{
  "published_before_scoring": "2026-11-09",
  "purpose": "Rules fixed before any product was scored. They define the cohorts and their applicability masks, what each dimension means, what a score of 0 to 5 requires, when a cell is n/e or n/a, which cells count, how availability, evidence grade and verification status are recorded, and how the Verdict Quotient summarises them without a league table.",
  "selection_rules": {
    "universe": "A reasoned market sample assembled by the vendor-universe sweep (workstream U) from the brief's seeds, the leads of external review round 1 and vendors found in workstreams A to J. Not a census. The platforms' own controls, researchers and auditors, and standards bodies are described in their own tables and never scored. Contextual products without a verification or blocking product belong to the contextual paper; curated deals belong to the curation paper.",
    "scored_set": "A named product enters the matrix when it is sold to buyers for one of the three cohort tasks, is live or beta at the freeze, and has at least inclusion_min_scored counted numeric cells; n/a, n/e and unverified cells do not count. data/product-register.json lists every scored, listed and historical product with its inclusion reason, and every exclusion carries dated, bounded wording (what was opened, when). A company can have rows in several cohorts; evidence is not inherited between rows, and no product is split, aliased or padded to reach a band threshold.",
    "deep_profiles": "Six to eight companies, chosen after workstream U by market weight and depth of public evidence. The seed list is DoubleVerify, Integral Ad Science, HUMAN, Pixalate, Zefr, Channel Factory and Adloox (Peer39), plus one platform programme described, not scored. Each profile states why it was chosen."
  },
  "scale": {
    "0": "the record documents that the capability is absent or has failed the criterion",
    "1": "adjacent or incidental",
    "2": "partial or indirect, for example via a partner",
    "3": "real but thinly evidenced, narrow in scope, or beta",
    "4": "documented and specific but from vendor documentation only, or independently observed but narrow",
    "5": "the capability is documented in operational detail from an official document, filing or independent observation"
  },
  "cell_rules": {
    "scale": "0 to 5 integer, evidence-bounded, not absolute; every value is defined in \"scale\" above, and every dimension carries its own anchors for at least 5, 3, 1 and 0.",
    "zero": "0 is used only when a source documents absence (a vendor statement that the feature is not offered, a retired product, or an independent finding of failure). Missing evidence is never scored 0. A 0 recorded on a dated adverse record stays when the source later goes offline (the archived copy is kept); it changes only when a later record supersedes it. Allegations, letters and complaints never produce a 0: only a finding (an accreditor's decision, a regulator's finding, a court's finding on the merits, or an independent test meeting the dimension's anchor) can.",
    "ne": "n/e (no evidence) when the opened record gives no basis to score either way. Distinct from 0.",
    "na": "n/a comes only from the cohort's applicability mask (cohorts[].applicable), never vendor by vendor. n/a never enters k, a group mean, the coverage numerator or any imputation.",
    "availability": "Every scored cell carries one label: live (generally available at the cutoff), beta (limited access or beta), announced (not yet observable), retired, unconfirmed (claimed but no observable evidence).",
    "evidence_grade": "A = filing, official transaction document, regulator or standards-body record; B = independent measurement or documentation by a party without financial interest; C = vendor technical documentation or help centre; D = vendor marketing, press release or case study; E = none.",
    "date": "The date of the most recent source the cell rests on (YYYY-MM or YYYY-MM-DD).",
    "verification_status": "Every scored cell carries one: verified (a verifier re-checked the cell's sources), qualified (re-checked; scored against the qualified claim) or unverified. Only verified and qualified cells count; an unverified cell is shown but treated as n/e for admission, k, the quotient and band eligibility.",
    "product": "The scoring unit is a named product that serves the cohort's buying task, one row per product, each mapped to one record in data/product-register.json. Evidence is not inherited between rows, and an alias or bundle never creates a second row.",
    "same_standard": "The same evidentiary standard applies to every vendor. Documentation volume is not product quality: a large help centre earns grade C, not a higher score.",
    "grade_caps": "Vendor documentation (grade C) caps a score at 4 and vendor marketing (grade D) at 3; a grade can cap a score but never raise it. scripts/check.mjs enforces the caps in grade_caps.",
    "maturity": "Live and beta products may be scored under the caps; announced and retired products are listed in the product register, not scored (check.mjs fails a scored announced or retired cell).",
    "first_party": "A platform's or seller's own measurement (first-party validation, for example Google's Active View or Amazon's traffic validation) is recorded in the platform table and in hypothesis H2; it never scores a third-party product's cell."
