Skip to content

Your AI makes financial claims. Artizar makes them provable.

Send the claim your model made. Get a deterministic verdict, the filing evidence, and a signed record that verifies offline.

Request API accessQuickstart: 5 steps, curl
curl -s https://api.artizar.ai/v1/verifications \
  -H "Authorization: Bearer $ARTIZAR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"ticker": "TSLA", "metric": "revenue",
       "period": "FY2023", "claimed_value": "81462000000",
       "unit": "USD"}'
201 Createdsealedlog position 2628
Claimed: Tesla revenue, FY202381,462,000,000
Filed with the SEC · 0001628280-26-00395296,773,000,000
Difference, claim less filing (15.8%)(15,311,000,000)
contradictedrule R2, primary-source contradiction, policy-v7

ed25519 12b8045ca1b000ca86467064b3ec9fb2cd321067f520608d… · record 466c727c-28df-48ee-8fc0-97e05dcdb335

A real sealed record: claude-haiku-4-5’s own answer for Tesla’s FY2023 revenue, from FilingsCheck v1.3.

  • 2530 sealed grades in FilingsCheck v1.3
  • 2032 of 2032 adversarial probes graded as expected
  • 0 false greens

The errors that survive grounding

Retrieval and grounding work on the input side: they decide what the model reads, and the same system that wrote the answer vouches for them. The sentence that reaches your user is the output side, and it can carry the wrong period, the wrong scale, or a dropped line item with a correct-looking citation underneath.

Artizar checks the claim that comes out, against the filing as filed. Three of the planted claims from the latest benchmark run, one for each kind of error:

  • Period

    Meta Platforms revenue, FY2023

    The company’s FY2022 revenue, given as FY2023

    Claimed
    116,609,000,000
    Filed
    134,902,000,000
    contradictedlog position 2672

  • Scale

    Alphabet revenue, FY2023

    The filed figure, a thousand times too large

    Claimed
    307,394,000,000,000
    Filed
    307,394,000,000
    contradictedlog position 2668

  • Line item

    Oracle gross profit, FY2024

    Gross profit as last tagged, for the year ended 2018-05-31, given as FY2024

    Claimed
    24,287,000,000
    Filed
    no current fact
    no evidencelog position 2793

A claim with no current fact behind it is never matched to an old one: Oracle stopped tagging gross profit years ago, so the verdict is no evidence, not a green.

How a verdict is made

One call in; a verdict, its evidence, and a sealed record back.

  1. Evidence

    Artizar resolves the issuer, the metric, and the fiscal period, then pulls the XBRL facts behind the claim (the structured data companies file with the SEC) straight from EDGAR, the SEC’s filing system. Every fact keeps its accession number (the SEC’s filing id), its concept, and a content hash.

  2. Rules

    A versioned scoring policy maps that evidence to a tier. Green means the filing confirms the claim, red means it contradicts it, yellow means the evidence could not settle it. A refusal names its reason and carries no tier. The scorer writes the rationale from the rule that fired; no language model is in the path, so the same claim gets the same verdict.

  3. Notes

    A verdict can carry up to 8 context notes: the caveats an analyst would add, such as which fiscal year a figure belongs to or which revenue concept the company reports. The same versioned rule table writes them, and they never change the tier.

  4. Record

    The verdict is sealed with an Ed25519 signature, and a commitment to it is entered in a public, append-only log. Once the log head is published and cosigned by a witness, the record’s position comes with a proof. The next section takes one record apart.

Same claim, two runs

On 2026-10-06, one planted claim (Apple’s FY2023 revenue, stated 1,000 times too high) was sealed in two separate benchmark runs, 6 minutes and 57 seconds apart.

The same claim sealed in two runs: every verdict field matches; only the record id, seal, log position, and time differ.
Fieldclaude-haiku-4-5 runclaude-sonnet-5 runMatch
claimed_value · same383285000000000 USD383285000000000 USDsame
outcome · samecontradictedcontradictedsame
rule · sameR2, primary‑source contradictionR2, primary‑source contradictionsame
verified_value · same383285000000 USD383285000000 USDsame
evidence[0] hash · same718cd90e0540b1b9…718cd90e0540b1b9…same
context_notes · samefiscal_frame, candidate_consultedfiscal_frame, candidate_consultedsame
scoring_policy · samepolicy‑v7policy‑v7same
record_id · differse048d5ec…4be27ace…differs
seal · differs921655a5b3789313…9afd2911fe8b1ab7…differs
log position · differs26652980differs
sealed at · differs21:17:23Z21:24:20Zdiffers
context_notes[0] · notes-v6

fiscal_frame

FY2023 is the filer's own fiscal label: the bound fact's period ends 2023-09-30, not a calendar-period boundary. Fiscal and calendar figures for the same label can differ.

