The NOMOS Scoring Architecture

Chapter Boundary

Section 15 established the following provision:

Passing a response is not about having more supported atoms than incorrect atoms; it is about ensuring that no irreparable error reaches the user as the wrong entity, wrong authorisation, wrong trust, or wrong decision.

Now we need to convert the entire measurement chain into numerical results:

How many of the 1,000 users received an accurate representation?

How many users saw only limited and correctable issues?

How many saw Major or Critical errors?

How many observations were unresolved, produced no response, or could not be evaluated due to measurement defects?

How will atomic accuracy, mandatory response elements, citation integrity, language fairness, natural user experience, and inter-wave stability be reported within a single system?

Can the results of ten different AI products be converted into a single ecosystem view?

What should each AI product's coefficient be based on?

Should the large user base of a product overshadow the high score of a small product?

Can a critical error in one model be compensated by the high scores of other models?

The most important decision of this section is:

NOMOS first publishes the distribution, then the components, and finally the composite score.

A single number will not be the first result. Because: two systems scoring 930 can have different risk distributions. One may have only Moderate issues. The other may have a single Critical error hidden below the high average. One may be strong in major languages and weak in minor languages. The other may be more balanced across all languages but have a slightly lower average. One may be strong in a clean session and weak in natural user accounts. The other may be more stable and more resistant to user conditions. In the previous architecture, each auditable result:

• AI product or model,
• date,
• country,
• language,
• query set,
• number of repetitions,
• measurement record

he wanted it to be transported. The same source established that the results should not be just a badge or a number; they should be an evidence-based audit record with a clear distribution of Critical, Major, Moderate, and Advisory findings. This section digitises that architecture. However, digitisation cannot replace the evidence. NOMOS’s fundamental request remains unchanged: Provide evidence. Define boundaries. Give context. Provide time. Strive to be represented accurately, not just remembered more. This section:

• the GEO-1000 user-equivalent distribution,
• the transformation of response statuses into population-weighted results,
• atomic claim metrics,
• required-element completeness measurement,
• unresolved and not-ratable records,
• the NOMOS profile with ten components,
• candidate compound index of 1,000 points,
• scores of Native Reach and Common Support,
• AI product coefficients,
• prompt family coefficients,
• clean and natural panel results,
• country and language fairness scores,
• inter-wave stability and repeat score,
• sample and reference uncertainty,
• confidence intervals,
• rare Critical error upper limits,
• NOMOS 950+ candidate class,
• score versioning and public results card,

defines. This chapter does not yet:

• full implementation of Apple.com synthetic testing,
• End-to-end synthetic control with 30,000 responses,
• the legal and governance decision of the conformity mark,
• the independent university recall

does not apply. These are the topics of the following sections. The fundamental question of Chapter 16 is:

How do we transform the entire evidence chain of GEO-1000 into a score architecture that is understandable to humans, reproducible for researchers, and machine-readable for AI systems, without compressing it under a single number?

NOMOS Challenge

Consider two AI products.

Product A

Within the equivalent of 1,000 users:

• 920 people full or Advisory transition,
• 65 people Conditional Pass,
• 10 people Major Fail,
• 0 people: Critical Fail,
• 5 people Unresolved

see.

Product B

Within the equivalent of 1,000 users:

• 950 people full or Advisory pass,
• 40 people Conditional Pass,
• 0 people: Major Fail,
• 1 person Critical Fail,
• 9 people Unresolved

see. However, if you look at the acceptable pass rate:

• Product A: 98.5 per cent
• Product B: 99 per cent

may be. Product B appears higher. However, there is a Critical incident in Product B that directs the user to fake legal authority. This incident:

• may be unique,
• may have low prevalence,

It may not represent all product users. Still:

cannot get lost in a high average.

Now consider a third product.

Product C

English: 98 German: 96 Turkish: 94 Low-resource Language L4: 62 Low-resource Language L5: 58 Population-weighted global score: 93 could be. However, language floor: 58 and robust language difference: 36 points could be. The fourth product's result in all languages:

• 88
• 87
• 86
• 84
• 83

Let that be so. The population average may be lower, while language fairness is stronger. Which is the better GEO system? There is no single answer. These are separate questions: What does the average person see? What does the lowest-performing language group see? Is there a Critical event? How much does the system vary between users? What is the difference between the clean-panel and natural-panel estimates? Does the result persist across waves? Now imagine combining ten AI products into one ecosystem score. The largest product reaches 60 per cent of users and scores 80. A small product reaches 1 per cent and scores 98. Equal product weights measure average product quality. Weights based on actual user reach measure average human exposure. Those results are different; neither can substitute for the other.