  },
  "grade_caps": {
    "C": 4,
    "D": 3
  },
  "inclusion_min_scored": 4,
  "dimensions": [
    {
      "key": "method_disclosure",
      "label": "Method disclosure",
      "short": "Method",
      "definition": "How each metric the product sells is determined (signals, models, thresholds, human review, sampling, update cadence), as published. An accreditation letter's scope paragraph is not method disclosure: accreditation scores only in dimension 10. A private audit's findings earn nothing here unless published.",
      "anchors": {
        "5": "method, thresholds and update cadence published for every metric the product sells, in operational detail, in an official document or an independent technical analysis of the method",
        "4": "method and thresholds published for every metric the product sells, in the vendor's own technical documentation",
        "3": "method summarised for at least half of the metrics the product sells, or published in detail for fewer than half",
        "2": "method described only in a platform's or partner's documentation of the integration",
        "1": "marketing description only: signals or models named, with no thresholds or cadence",
        "0": "an accreditor's record, a regulator's finding or a court's finding on the merits documents that the stated method was not applied"
      }
    },
    {
      "key": "detection_evidence",
      "label": "Independent detection evidence",
      "short": "Detection",
      "definition": "Published tests of the product's error rates for a category it sells (invalid traffic, a safety or suitability category, made-for-advertising where the product sold that detection during the test's observation period): miss rate, false-positive rate or precision. A test counts only with a labelled test set, a denominator, the product's settings and the observation period. Case reports that show a failure mode without a denominator, studies of agreement between vendors and prevalence studies of a sample do not score here; they are recorded in the evidence register and the chapters. A response that corrects a test's scope, data, decoding or denominator counts against the test without a new trial of equal size. The cell records the test's design and result direction in its reason.",
      "anchors": {
        "5": "two or more independent tests (by parties without a financial interest), out of sample, with labelled test sets and denominators, find the product within its stated error for the category",
        "4": "one such independent test finds the product within its stated error",
        "3": "a vendor or commissioned test with a published method, a labelled holdout and a denominator",
        "2": "a vendor or commissioned test with summary results only, no denominator or holdout published",
        "1": "accuracy or catch-rate claims only",
        "0": "an independent test with a labelled test set and a denominator finds the product missed 25% or more of a category it sold at the time, under the documented settings, and no response corrects the test's scope or data"
      }
    },
    {
      "key": "ivt_method",
      "label": "Invalid-traffic method",
      "short": "IVT",
      "definition": "What the product's invalid-traffic detection covers in each environment where it sells IVT measurement or protection: general and sophisticated invalid traffic (GIVT and SIVT), pre-bid and post-bid, and the method class (tag, log, server-side, platform feed). Accreditation is not counted here (dimension 10) and test results are not counted here (dimension 2).",
      "anchors": {
        "5": "GIVT and SIVT detection, pre-bid and post-bid, documented for every environment where the product sells IVT measurement, with the method class, in an official document or an independent technical analysis",
        "4": "the same coverage documented in the vendor's own technical documentation",
        "3": "SIVT detection documented with the method summarised, or documented for some but not all environments sold",
        "2": "SIVT detection available only through a partner's documented integration",
        "1": "GIVT only, or IVT detection asserted with no GIVT or SIVT scope stated",
        "0": "the product is sold as IVT measurement or protection in an environment where its own documentation or an accreditor's record shows no IVT filtration"
      }
    },
    {
      "key": "suitability_method",
      "label": "Safety and suitability method",
      "short": "Suitability",
      "definition": "The framework and dated edition the product implements (the archived June 2022 GARM edition, a successor or a custom taxonomy), whether classification is content-level (the page, video, episode or post) or placement-level (domain, app or channel), and the languages and media types covered. Scored against the media types the product sells: a single-medium specialist is not marked down for media it does not sell, and breadth is recorded as an attribute. 'GARM-aligned' after August 2024 means aligned to the archived June 2022 edition; who maintains a framework is recorded separately (hypothesis H6).",
      "anchors": {
        "5": "framework and dated edition named, category definitions published, content-level classification documented for every media type and language the product sells, with the update cadence",