AAPL revenue FY2023, confirmed · record 9e921248…, FilingsCheck v1.3

A record you can hand to anyone

Every verification is sealed, and a commitment to it is entered in the public log. Months later, after the model and the prompt have changed, the record still shows exactly what was checked, against what, and why. This one is a confirmed claim from FilingsCheck v1.3.

Entry 2664 · sealed 2026-10-06 21:17:23 UTCclaude-haiku-4-5 · USD
Claimed: Apple revenue, FY2023a planted control claim
383,285,000,000
Filed with the SEC, 2022-09-25 to 2023-09-30us-gaap:RevenueFromContractWithCustomerExcludingAssessedTax
383,285,000,000
Difference, claim less filing
0
confirmedtier green · rule R4 · policy-v7

ed25519 330fb6aeb2cb3437ac293353bc07e27b9880a9580d894c76498b6dae7d016f09976aad4dca1dbccf3e0679d32d2500f3b3a5fa1ff2627af0237ec53d073a7809commitment 0bcfb7c50563c298c3a974568ec5b40caed424065ad54a193eaee08beee0068erecord 9e921248-61f3-4e85-b998-fdac5cd3da0d · log position 2664

  1. ClaimAAPL revenue, FY2023

    What the model said, as structured fields: issuer, metric, period, value.

  2. Evidenceaccession 0000320193-25-000079 · hash 718cd90e0540b1b9…

    The XBRL fact as the company filed it, with a content hash of exactly what was compared. The chain never passes through a data vendor’s copy or a news story about the 10-K.

  3. RuleR4, primary-source confirmation · policy-v7

    The versioned rule that decided the verdict and wrote its rationale.

  4. Notesfiscal_frame, revenue_tag

    The caveats an analyst would add. They never change the tier.

  5. Sealed25519 330fb6aeb2cb3437…

    A signature over the verdict, the record id, and the commitment. Check it against the published seal key.

  6. Commitment0bcfb7c50563c298…

    The hash of the record you hold, entered in the public log. If it matches, nothing was edited after the fact.

  7. Log position2664

    An inclusion proof places the entry under a log head a witness has cosigned, so the record cannot quietly disappear. The proof is available once that head is cosigned, shortly after sealing.

Check it without trusting us

Download this record with its proof and check it yourself. A short Python script, using only the widely used cryptography package, checks the seal against the published key, recomputes the commitment from the record, and confirms that the witness-cosigned public log contains it. No Artizar account, software, or service is involved, and changing one character of the record makes a check fail.

Design partners receive the sealed benchmark bundles, and one command re-checks an entire run the same way.

Offline, against the published keys

$ curl -sO https://artizar.ai/records/9e921248-61f3-4e85-b998-fdac5cd3da0d.json$ curl -sO https://artizar.ai/records/verify_record.py$ curl -sO https://artizar.ai/.well-known/artizar-keys.json$ pip install cryptography$ python verify_record.py 9e921248-61f3-4e85-b998-fdac5cd3da0d.json artizar-keys.jsonrecord 9e921248-61f3-4e85-b998-fdac5cd3da0d: revenue FY2023, claimed 383285000000 USD, confirmed  ok   seal  ok   commitment  ok   tree head signed  ok   witness cosigned  ok   included in the log

No verification system can promise that a model is right every time. Artizar promises something you can check: every claim gets a sealed verdict, every verdict with evidence traces to the filing, and a refusal says why it could not. The engine is closed; the method is not. The keys are public, and the scoring rules and record schemas ship to every design partner.

Calculated EBITDA, from the figures the company filed

EBITDA and free cash flow are checked against a published formula over the GAAP figures the company filed. A company’s own Adjusted EBITDA adds back items the company chooses, so it is a different number.

Here a claim quotes Uber Technologies’ reported Adjusted EBITDA for FY2024. The filed components sum to a different figure, so the verdict is contradicted, and a note says why. Uber Technologies’ figure is a legitimate company metric; calculated EBITDA is a different one.

sealed 2026-10-06 21:18:18 UTC · log position 2828UBER · ebitda_calculated · FY2024 · USD
us-gaap:NetIncomeLossEDGAR 0001308179-26-000303
9,856,000,000
us-gaap:IncomeTaxExpenseBenefitEDGAR 0001543151-26-000015
(5,758,000,000)
us-gaap:InterestExpenseNonoperatingEDGAR 0001543151-26-000015
523,000,000
us-gaap:DepreciationDepletionAndAmortizationEDGAR 0001543151-26-000015
711,000,000
Calculated EBITDA, the verified value
5,332,000,000
Claimed: Uber Technologies’ Adjusted EBITDA
6,484,000,000
contradictedtier red · rule R2 · policy-v7

calculated_vs_adjusted

The verified figure is the published ebitda_calculated formula over GAAP components, not the filer's own non-GAAP adjusted figure (e.g. Adjusted EBITDA), which adds back filer-chosen items and can differ materially.