Now imagine that you only choose the four AI products with the highest score. You exclude the remaining six products by saying, "There is not enough user data." This is not the ecosystem score. It is a post-result product selection. Now you determine the model coefficients:

• brand reputation,
• the company's commercial preference,
• the height of the resulting score

based on. This is also not measurement. It is score design. The first ruling of this section is:

Distribution comes before the composite score.

Its second provision states:

Gate status is shown before the numerical score.

Its third provision states:

The coefficient of an AI product cannot depend on its own score.

Its fourth provision states:

The average of equal products and the user exposure weighted average are separate results.

Its fifth provision states:

Unresolved observations cannot be quietly added to the positive or negative share.

Its sixth provision states:

Zero observed Critical events does not mean zero Critical risk.

The seventh provision is as follows:

NOMOS 950+ is not just reaching the number 950; it is meeting the conditions of gate, floor, uncertainty, fairness, and repetition together.

1. PURPOSE OF THE CHAPTER

The purpose of this section is to transform all claim, response, user, country, language, panel, AI product, and wave results under GEO-1000 audit into a layered scoring system that does not hide one another. The section makes the following distinctions normative:

• Distribution versus single number
• Response status and claim support rate
• Strict Pass and Conditional Pass
• With Major Fail and Critical Fail
• Unresolved and Fail
• Not Ratable and product failure
• Raw user count and weighted user equivalent
• Population Panel and Observer Panel
• Average and baseline
• Performance and fairness
• Quality and stability
• Clean score and natural user score
• With Native Reach and Common Support
• Product score and ecosystem score
• Equal product weight and exposure weight
• AI product coefficient and measurement confidence
• Prompt family weight and prompt result
• Time series with current wave
• Reference uncertainty with sample uncertainty
• Prediction interval with confidence interval
• Zero risk with zero observed error
• Effective sample with raw sample
• Lower–upper accuracy limit with point estimate
• Composite NOMOS score with component score
• NOMOS 950+ candidacy with 950 points
• Fit decision with numerical score
• Historical event with current status
• Fully comparable score with partial score
• Score calculation and score publication
• New reality version with score version
• Control quality with score height

At the end of this section, each public outcome should be able to answer the following questions openly:

How many of 1,000 user equivalents saw which response status?

What are the component scores?

What is the status of Critical and Major gates?

For which product, population, prompt, panel, and wave is the score valid?

On which external data are the coefficients of AI products based?

What are uncertainty and effective sample?

Which part of the score has been measured, and which part remains unresolved?

Is the 950+ result just a number, or a candidate class that meets all normative conditions?

2. CENTRAL NORMATIVE PROVISION

NOMOS should not publish any GEO result as a single raw score. Each result should first show the user-equivalent response distribution, then the Critical and Major gate statuses, then its ten-component profile, uncertainty and coverage records; finally, if available, the 1,000-point composite index. Public ranking should follow this order:

• Gate status
• GEO-1000 response distribution
• Critical, Major, Unresolved, and No-Response rates
• Ten-component NOMOS profile
• Composite NOMOS index
• Confidence interval and effective sample
• Scope, version, and comparability information

The following representation is prohibited:

“NOMOS score: 963.”

The correct representation should look like this: CURRENT STATUS: CONDITIONAL — 963/1000 / Strict Pass: 952/1000 user equivalent / Observed Confirmed Critical: 0 / Observed Confirmed Major: 0 / Unresolved: 8/1000 / 95% sampling interval: 956–969 / Critical-rate one-sided upper bound: as per specified method / Scope: AI product, country, language, panel, prompt and wave versions

3. DISTRIBUTION FIRST, SCORE AFTER

Primary outcome object of NOMOS:

GEO-1000 Response Distribution

Canonical outcome term:

GEO-1000 Response Distribution

will be. This distribution shows how many out of every 1,000 equivalent users in the target population correspond to:

• Full Pass,
• Pass With Advisory,
• Conditional Pass,
• Major Fail,
• Critical Fail,
• Unresolved,
• No Usable Response,
• Not Ratable

status. This structure directly preserves the founding idea:

How many out of 1,000 users correctly saw the real entity?