        "4": "framework and edition named and content-level classification documented for every media type sold, with category definitions or languages unpublished",
        "3": "framework named, but classification placement-level for some media types sold, or category definitions unpublished",
        "2": "classification available only through a partner's documented integration",
        "1": "'brand safe' or 'suitable' asserted, with no framework or classification level stated",
        "0": "the product is sold as content-level classification while its own documentation or an accreditor's record shows domain-level or app-level classification only"
      }
    },
    {
      "key": "verdict_authority",
      "label": "Verdict uses and authority",
      "short": "Authority",
      "definition": "How far the product's verdicts on its main sold metric travel toward settlement, as documented in named platform terms, exchange terms or public contracts: report only, pre-bid exclusion, post-bid blocking, a term giving the verdict determining effect on billable quantity, a fee, a credit or a remedy, and a documented settlement. A term that names a class (for example MRC-accredited third-party measurement services) counts for a product inside that class for that metric. The ladder measures how far a verdict travels toward money, which is the paper's question; it is not a measure of value to a buyer. A clause lifts the cell only if it covers the product's main sold metric.",
      "anchors": {
        "5": "a public term gives the product's verdict determining effect on billable quantity, a fee, a credit or a remedy, and a settlement paid on its verdict is documented",
        "4": "such a term is documented on a named platform or in a public contract, with no settlement documented",
        "3": "pre-bid exclusion and post-bid blocking both documented on named platforms",
        "2": "pre-bid exclusion or post-bid blocking documented",
        "1": "report only",
        "0": "the product's own terms state that its measurements may not be used to determine billing, credits or remedies"
      }
    },
    {
      "key": "environment_coverage",
      "label": "Environment coverage",
      "short": "Environments",
      "definition": "Environments in which the product's main sold metrics are live: desktop web, mobile web, mobile in-app, CTV, audio and DOOH. Each counts once. Formats inside an environment are attributes, and the large platforms are counted only in platform access. Live means available as a product (generally or in beta) at the freeze, not a one-off study.",
      "anchors": {
        "5": "five or six environments live",
        "4": "four environments live",
        "3": "three environments live",
        "2": "two environments live",
        "1": "one environment live",
        "0": "a source documents that the product has no live environment (it is then listed, not scored)"
      }
    },
    {
      "key": "platform_access",
      "label": "Platform access",
      "short": "Platforms",
      "definition": "The nine large platforms (Meta, YouTube, TikTok, Amazon, X, Snap, Pinterest, Reddit, LinkedIn) on which the product is a documented third-party partner for a metric it sells, with the scope recorded per the platform table: content-level or placement-level, event-level or aggregate, pre-bid or post-bid. A partner's name alone is n/e, and undocumented access is n/e, not 1.",
      "anchors": {
        "5": "five or more platforms documented, at least two with content-level or event-level scope",
        "4": "five or more platforms with placement-level or aggregate scope only, or three or four with at least one content-level or event-level",
        "3": "three or four platforms documented",
        "2": "two platforms documented",
        "1": "one platform documented",
        "0": "the vendor or the platforms document that the product is not supported on any of the nine"
      }
    },
    {
      "key": "measured_scale",
      "label": "Measured scale",
      "short": "Scale",
      "definition": "Impressions or media transactions measured by the product in a stated 12-month period, as defined in a filing or an official statement, with the unit and its definition recorded. Bid requests, pages crawled or 'interactions' are different units: recorded as attributes, not scored. A figure without a unit or a period is n/e. A shorter period is scaled linearly to a year and marked derived; a figure dated before 2024 is capped at 2. Thresholds are inclusive.",
      "anchors": {
        "5": "1 trillion or more measured a year",
        "4": "100 billion or more a year",
        "3": "10 billion or more a year",
        "2": "1 billion or more a year",
        "1": "a stated figure below 1 billion a year",
        "0": "a source documents that the product is not in production"
      }
    },
    {
      "key": "us_availability",
      "label": "US availability and use",
      "short": "US",
      "definition": "Documented US availability and US production use only. International breadth is an attribute and earns no points.",
      "anchors": {
        "5": "generally available in the US, with US production use documented by an official record or independent observation (a filing, an accreditor's record of the US service, a platform's partner listing for US campaigns)",
