groundgate

LLMs propose facts. groundgate decides which ones you can trust.

groundgate is a deterministic admission layer for LLM document extraction. You give it a document, a schema, and the facts an extractor proposed. Every fact comes back with one of three outcomes:

• admitted: every check passed;
• needs_verification: nothing failed, but a person should look, and the decision says where;
• rejected: a check failed, with a stable reason code.

Every run writes a receipt with hashes of every input. Anyone can re-derive it byte for byte.

The goal is to admit as many fields as possible with the right value, while wrong values almost never get in. I judge every rule by two numbers: the facts admitted with a right value, and the wrong extractions admitted (escapes). Reviews are the cost between them: each one is a value that a person confirms by hand.

Why this exists

Current models get most extracted values right. In the benchmark below, 4,233 of 4,350 extractions were correct. The other 117, about 1 in 37, each cited a real place in the document, so nothing on the page tells you which ones are wrong.

Grounded extraction tools check that the quoted text exists in the source. They don't check the value you store. LangExtract is the clearest example, and a good tool. It aligns each extraction_text to the source and reports MATCH_EXACT, MATCH_FUZZY and so on. It never looks at the attributes that hold the typed value.

A real case from the benchmark: IRS Publication 560 states two limits a few lines apart.

The limit on elective deferrals, other than catch-up contributions, is $23,500 for 2025 ...
The limit on salary reduction contributions, other than catch-up contributions, is $16,500 for 2025 ...

The first is the 401(k) limit. The second is the SIMPLE plan limit. Asked for the 2025 elective deferral limit, 4 of 14 runs proposed both, and LangExtract marked both quotes MATCH_EXACT. groundgate sent both to a person with CONFLICTING_CANDIDATES, because two proposals for one field disagree. In Publication 15-B, one model stored "$1 million" as 1, also an exact match, and groundgate flagged it SCALE_WORD.

The same gap shows up on the document I used while building the library. I ran Gemini 3.6 Flash and GPT-OSS 120B through LangExtract on two pages of IRS Publication 590-A. LangExtract aligned all 48 extractions as MATCH_EXACT. groundgate then:

• rejected Gemini's reading of a misprinted $252,0000 as 2,520,000;
• flagged two GPT-OSS facts that took a Roth IRA limit from the spousal-deduction sentence; the numbers match, but the text says "more than" where the field says "at least";
• listed a required field both models missed;
• admitted 44 facts, all with the right value.

One admitted fact still cites the wrong rule. Its number is correct, its comparator matches, and no span check can tell. I checked every decision by hand; the details are in examples/irs-590a, which is also the page in the screenshot.

I measured the gap on purpose before writing the library. On 51 facts from two public-domain government documents, LangExtract's strictest setting (MATCH_EXACT only) accepted 100% of planted value and unit errors whose quote was correct. groundgate caught all of them and rejected none of the correct facts. The method and full tables are in spikes/m0_5.

LangExtract locates the text. groundgate checks the number in it.

Benchmark

Seven models (Gemini, GPT and Claude, each through its own CLI) ran through LangExtract on 30 public-domain FDA drug labels, NTSB accident reports and IRS publications, at two chunk sizes. A person checked every gold fact. These are spec 0.1's numbers on the first set, as published. Of 4,350 extractions, 117 were wrong: the wrong value, a value for a field the document doesn't state, the wrong unit, or not a number.

A gold fact is one value that a document states for a field, so a field with 12 strengths counts 12 times, and each fact counts once in each run. groundgate admits fewer right facts than MATCH_EXACT, and lets in 8 wrong values instead of 115. With the reviews, 96.5% (3,742/3,878) of gold facts get a right value.

Of the 526 extractions groundgate sent to a person, 92 were wrong; a random sample that size would have held about 14. groundgate also rejected 71 correct values that LangExtract had aligned to a place that doesn't state them: the value was right, the evidence wasn't.

Most catches (84 of 109) were CONFLICTING_CANDIDATES: a model proposed two values for a single-valued field, and groundgate sent both to review instead of picking one.

Six of the 8 that got through are the case groundgate says it cannot catch (see What it does not do). Five models took metoprolol's heart failure maximum (200 mg) as the hypertension maximum, and one took lisinopril's renal impairment dose as the usual starting dose. Each number is real and cited correctly; it belongs to another condition. The other two read levothyroxine's 1.6 mcg/kg/day as a flat 1.6 mcg, which the unit check allows. Putting all seven models through one gate caught one more, because when they were wrong they mostly agreed. Spec 0.2 goes after both: keyed fields check the condition, and a dose per kilogram no longer passes as a flat dose.