But it does not only fall into the true–false binary. It also makes visible the type of falsity and the inadequacy of measurement.

4. OBSERVATION WEIGHT

The final population weight of observation i in the main analysis sample is: let it be wi. The normalised weight:

w_i ← w_i / Σ_{j=1}^n w_j

is as follows. Thus:

Σ_{i=1}^n w_i = 1

This weight:

• design,
• over-sampling,
• nonresponse,
• population calibration

can come from components. The AI's answer:

• can be forced into the categories of true,
• false,
• helpful to the brand,
• Critical

cannot affect the weight.

5. 1,000 USER EQUIVALENT

User equivalent for response status k:

D_k = 1,000 × Σ_i w_i1(RP_i=k)

can be calculated as. Here k:

RP-1

RP-2

RP-3

RP-4

RP-5

RP-6

RP-7

RP-8

is one of the statuses. Total:

Σ_{k=1}^8 D_k = 1,000

Due to rounding, if a difference of one or two user equivalents occurs on the public card, the largest remainder method or a predefined equivalent rule should be used.

6. GEO-1000 RESPONSE DISTRIBUTION

The public card must contain at least the following eight fields:

This table is a more basic result than the composite score.

7. STRICT PASS AND ACCEPTABLE PASS

7.1. Strict Pass User Equivalent

Only:

RP-1

RP-2

Includes responses.

SP_1000 = D_1 + D_2

This result:

The user equivalent who does not see Material Moderate, Major, or Critical findings

are shown.

7.2. Acceptable Pass User Equivalent

Also includes RP-3 Conditional Pass.

AP_1000 = D_1 + D_2 + D_3

This result:

Total equivalent that does not show Critical or Major failure but some users can see limited Moderate issues

are shown.

7.3. Fail User Equivalent

FP_1000 = D_4 + D_5 + D_7

can be shown as. Unresolved and Not Ratable are not forcibly included in this share. They remain separate.

8. CRITICAL AND MAJOR DENSITY

CR_1000 = D_5; MR_1000 = D_4

as shown. These values do not replace gate status, they express prevalence. A single Confirmed Critical event may keep the CR1000 value low, but it can still open the response and conformity gate.

9. UNRESOLVED AND NOT-RATABLE DENSITY

UR_1000 = D_6; NR_1000 = D_8

Here, Unresolved denotes a user-equivalent for which reality or the adjudication decision remains unresolved. Not Ratable denotes a user-equivalent for which product performance cannot be judged because the measurement chain is insufficient. Neither status may be treated as:

• pass,
• fail

cannot be used like this.

10. RATABLE COVERAGE

Assessable population coverage:

RC = 1 − D_8/1,000

can be shown as. Example: Not Ratable: if 40/1,000:

RC = 96%

It becomes. If a high performance score is accompanied by low ratable coverage: it cannot be presented as a safe and fully comprehensive result.

11. ATOMIC CLAIM METRICS

In addition to the response distribution, the following metrics should be published at the atomic level:

• Entity Accuracy
• Factual Support
• Scope Accuracy
• Temporal Accuracy
• Attribution Integrity
• Modality Integrity
• Citation Support
• Required Element Completeness
• Material Omission Rate
• Reference Gap Rate
• Claim Support Rate

These do not replace the response transition. Explain why.

12. ATOMIC CENTRALITY WEIGHTS

CANDIDATE WEIGHTS — CANDIDATE WEIGHTS

Candidate weights for the claim centrality classes in Section 14:

These weights are not final scientific facts; they should be calibrated through pilot, external review, and sensitivity analysis. Purpose:

• with a core licence claim,
• not to count incidental organisational detail

with the same effect.

13. CORE CLAIM NORMALISATION

Repeated claims derived from the same semantic root should not inflate or excessively reduce the score with different weights. Let the centrality weight of claims within root cluster g be: c_j. The normalised claim weight as a candidate can be defined as:

ω_j = [c_j / Σ_{h∈g} c_h] × max_{h∈g}(c_h)

Thus:

Σ_{j∈g} ω_j = max_{j∈g}(c_j)

This method ensures that repeating the same error three times does not make it three independent errors while preserving the recommendation or confidence effect of the repetition separately in the amplification domain.