        "4": "generally available in the US, with US production use documented in vendor materials (named US clients or US campaigns)",
        "3": "available in the US, with no US production use documented, or US availability in beta",
        "2": "available in the US only through a partner or reseller, with the arrangement documented",
        "1": "US mentioned incidentally, with no documented US availability",
        "0": "a source documents that the product is not available in the US"
      }
    },
    {
      "key": "accreditation",
      "label": "Accreditation",
      "short": "Accredited",
      "definition": "MRC accreditation of the units the product sells, read from MRC's accredited list and letters at the freeze. A unit is product x metric family x environment as recorded in the product register; the metric families are display viewability, video viewability, GIVT, SIVT, property-level ad verification and content-level brand safety or suitability, and a placement or language exclusion in a letter narrows the unit. Score = accredited units / sold units. An audit in progress counts only when MRC lists it, and only as the anchors state; a vendor's announcement is a vendor assertion. Withdrawal from a unit no longer sold is recorded, not scored; a unit still sold after a withdrawal or a denial is unaccredited. TAG certification is entity-level and recorded as an attribute. Attention metrics belong to the attention paper and are not units here.",
      "anchors": {
        "5": "90% or more of sold units accredited",
        "4": "60% to under 90% of sold units accredited",
        "3": "30% to under 60% of sold units accredited",
        "2": "more than 0% and under 30% of sold units accredited",
        "1": "no sold unit accredited, with an MRC audit of at least one sold unit listed as in progress",
        "0": "no sold unit accredited and none under review on MRC's lists"
      }
    },
    {
      "key": "interest_disclosure",
      "label": "Interest disclosure and safeguards",
      "short": "Interests",
      "definition": "Disclosure of revenue by customer type, of sell-side products and of platform-paid programmes, and published safeguards (a conflict or separation policy) for any owner or customer with an interest in the inventory or the currency the product measures. Ownership itself is recorded descriptively and earns or loses nothing; an interested owner matters here only through the safeguards published for it.",
      "anchors": {
        "5": "revenue by customer type in a filing for the last fiscal year, sell-side and platform-paid products disclosed, and a published conflict or separation policy covering every such interest",
        "4": "revenue by customer type in a filing and sell-side products disclosed, with no published conflict or separation policy",
        "3": "sell-side products or platform-paid programmes disclosed by the vendor, with no revenue shares",
        "2": "sell-side relationships disclosed only by partners or platforms",
        "1": "no disclosure of revenue mix, sell-side products or platform-paid programmes found in the vendor's pages and filings screened",
        "0": "a filing, a regulator's finding or a court's finding on the merits documents an undisclosed conflict"
      }
    },
    {
      "key": "dispute_process",
      "label": "Dispute and correction process",
      "short": "Disputes",
      "definition": "What a buyer or a seller can do when it disputes a verdict: a published dispute or appeal process with time limits, a published restatement or correction policy, and documented corrections or dated method changes. The legal and incident register (letters, complaints, rulings, independent reports) sits outside the VQ: each item is described with its posture and never scored, because a complaint is not a judgment and an unexamined product has no record to show.",
      "anchors": {
        "5": "a published dispute or appeal process with time limits, a published restatement or correction policy, and at least one documented correction or dated method change in the 36 months before the freeze",
        "4": "two of those three documented",
        "3": "one of the three documented in the vendor's own terms or help pages",
        "2": "a dispute route documented only by a platform or a partner",
        "1": "an account-manager or contact route only",
        "0": "the vendor's terms state that its measurements are final and not subject to dispute or correction"
      }
    }
  ],
  "peer_groups": [
    {
      "key": "verification",
      "label": "Verification: third-party verification companies",
      "definition": "Companies whose primary business is third-party ad verification (for example DoubleVerify, Integral Ad Science, Adloox, Protected Media)."
    },
    {
      "key": "fraud",
      "label": "Fraud: invalid-traffic and bot-security companies",
      "definition": "Companies whose primary business is invalid-traffic detection, fraud prevention or bot security, scored on the product that measures or filters ad impressions (for example HUMAN, Pixalate)."
    },
    {
      "key": "contextual",
      "label": "Contextual: content classification companies",
      "definition": "Companies whose primary business is content classification, scored only on a verification or blocking product (for example Zefr, Channel Factory, Peer39)."
