common.skip_to_content
common.brand common.back_to_home
disclosure.banner disclosure.banner_link.

figures.title

figures.lede

figures.provenance.measured_at 02-09-2026 14:53 · figures.provenance.revision 1cd55feac3fd748e · figures.provenance.backend gliner · figures.provenance.threshold 0.5 · figures.provenance.raw

figures.recall.title

figures.recall.nl: figures.recall.found_of

figures.recall.kind figures.recall.found figures.recall.rate
PERSON 7/7 1
DATE_TIME 4/4 1
NL_BSN 3/3 1
ADDRESS 6/6 1
PHONE_NUMBER 3/3 1
EMAIL_ADDRESS 2/2 1
IBAN_CODE 2/2 1
NL_LICENSE_PLATE 1/1 1
ORGANIZATION 1/1 1
NL_KVK 1/1 1
NL_BTW 1/1 1

figures.recall.en: figures.recall.found_of

figures.recall.kind figures.recall.found figures.recall.rate
PERSON 2/2 1
DATE_TIME 1/1 1
EMAIL_ADDRESS 2/2 1
IBAN_CODE 1/1 1
CREDIT_CARD 1/1 1
ADDRESS 1/1 1

figures.over_masking.title

figures.over_masking.body

figures.over_masking.kind figures.over_masking.kept figures.over_masking.wrongly_masked
isbn 2/2 0
imei 0/1 1
ean 1/1 0
order_reference 2/2 0
invoice_number 1/1 0
iso_date 1/1 0
place 0/1 1
product 0/1 1

figures.over_masking.note

figures.cost.title

This is how much text changed, not how much worse the answer got. The answer-quality figure needs a task suite run scrubbed and unscrubbed and judged, and it is not built.

figures.cost.kind figures.cost.share
juridisch0.3189
zorg0.5434
hr0.4278
financieel0.3972
gemeente0.4326

figures.corpus.title

figures.corpus.body

{
    "name": "sluiswerk-nl-en-v1",
    "version": "2026-09-02b",
    "note": "Every value here is invented. Checksums are valid so the deterministic recognizers really fire rather than appearing to.",
    "public": {
        "nl": {
            "documents": 6,
            "domains": [
                "juridisch",
                "zorg",
                "hr",
                "financieel",
                "gemeente"
            ]
        },
        "en": {
            "documents": 2,
            "domains": [
                "hr",
                "legal"
            ]
        },
        "hard_negatives": {
            "documents": 10,
            "kinds": [
                "isbn",
                "imei",
                "ean",
                "order_reference",
                "place",
                "product",
                "invoice_number",
                "iso_date"
            ],
            "note": "place and product are here because a village name and a product name are what a person-recogniser most wants to claim. They were declared in this file before they existed in the corpus, which is the kind of gap this measurement is supposed to close rather than create."
        },
        "note": "Document counts here were overstated (12 nl, 6 en) against a corpus of 6 and 2. They are corrected to what is measured. The recall figures were always computed over the real corpus, so they did not change; what changed is that the sample size a reader is told about is now the sample size. measure.py refuses to run when this file stops describing the corpus."
    },
    "held": {
        "note": "Real Dutch text obtained under an agreement is measured here and never leaves this machine. Only figures derived from it are published.",
        "present": false,
        "documents": 0
    }
}

figures.history.title

figures.history.measured_at figures.history.revision figures.history.recall_nl figures.history.over_masked
2026-09-02T14:53:05+0000 1cd55feac3fd748e 1 0.3
2026-09-02T10:00:37+0000 615f905f2f7cd8a5 1 0.125