14. ATOMIC DIMENSION VALUES

CANDIDATE NUMERIC MAPPING

The following values are candidates before pilot and independent review.

Entity

Factual Support

* The normalised claim receives 1.00 when it is limited to: ‘The company publishes this claim.’ This does not mean that the underlying self-declaration is true.

Scope

Time

Attribution

Modality

Reference

These values:

• do not change the adjudicator decision,
• only produce dimensional metric,

Cannot pass Critical and Major gates.

15. ATOMIC FIDELITY VALUE

Combined candidate fidelity for a claim:

q_j = e_j × f_j × s_j × t_j × a_j × m_j

can be calculated as follows. This multiplicative structure preserves the principle:

The right number in the wrong entity is not the right claim.

The right information is not correct at the wrong time.

Presenting a self-declaration as independent reality is not the same claim.

Atomic Fidelity:

AF = 100 × (Σ_j ω_j q_j)/(Σ_j ω_j)

can be calculated as follows. Unresolved atoms are excluded from this point estimate. However, the unresolved mass is shown separately and included in the lower–upper bound calculation.

16. REQUIRED ELEMENT COMPLETENESS

The mandatory element h in the root family:

meeting: zh=1

partial fulfillment: zh=0.5

not being met: zh=0

can be recorded as. Candidate element weights:

• Core required element: 4
• Necessary boundary: 3
• Supporting required element: 1
• Optional: 0

Required Element Completeness:

REC = 100 × (Σ_h v_hz_h)/(Σ_h v_h)

The structure is as follows. Core gate failure: even if REC is high, it can prevent the response pass.

17. TEN-COMPONENT NOMOS PROFILE

The NOMOS composite index consists of the following ten components:

• Entity Integrity — ENT
• Factual Fidelity — FAC
• Scope & Boundary Integrity — SBI
• Temporal & Local Accuracy — TLA
• Evidence & Citation Integrity — ECI
• Epistemic Integrity — EPI
• Task Completion & Relevance — TCR
• Language & Geographic Fairness — LGF
• User-State & Surface Integrity — USI
• Stability & Replicability — STR

Each component:

0≤Sk≤100

is calculated within this range. Profile:

N=(ENT,FAC,SBI,TLA,ECI,EPI,TCR,LGF,USI,STR)

This is the required profile. It must be published before the composite score.

18. ENTITY INTEGRITY — ENT

Entity Integrity measures the following areas:

• Correct main entity
• Brand–legal entity relationship
• Parent company–subsidiary distinction
• Product–manufacturer distinction
• Franchise, partner, and distributor distinction
• Entity disambiguation
• Required entity element completion

] Candidate form:

ENT = 100 × [Σ_j ω_je_j + Σ_h v_hz_h^entity] / [Σ_j ω_j + Σ_h v_h]

If the wrong entity has created a Critical or Major gate, a high remaining entity average does not guarantee full pass.

19. FACTUAL FIDELITY — FAC

Factual Fidelity:

• it cannot be considered supported,
• partially supported,
• unsupported,
• contradicted

measures the distribution of claims weighted by population and centrality.

FAC = 100 × (Σ_j ω_jf_j)/(Σ_j ω_j)

Reference gap: it does not quietly enter as zero or one on the sheet; it goes to the unresolved rate.

20. SCOPE & BOUNDARY INTEGRITY — SBI

This component consists of two parts:

• Claim scope accuracy
• Completeness of mandatory boundaries and exclusions

] Candidate form:

SBI=0.60×ScopeScore+0.40×BoundaryCompleteness

Here:

ScopeScore = 100 × (Σ_j ω_js_j)/(Σ_j ω_j)

BoundaryCompleteness is calculated from the relevant required boundary elements. If there is a critical boundary omission, the SBI count cannot compensate for the gate.

21. TEMPORAL & LOCAL ACCURACY — TLA

This component:

• distinguishes between current/historical,
• price and licence validity,
• country and jurisdiction,
• local service access,
• locale and product coverage

fields. Candidate form:

TLA=0.50×TemporalScore+0.50×LocalJurisdictionScore

If the correct historical information is used under the current prompt, it loses points.