The benchmark also found a bug. When LangExtract's chunker cut "29.97" after "29.", groundgate read the span as 29, which its own spec forbids. It's fixed, and three wrong altimeter settings no longer get through.

Tables and planted-error results: bench/RESULTS.md. Method and caveats: bench/README.md. The write-up, with what got through and why: Grounding is a contract, not a citation.

Spec 0.2 on a second set

Spec 0.2's rules were written from the escapes above, so they can't be measured on the same documents. I picked 101 new ones by rules frozen before anyone read them: FDA labels with doses per indication or per kilogram, IRS publications and Federal Register rules with "increased from", "must not be more than" and "$1.2 million". The same seven models ran at both chunk sizes. Every candidate was decided twice, by spec 0.1 (the released 0.1.0 wheel) and by spec 0.2. Pooled over 14 runs on the 74 documents checked in full:

642 of 0.1's 690 false rejects are in the 19 documents picked for amounts like "$1.2 million", which 0.1 could not match. 82 of its 253 escapes were a right value under the wrong indication; 0.2 lets 10 of those through. It pays in review load. The 72 that still got through are listed one by one in bench/set2/RESULTS.md, and two rules that need work are open for the next version (#51, #52).

This set shows the gap that matters most. Spec 0.2 admits a right value for 53.2% of gold facts, and the reviews bring it to 86.8% (4,873/5,614). Closing that gap without new escapes is the work of the next versions: the 0.4 key clear and a basis that the extractor states (#125) both aim at it.

Spec 0.5 on new 10-K filings

Spec 0.5 lets the extractor cite each part of a value: the sign, the scale, the unit, the field and the key. I measured it on 20 10-K filings that the rules did not come from, and on the IRS filing-status set. One extractor (Claude Haiku 4.5) ran once per set, and every candidate was decided by spec 0.4 (the released 0.4.0 wheel) and by spec 0.5. A person labeled every admitted value blind.

Two rules still cost right values, and neither adds an escape: a unit item fails when the unit is already next to the number, and a scale quote must match the case of the text. Without the unit check, spec 0.5 admits 33 right values on the 10-K set. Both rules are open for 0.6 (#145). bench/evidence/RESULTS.md has every number.

Quickstart

pip install groundgate

import groundgate as gg

text = "The IRA contribution limit is $7,000. The deduction phases out above $79,000."
schema = {
    "fields": {
        "ira_limit": {"type": "integer", "unit": "USD"},
        "phaseout_start": {"type": "integer", "unit": "USD"},
    }
}


def cite(field, value, quote):
    """A candidate citing the first place `quote` appears, as UTF-8 byte offsets."""
    start = text.encode().index(quote.encode())
    span = {"start": start, "end": start + len(quote.encode()), "text": quote}
    return {"field": field, "value": value, "unit": "USD", "evidence": span}


candidates = [
    cite("ira_limit", "7000", "$7,000"),
    cite("ira_limit", "70000", "$7,000"),  # a digit too many
    cite("phaseout_start", "79000", "$79,000"),  # "above" is not "at"
]
receipt = gg.admit(text, schema, candidates)
for d in receipt.decisions:
    print(f"{d.outcome:<19} {d.field:<15} {d.value:<6} {' '.join(d.codes)}".rstrip())

print(gg.verify(receipt.to_dict(), text, schema, candidates).ok)

admitted            ira_limit       7000
rejected            ira_limit       70000  VALUE_NOT_IN_EVIDENCE
needs_verification  phaseout_start  79000  QUALIFIED_VALUE
True

The core has no dependencies. groundgate[pdf] adds pdfminer.six for PDF text, and groundgate[langextract] installs LangExtract. docs/guide.md covers schemas, policies, candidates and the Python API.

With LangExtract

import langextract as lx
from groundgate.adapters.langextract import admit_document

result = lx.extract(text_or_documents=text, prompt_description=prompt, examples=examples)
receipt = admit_document(result, schema)

Each extraction's value and unit attributes are checked at the place LangExtract aligned its extraction_text. An extraction LangExtract could not align is rejected NO_EVIDENCE.