    },
    {
      "key": "measurement",
      "label": "Measurement: audience and media measurement companies",
      "definition": "Companies whose primary business is audience or media measurement, with verification as one product line (for example Comscore)."
    }
  ],
  "cohorts": [
    {
      "key": "suite",
      "label": "A · Full verification suites",
      "definition": "Products sold to buyers that measure viewability, invalid traffic (GIVT and SIVT) and at least one of brand safety or suitability, in at least two environments. Attention is the attention paper's and is not scored here.",
      "applicable": [
        "method_disclosure",
        "detection_evidence",
        "ivt_method",
        "suitability_method",
        "verdict_authority",
        "environment_coverage",
        "platform_access",
        "measured_scale",
        "us_availability",
        "accreditation",
        "interest_disclosure",
        "dispute_process"
      ],
      "banded": true
    },
    {
      "key": "ivt",
      "label": "B · Invalid-traffic and fraud specialists",
      "definition": "Products whose primary output is invalid-traffic detection or fraud prevention that measures or filters ad impressions; click, install, session and form fraud tools enter only where impression measurement is documented. The IVT products of cohort A are scored in A.",
      "applicable": [
        "method_disclosure",
        "detection_evidence",
        "ivt_method",
        "verdict_authority",
        "environment_coverage",
        "measured_scale",
        "us_availability",
        "accreditation",
        "interest_disclosure",
        "dispute_process"
      ],
      "banded": true
    },
    {
      "key": "suitability",
      "label": "C · Content safety and suitability specialists",
      "definition": "Products whose primary output is content-level safety or suitability classification sold for verification or blocking. When the same signal is sold both as an avoidance service and as a curated deal, the avoidance service is scored here once per buying task and the curated deal belongs to the curation paper, with the shared origin recorded.",
      "applicable": [
        "method_disclosure",
        "detection_evidence",
        "suitability_method",
        "verdict_authority",
        "environment_coverage",
        "platform_access",
        "measured_scale",
        "us_availability",
        "accreditation",
        "interest_disclosure",
        "dispute_process"
      ],
      "banded": true
    }
  ],
  "cohorts_note": "Each cohort lists its applicable dimensions (the mask). A cohort that carries bands (banded: true) needs at least six applicable dimensions and at least one in each group. A dimension may carry anchors_by_cohort { <cohort>: {5, 3, 1, 0} } where the anchors differ by cohort; every applicable dimension in every cohort needs anchors for at least 5, 3, 1 and 0.",
  "row_order": [],
  "matrix_title": "The capability matrix",
  "matrix_note": "Each cell scores one documented capability from 0 to 5 against fixed anchors; n/e means no evidence, n/a means the dimension does not apply to the cohort. Click a cell for its rationale, availability, evidence grade, date, sources and verification status. Rows are compared only within their cohort, and a score is documented capability, not accuracy.",
  "matrix_foot": "",
  "composite_index": {
    "used": true,
    "name": "Verdict Quotient (VQ)",
    "added": "Defined on 9 November 2026, after external review round 1 of the brief and before any product was scored: the dimensions, anchors and masks were fixed before any cell was opened.",
    "groups": {
      "Depth": {
        "label": "Depth",
        "meaning": "What the verdict is: how its method is disclosed, whether independent tests support it, what it detects and classifies, and how far it travels toward settlement.",
        "dims": [
          "method_disclosure",
          "detection_evidence",
          "ivt_method",
          "suitability_method",
          "verdict_authority"
        ]
      },
      "Reach": {
        "label": "Reach",
        "meaning": "Where it works: environments, large platforms, measured volume and US use.",
        "dims": [
          "environment_coverage",
          "platform_access",
          "measured_scale",
          "us_availability"
        ]
      },
      "Standing": {
        "label": "Standing",
        "meaning": "Whether a buyer can trust and contest it: accreditation, interest disclosure and safeguards, and the dispute and correction process.",
        "dims": [
          "accreditation",
          "interest_disclosure",
          "dispute_process"
        ]
      }
    },
    "weights": {
      "Depth": 0.45,
      "Reach": 0.25,
      "Standing": 0.3
    },
    "field_mean_method": "raw_mean",
    "prior_strength": 4,
    "shrink_alternatives": [
      0,
      2,
      8
    ],