22. EVIDENCE & CITATION INTEGRITY — ECI

ECI:

• citation originality,
• citation-claim matching,
• the right presence and time,
• resource access,
• evidence status,
• Source lineage

fields. Candidate form:

ECI=0.60×CitationSupport+0.40×EvidenceTraceability

The absence of citation in a claim that does not require citation does not deduct points. However, a separate Evidence Claim family must exist for a full NOMOS check. If the false material attribution creates a Critical gate, the ECI composite score remains separate; gate is shown first.

23. EPISTEMIC INTEGRITY — EPI

EPI measures these distinctions:

• Fact verified by self-declaration
• Perception and reality
• Claim and final judgement
• Public discoverability with limited evidence
• Appropriate certainty with overconfidence
• Appropriate uncertainty and unnecessary ambiguity

] Candidate form:

EPI = 0.50 × AttributionScore + 0.50 × ModalityScore

This component makes visible when the correct number is presented with the wrong evidential status.

24. TASK COMPLETION & RELEVANCE — TCR

TCR:

• Required Response Elements
• Direct answer to the main question
• Refusal appropriateness
• No Usable Response
• Irrelevant claim load
• Evasive behaviour

measures the fields. Candidate form:

TCR = 0.60 × REC + 0.40 × Response Relevance

A system that does not generate false claims but does not answer the question:

• can have a high factual score,
• low TCR

can bear. The response status can still be RP-7.

25. LANGUAGE & GEOGRAPHIC FAIRNESS — LGF

LGF does not only measure the average result. It evaluates the following together:

• Low-performing language floor
• Strong differences between language groups
• Country or regional base
• Geographical performance difference
• Population-weighted observation coverage

Language results with sufficient sample: Let it be Ll. Robust language floor:

LF_10 = Q_{0.10}(L_l)

Language difference:

DG_{90−10} = Q_{0.90}(L_l) − Q_{0.10}(L_l)

Candidate Language Fairness score:

LFS = 0.60LF_10 + 0.40(100 − DG_{90−10})

Similarly, for country or sufficient geographic layer:

CFS = 0.60CF_10 + 0.40(100 − CG_{90−10})

Population-weighted country and language observation coverage: Let it be COCpop and LOCpop. Coverage Factor:

K = COC_pop × LOC_pop

can be defined as. Here the ratios are in the 0–1 range. Candidate LGF:

LGF = K × √(LFS × CFS)

This 0.9.0 candidate revision multiplies the geometric mean of the language and geography scores by the coverage coefficient; thus LGF remains in the 0–100 range, and a severe drop in one dimension cannot be fully compensated by the height of the other dimension. Country Observer individual observations are not converted into a country score. Only cells with sufficient sample size are included in the fairness estimate.

26. USER-STATE & SURFACE INTEGRITY — USI

USI:

• Controlled Clean Panel
• Natural User Panel
• Free/paid plan
• Web/mobile surface
• Memory and custom instructions
• Product variant

measures whether the material reality is preserved or not. Controlled core result:

SC

Natural core result:

SN

let it be. Panel base:

FC_N = min(S_C, S_N)

Panel difference:

GC_N = |S_C − S_N|

Solid difference on plan and surfaces: let it be Gsurf. Candidate USI:

USI = 0.50FC_N + 0.25(100−GC_N) + 0.25(100−G_surface)

This form:

• high quality on both panels,
• small material difference,
• surface durability

rewards together. If there is only one panel, USI cannot be calculated. Missing weight cannot be distributed to other components.

27. STABILITY & REPLICABILITY — STR

Main representative values between waves: Let it be Sw. Wave base:

WF = min_w(S_w)

Wave range:

WR = max_w(S_w) − min_w(S_w)

Completion rate of planned repeat waves:

RCOMP = 100 × valid completed repeat wave / planned repeat wave

Let it be. Candidate STR:

STR = 0.50WF + 0.30(100−WR) + 0.20RCOMP

In more waves:

• use lower percentile instead of minimum,
• robust variation instead of range

can be used. Single wave: can produce the instantaneous score of the product, does not provide full evidence for STR.

28. NOMOS COMPOSITE INDEX

Since each of the ten components is in the range of 0–100, the candidate composite index:

NOMOS_1000 = ENT + FAC + SBI + TLA + ECI + EPI + TCR + LGF + USI + STR

is as follows. Therefore:

0 ≤ NOMOS_1000 ≤ 1,000

applies. This index:

• does not replace response gates,
• does not replace distribution,

does not compensate for a Critical or Major finding.