What it verifies, and what it refuses on the record

Every verdict has one of four outcomes: confirmed green, contradicted red, no evidence yellow, or not verifiable, which is a refusal and carries no tier.

Verifies

Financial-statement figures for US public companies, by ticker or CIK (the SEC’s company id), read from the XBRL data in their SEC filings. Annual periods use the company’s own fiscal label (FY2023) or the calendar year (Y2023); quarters work the same way (Q1FY2024, Q1Y2024). Balance-sheet figures are checked as of a date.

Income statement
  • revenue
  • gross_profit
  • operating_income
  • rnd_expense
  • interest_expense
  • income_tax_expense
  • net_income
Per share
  • eps
Balance sheet, as of a date
  • cash
  • total_assets
  • stockholders_equity
  • shares_outstanding
Cash flow
  • operating_cash_flow
  • capex
  • depreciation_amortization
Calculated, not adjusted
  • ebitda_calculated
  • free_cash_flow_calculated

Computed from the filed GAAP components by a published formula. A company’s own “Adjusted EBITDA” is a different number, and a verdict says so.

17 metrics in the metric lexicon, currently lexicon-v7. It widens only through a reviewed change, with adversarial probes for every addition.

Refuses, and says why

A refusal is a sealed verdict too: outcome not verifiable, no tier, and a reason you can branch on.

ReasonMeaning
unresolved_subjectThe ticker doesn't identify one company. Retry with the CIK.
unmapped_metricThe metric has no published definition yet (total_debt today).
unsupported_valueThe claimed value can't be typed for the metric, so nothing is compared.
unsupported_marketThe claim is about a filing outside US SEC filings.
ambiguous_fiscal_labelThe company's own annual filings give the fiscal period conflicting labels. Refused, never guessed.

Answers “no evidence” rather than guessing

  • A fiscal fourth quarter, which has no standalone XBRL fact, or a quarter not filed yet.
  • A concept the company stopped tagging, where a naive parser would serve the last stale value.
  • A calculated metric with a missing component. It is never computed from a partial sum.

Measured in public: FilingsCheck v1.3

FilingsCheck asks 8 models for figures reported in SEC filings and runs every answer through the production API. The scorecard measures Artizar itself: planted falsehoods the engine must catch, and true controls (correct figures) it must not flag. The model headlines are supported over asserted; abstentions leave the denominator.

A low headline means the engine caught that model’s errors, and each catch has its own sealed record. Tasksets change between editions, so compare headlines within one edition only. Read the full edition.

FilingsCheck v1.3, published 2026-10-07: engine scorecard and model headlines
FilingsCheck v1.3 · 2026-10-07Figure
adversarial probes graded as expected2032 / 2032
false greens: planted falsehoods graded confirmed0
sealed grades, each with its own proof2530
claude-haiku-4-5, supported of asserted12.7% (8/63)
claude-sonnet-5, supported of asserted71.7% (43/60)
gpt-5.4-mini, supported of asserted41.1% (23/56)
gpt-5.4, supported of asserted84.4% (38/45)
claude-sonnet-5-5, supported of asserted94.3% (66/70)
claude-opus-5-5, supported of asserted95.8% (69/72)
claude-fable-5-1, supported of asserted94.4% (67/71)
gpt-5.5, supported of asserted82.3% (51/62)

Built for teams shipping AI over financial data

AI agents and copilots
Verify a figure before the answer leaves the pipeline, and attach the verdict, the evidence, and a record id.
Fintech products embedding AI
A control where AI output reaches customers, with a sealed record for every claim you surfaced.
Research and investment platforms
Each supported figure carries its evidence back to the filing; a reader you hand the record to can re-check it offline.
Compliance, model risk, and audit
A record per claim, stamped with the policy version that produced it: what was checked, against what, under which rules.

Don’t ask users to trust your AI. Give them the record.

The API is live and onboarding design partners. Bring a claim your model made and leave with a sealed record you can verify offline.

Request API accessQuickstart: 5 steps, curl

Design-partner program, keys issued by the founders.