`	string or null	Production or shoot date on the document.
`form_date`, `form_date_parsed`	string or null	The date the form itself is dated or signed.

id_crops[]

After the answer, each ID copy is cut out of the PDF (with a 20% margin around its outline, at 300 dpi) and shown to the model again on its own: is this an ID, and does it show the copy's recorded date of birth? A crop that fails is re-cut from the neighbouring pages and checked again. The checks never change the answer; they tell you whether the image can be trusted as evidence. Keep the images: they disappear with the job.

Field	Type	Meaning
`performer`, `copy`	integer	Which ID copy: `answer.performers[performer-1].id_copies[copy-1]`.
`page`	integer	The page the crop was cut from: for a repaired crop, the corrected page.
`status`	string or null	`verified`: an ID, showing the recorded date of birth. `repaired`: the first cut wasn't, but a neighbouring page was, and the crop was taken from there. `failed`: nothing could be vouched for. `null`: not checked.
`note`	string or null	One line for a reviewer: the ID type and name read off the crop (`"passport; DOE JANE A"`), where a repaired crop moved to, `back side` for the back of a card, or what a failed check saw instead (`"signed code of conduct page"`).
`jpeg_base64`	string	The crop, a base64-encoded JPEG.

How dates are parsed

Every date field in answer appears twice: as printed, and as …_parsed, an ISO YYYY-MM-DD date, or null when the printed text can't be read as one date. Use the printed value as the evidence and the parsed one for arithmetic.

Printed	Parsed	Rule
`5/4/21`	`2021-05-04`	Numeric dates are read month first...
`21/11/1986`	`1986-11-21`	...unless the first number can't be a month: then day first.
`2019-05-04`	`2019-05-04`	Year-first dates are read as written.
`15 MAR /MARS 99`	`1999-03-15`	A second-language month (Canadian IDs) is ignored. Two-digit years 50–99 are 1900s, 00–49 are 2000s.
`01/02/49` as a birth date	`1949-02-01`	A birth date is never in the future: one that would be is moved back a century.
`2027 08`	`2027-08-01` (expiration) / `2027-08-31` (birth date)	A date without a day rounds the cautious way: birth dates to the month's last day, expiration and production dates to its first.
`08171990`	`1990-08-17`	Eight digits: month-day-year, else day-month-year.
`May 4, 2021 & May 5 2021`	`2021-05-04`	A production date listing several shoot days gives the earliest: the performer must be 18 on every one. If any listed day can't be read, the parsed value is `null`.

Which page number to trust

All page numbers refer to the combined PDF. The model's own page numbers, in id_copies[].page and documents[].pages, can be off by one on long packets. For an ID, prefer id_crops[].page: it is where a crop was actually checked. To trace a page back to the file you uploaded, look it up in page_map.