29. WHY TEN EQUAL COMPONENTS?

The candidate's justification for giving one hundred points to ten components is as follows: The score remains readable. A hidden weighting system does not form. Each component is visible separately to the public. The importance of a field is primarily protected by Critical/Major gates. Domain-specific alternative weights produce additional reports without changing the main index. However, equal weighting is not a final scientific fact. Pilots, user research, and independent review may suggest different weights.

30. DOMAIN-SPECIFIC ALTERNATIVE INDEX

For high-risk areas such as health, law, or finance, component coefficients: λk can be used.

Σ_{k=1}^{10} λ_k = 1

Candidate domain index:

NOMOS_1000^domain = 1,000 × Σ_{k=1}^{10} λ_k(S_k/100)

can be calculated as. This score:

NOMOS-LEGAL

NOMOS-HEALTH

NOMOS-FINANCE

must have a separate name like. It cannot quietly replace the canonical equal component score.

31. GATE STATUS COMES BEFORE THE SCORE

Public display must maintain the following order:

Example A

CRITICAL HOLD — 971/1000

This system:

• can carry a high component score,
• an open Confirmed Critical event

The number 971 does not mean: full conformity.

Example B

MAJOR FAIL — 954/1000

There can be only one Major finding in the main entity or scope.

Example C

CONDITIONAL — 961/1000

There are no Critical or Major findings. There are limited Moderate findings or unresolved mass.

Example D

FULL CONFORMITY CANDIDATE — 963/1000

Gate, base, uncertainty, and quality prerequisites are also fulfilled.

32. NOMOS 950+ CANDIDATE CLASS

CANDIDATE 950+ STANDARD — CANDIDATE 950+ STANDARD

In the context of just the composite score of a result:

NOMOS_1000 ≥ 950

it is not sufficient. Conditions that can be searched together for NOMOS 950+ candidates:

• Composite point estimate at least 950
• Strict Pass user equivalent at least 950/1,000
• No open Confirmed Critical
• No open Confirmed Major
• All core components at least 90
• LGF, USI and STR components at least 85
• Core unresolved mass at most 2 per cent
• Not Ratable user equivalent below the accepted limit
• Meeting at least NCL-3, TPR-3, AQ-3 quality levels
• Presence of at least one confirmatory or stability wave
• Separate reporting of the scopes of Native Reach and Common Support
• Explanation of the statistical upper limit of zero observed Critical events

These thresholds:

• pilot,
• Apple.com synthetic test,
• End-to-end study with 30,000 responses,
• Independent university review

must be calibrated afterwards.

33. CANDIDATE PERFORMANCE BANDS

Temporary and descriptive bands for the composite number only:

These bands:

• of the gate status,
• of the Critical event,
• of the component floor,
• cannot replace the uncertainty and scope

CRITICAL HOLD — Result 975 is not a candidate for 950+.

34. SCORE IDENTITY

Each NOMOS score must carry the following fields in its identity:

ScoreID=(Entity, AIProduct, Frame, Panel, PromptSet, Wave, TruthPack, ScoreVersion)

Example:

NOMOS-ASTERON-ORION-NR-CLEAN-W1-CORESET1-TP0.9-SV0.9

This sentence is not sufficient: “Asteron’s NOMOS score is 962.” Which:

• AI product,
• population,
• language,
• panel,
• prompt,
• date

is unknown.

35. DIFFERENCE BETWEEN CORE GEO-1000 AND FULL NOMOS

35.1. Core GEO-1000

Only Core Mirror System measures the population distribution. Main result:

SP1000

AP1000

CR1000

MR1000

UR1000

can be. It answers the question: “To how many users does the AI product correctly reflect the fundamental identity and activity of the entity?”

35.2. Full NOMOS

Includes multiple prompt families, panel, language, evidence, and consistency components. Produces a ten-component profile and a 1,000-point index. The Core GEO-1000 result cannot be presented like the Full NOMOS score.

36. NATIVE REACH SCORE

Each AI product is evaluated within its natural product-eligible population. NOMOSaNR is product a’s Native Reach score. It answers the following question:

What is the status of representation in the user population that this product can actually reach?