From the command line

groundgate extract p590a.pdf --pages 1-2 -o doc.txt --layout layout.json
groundgate admit   doc.txt schema.json candidates.json -o receipt.json
groundgate verify  receipt.json doc.txt schema.json candidates.json
groundgate report  receipt.json doc.txt schema.json candidates.json --layout layout.json -o report.html

extract turns a PDF, HTML, XML or text file into the NFC text everything else reads, and records the page and box of every word. Text a reader can't see (proof marks drawn off the page, hidden HTML) is dropped, and superscripts are marked so that 10⁹ reads 10^9, never 109. Pass - as the document to read it from a pipe. report refuses to render a receipt that doesn't re-derive from its inputs. examples/irs-590a runs the whole pipeline on an IRS publication with cached model output, so it reproduces without an API key.

What it checks

Checks run in a fixed order. The first failure rejects the fact.

A fact that passes every check can still be flagged for a person:

Number matching is collision-safe. The span 500 mg inside 1,500 mg never reads as 500, a span that stops at 29. inside 29.97 never reads as 29, and a malformed number like the $252,0000 printed in IRS Publication 590-A never equals 252,000 or 2,520,000. When a span misses the value but its quote occurs exactly once elsewhere with the right value and unit, groundgate moves the evidence there and records EVIDENCE_REANCHORED.

The rules are in SPEC.md. conformance/ holds 31 language-neutral vectors that pin every code, so another implementation can prove it agrees.

Receipts

{"groundgate": "0.5",
 "document": {"id": "irs-p590a-2025-pages-1-2", "sha256": "sha256:2fb2495b5dfb...", "source": null},
 "references": [],
 "schema_sha256": "sha256:...", "policy_sha256": "sha256:...",
 "decisions": [{"candidate_id": "Gemini_3.6_Flash_Medium/0", "field": "ira_limit_2025",
                "outcome": "admitted", "codes": [], "value": "7000", "unit": "USD",
                "source": "document", "ref": null, "url": null,
                "evidence": {"start": 2317, "end": 2323}, "parts": [], "missing": [],
                "candidate_sha256": "sha256:a15c9d8b..."},
               "..."],
 "coverage": [{"field": "roth_phaseout_single_2025_end", "code": "REQUIRED_FIELD_MISSING"}],
 "judgments": [],
 "summary": {"admitted": 44, "needs_verification": 3, "rejected": 1},
 "receipt_sha256": "sha256:..."}

Hashes are SHA-256 over RFC 8785 canonical JSON, so a receipt means the same thing in any language. groundgate verify re-runs every decision from the inputs and fails on any difference: a changed document, schema, policy, candidate or outcome.

What it does not do

• It does not judge meaning. If the model reports a number that really is in the sentence but belongs to another field, every span check passes: a model that reads an age-50 limit of $8,000 as the IRA limit cites a real "$8,000" with the right unit. A second proposer plus CONFLICTING_CANDIDATES catches it only when the models disagree, and in the benchmark they mostly agreed: six of the eight escapes were this case. When the field depends on a condition the schema can name, such as a drug's indications, a keyed field checks it (spec 0.2).
• No dates, arrays of records, or cross-document checks in spec 0.5.
• English number formats only. A decimal comma (1.234,56), space groups (1 234) and number words (five) have no value, so a fact that cites them is rejected. The guide lists what is read. Locale packs for other formats are a design (docs/design/locale-packs.md, #60).
• No OCR. Scanned PDFs need a text layer first (for example ocrmypdf).
• No model calls in the decision. groundgate never asks an LLM whether an LLM was right. Spec 0.4 reads a judge's answers only as recorded inputs, so a receipt still re-derives byte for byte. This judge block is experimental: its one measure on real extractor output is the key clear on 10-K filings (#120). groundgate-calibrate (#112) measures the thresholds on your own documents.

Status

Alpha. groundgate 0.5.0 implements spec v0.5. CHANGELOG.md lists what each version changes, with its measure. 0.5 makes the evidence a list of items (#141): a value can cite its sign, scale, unit, field and key, and outside sources, under the app's policy. 0.4 added recorded judgments (#116), an experimental way for a judge's answers to clear a key flag or add doubt, with no model call in the decision. 0.3 brought a wider change rule (#51), abbreviation dots that no longer end a sentence for qualifiers (#63, #77), Indian digit grouping (#59), ASCII digits (#94), and keys that stop at a label line (#52) or read a table by its lines (#87). A receipt names the spec version it was decided under, and only that version verifies it.

Locale packs for number formats outside English are a design, not yet code. Of the hybrid decisions design, 0.4 has the recorded judgments for the key and field questions. The audit sample and model-proposed locations are not code yet.

License

Apache-2.0.