    "missing_scenario_values": [
      1,
      4
    ],
    "missing_envelope_values": [
      0,
      5
    ],
    "bands": [
      [
        "Most documented",
        12
      ],
      [
        "Well documented",
        1
      ],
      [
        "Partly documented",
        -8
      ],
      [
        "Least documented",
        -1000
      ]
    ],
    "thin_record_min": 6,
    "cohort_min_banded": 6,
    "method": "Within each cohort, over its admitted rows (thin rows included; listed and historical products never count): each group score is the mean of the row's counted numeric cells in the group (n/e, n/a and unverified cells skipped, never zero). Raw = 0.45 Depth + 0.25 Reach + 0.30 Standing, renormalised over the groups present. F = the arithmetic mean of the cohort's Raw scores (field_mean_method raw_mean). Shrunk = (k x Raw + K x F) / (k + K), k = the row's counted numeric cells, K = prior strength 4. Quotient = 100 x Shrunk / 5 at full precision; only the displayed value is rounded, halves up. An empty cohort has no quotient.",
    "reweightings": [
      {
        "name": "Equal thirds",
        "weights": {
          "Depth": 0.3333333333333333,
          "Reach": 0.3333333333333333,
          "Standing": 0.3333333333333333
        }
      },
      {
        "name": "Depth-led",
        "weights": {
          "Depth": 0.6,
          "Reach": 0.2,
          "Standing": 0.2
        }
      },
      {
        "name": "Reach-led",
        "weights": {
          "Depth": 0.3,
          "Reach": 0.45,
          "Standing": 0.25
        }
      },
      {
        "name": "Standing-led",
        "weights": {
          "Depth": 0.3,
          "Reach": 0.2,
          "Standing": 0.5
        }
      }
    ],
    "why_not_a_sum": "A sum counts an unscored dimension as zero and so penalises a thin public record rather than a weak product; a plain average of scored cells rewards missing evidence. Group means plus shrinkage avoid both.",
    "limits": "The index summarises publicly documented capability within a cohort for one buying task. It is not a measure of accuracy, commercial independence or legal safety, and it inherits the single-scorer limitation. Bands are reported because exact positions depend on the weights; a band that a single-cell deletion or a missing-evidence scenario changes is labelled band-sensitive, and the paper draws no buyer recommendation from an ordering those tests change.",
    "bands_rule": "Bands come from the unrounded offset 20 x (Shrunk - F): Most documented at +12 or more, Well documented at +1 or more, Partly documented at -8 or more, Least documented below; equal values share a band. These are the shared rules' offsets (Front rank, Strong, Mid, Narrow), renamed as in the attention paper to say what they measure: how much capability is documented in public. The thresholds did not change.",
    "thin_rule": "A row is band-eligible with k >= thin_record_min and at least one counted cell in every group that applies under the cohort's mask; a row missing a whole applicable group is shown with that group named and no band. Count the eligible rows first: a cohort with fewer than cohort_min_banded of them shows quotients and coverage without bands.",
    "shrink_rationale": "Prior strength K = 4 is an author-defined heuristic, not a statistical estimator: the dimensions are different capabilities, not repeated measurements of one quantity. The build reruns the index with K = 0 (Raw), 2 and 8 and reports how far positions and bands move.",
    "missing_rule": "Missing evidence is not neutral. Deletion: each counted cell turned into n/e in turn, recomputing the row's k and eligibility; quotient change, admission loss and eligibility loss are reported separately. Scenarios (not bounds): row by row, with F and every other row frozen, all of the row's applicable n/e cells filled with missing_scenario_values[0], then [1]; a banded row whose band changes under either is flagged band sensitive to missing evidence. Fills of missing_envelope_values are reported separately as a hypothetical full-scale envelope, never used for the flag. Imputed cells never create admission or band eligibility: the row keeps its observed k and eligibility. Publication consequence (round 1): a band that changes when any single counted cell is deleted is reported as band-sensitive beside the band in the text, like a band sensitive to missing evidence, and no ordering that either test changes is presented as a buyer recommendation.",
    "reweighting_rule": "Each reweighting recomputes Raw and F across the fixed cohort."
  },
  "scorer": {
    "independent_second_scorer": false,
    "disclosure": "One scoring pass by the author's AI-assisted research process, applied after the verifier records were read; facts were adversarially verified, scores were not independently re-scored. No human second scorer and no inter-rater statistic exist."
  }
}