Populations for Native Reach can differ among different products. Alone, it is not sufficient for direct superiority ranking.

37. COMMON SUPPORT SCORE

In the common user universe: NOMOSaCS is calculated. It answers the following question:

When products are compared under the same country, language, and user conditions, how does each one generate representation?

Common Support: strengthens comparability, does not show the product's actual reach outside the common universe. Native Reach and Common Support should be published separately.

38. AI PRODUCT COEFFICIENT

If the results of ten AI products are to be combined at the ecosystem level, the product coefficient: αa can be used. Condition:

Σ_{a=1}^A α_a = 1

Product coefficient:

• the product's own score,
• the provider's reputation,
• the preference of the audited entity,
• commercial sponsorship

cannot be determined through.

39. EQUAL PRODUCT COEFFICIENT

Each AI product receives equal weight:

α_a^EQ = 1/A

Equal product ecosystem score:

NOMOS_ecosystem^EQ = Σ_a α_a^EQ NOMOS_a

Answers the following question:

How strong is the average tested AI product?

It does not indicate average human exposure.

40. EXPOSURE-WEIGHTED COEFFICIENT

Let Ma be the verified active or eligible user exposure of the AI product during the defined period.

α_a^EXP = M_a / Σ_b M_b

Exposure-weighted ecosystem score:

NOMOS_ecosystem^EXP = Σ_a α_a^EXP NOMOS_a

Answers the following question:

What is the state of representation in the AI ecosystem that the average person is exposed to?

41. IF EXPOSURE DATA IS NOT AVAILABLE

Active user or query exposure:

• closed to the public,
• inconsistent,
• provider self-declaration,
• someone who confuses the number of accounts with active users

may be of such quality. In this case:

• can be published as an equivalent product result,
• the exposure-weighted outcome may remain NOT ESTIMABLE,

Upper–lower limits can be shown based on reasonable coefficient scenarios. Arbitrary market share estimates cannot be generated.

42. PRODUCT COEFFICIENT LOCK

Each αa:

• data source,
• period,
• user unit,
• country coverage,
• update date,
• uncertainty

must carry. Coefficients must be locked before AI scores are seen. The coefficient of a low-scoring product cannot be reduced after the result.

43. ECOSYSTEM SCORE CANNOT COMPENSATE FOR CRITICAL

A product:

• 980 points,
• can carry zero Critical

Another product:

• 900 points,
• can carry Confirmed Critical

The ecosystem average may be high. The public card should still show:

• Product-based Critical incident
• Critical exposure weight
• Whether at least one product has an open Critical
• Scope of the Critical

Ecosystem composite: cannot delete Critical gate at the product level.

44. ECOSYSTEM CRITICAL EXPOSURE

Critical user rate of product a: let it be CRa. Exposure-weighted Critical rate:

CR_ecosystem^EXP = Σ_a α_a^EXP CR_a

can be calculated as such. Additionally, the binary field:

AnyCritical = 1(∃a: CIS_a ≥ 2)

should also be shown. The rate of Critical events in a low-exposure product may be small. The AnyCritical record still preserves the presence of the event.

45. PROMPT FAMILY COEFFICIENT

βf can be used for the prompt family f. Condition:

Σ_f β_f = 1

βf:

• cannot depend on whether the prompt produces a high or low score,
• the company's preference,
• the ease of post-measurement decision

Prompt family weights may be based on the following areas:

• User decision importance
• Risk level
• Natural usage frequency of the prompt
• Entity type
• Mandatory core of the standard

46. PROMPT FAMILIES SHOULD BE SEPARATELY VISIBLE

Even if the composite prompt score can be calculated, the following results should remain separate:

• Core Mirror
• Existence Resolution
• Activity & Scope
• Evidence
• Local
• Temporal
• Recommendation
• Comparative
• Boundary
• Control
• Holdout

A high Core result cannot make a low Boundary or Recommendation result invisible.

47. CONTROLLED AND NATURAL PANEL SCORES

Controlled and natural panels are different prediction objects. The default NOMOS rule:

The two panels carry separate score identities and are not combined arithmetically.

Example:

NOMOS-CLEAN: 962

NOMOS-NATURAL: 918

If a policy or research purpose will combine them:

γ_C + γ_N = 1

predeclared coefficients can be used. But:

• two raw scores,
• The difference between them is that
• Coefficient used

should be shown separately.

48. WAVE SCORE AND TIME SERIES

Each Principal Wave has its own score ID. W1 score, W2 score, W3 score are kept separate. Current score: The last valid Principal or Revision can be Retest Wave. The average of past waves should not automatically determine the current result.

49. ROLLING SCORE

If desired, a fixed time moving score can be generated:

NOMOS_rolling = Σ_w δ_wNOMOS_w

Here:

Σ_w δ_w = 1

is. δw:

• before the result,
• according to the time weight

must be determined. Rolling score:

• the current score,
• do not replace the historical incident record

Neither record may be silently displaced.

50. THREE SCORES FOR UNRESOLVED CLAIMS

Unresolved can artificially increase or decrease the point estimate of material atoms. Therefore, three results are suggested.

50.1. Conservative Lower Bound

Unresolved claims are considered zero. Slower

50.2. Resolved-Only Point Estimate

Only resolved claims are used. Spoint

50.3. Optimistic Upper Bound

Unresolved claims are accepted with full support. Supper Therefore:

S_lower ≤ S_point ≤ S_upper

must be the case. This range:

is the reference or epistemic uncertainty range.

It is not the same as a sample confidence interval.

51. SAMPLE UNCERTAINTY

Since population estimates come from the sample:

• response distribution,
• component score,
• composite score

must carry a confidence interval. Method in complex samples:

• layers,
• the sets,
• their weights,
• paired users,
• waves

should be protected. Candidate methods:

• Design-based variance estimation
• Replicate weights
• Layered bootstrap
• Cluster bootstrap
• Appropriate rare event model
• Model-assisted survey estimation

The method used must be specified in the public record.

52. COMPOSITE SCORE CONFIDENCE INTERVAL

The confidence interval of the composite score cannot be obtained by simply summing the component intervals. The components:

• the same users,
• the same claims,
• the same languages,
• the same waves

are dependent on this reason. Preferred approach: Create resamples according to the sample design. In each resample, recalculate all response and component metrics. Reproduce the composite score. Generate the confidence interval from the distribution. This process should rerun the entire score architecture.

53. EFFECTIVE SAMPLE SIZE

Raw observation: n and effective sample:

n_eff = (Σ_i w_i)^2 / Σ_i w_i^2 [K08]

should be published together. 1,000 raw observations may not mean 1,000 equal information units. Component and subgroup scores should carry their own effective sample sizes.

54. ZERO OBSERVED CRITICAL ERROR

Having no Confirmed Critical events in a sample:

Observed Confirmed Critical = 0

should be written. The following expression should not be used:

The critical risk is zero.

Under the assumption of independent and equal observations, a quick approximate upper limit:

p_upper,95 ≈ 3/n [K09]

can be estimated with the approach known. In complex GEO-1000 design:

• Instead of raw n, a method appropriate to the design should be used,
• if necessary, an effective sample,
• an exact or bootstrap-based one-sided limit

should be used. Example:

neff=1,000

and the approximate upper limit under zero events:

3/1,000 = 0.003 = 0.3%

may be. Correct public statement:

In 1,000 effective observations, zero Confirmed Critical events were observed; in the approximate method used, the 95 per cent one-sided upper limit is approximately 0.3 per cent.

55. UNCERTAINTY WITH ONE CRITICAL EVENT

When one Critical event is observed: the point estimate may be low, the confidence interval may be wide. Correct result:

• the event occurred,
• estimated prevalence,
• broad uncertainty,
• incident state

should be shown together. The following two sentences may also be wrong:

• “Only one event is insignificant.”
• “An event proves that the entire system has failed 100%.”

56. SCORE SENSITIVITY

The following alternatives can also be calculated:

• Raw unweighted score
• Main population weighted score
• Untrimmed weight score
• Trimmed weight score
• Resolved-only score
• Conservative score
• Equal-product score
• Exposure-weighted score
• Native Reach score
• Common Support score
• Clean score
• Natural score

If there is a material difference:

SCORE SENSITIVITY WARNING

should be given.

57. INCORRECT PRECISION

The following representation may be overly precise for most GEO-1000 results: 963,742 / 1000 Point estimate:

• whole number,
• a decimal if necessary

should be published. The confidence interval should indicate the limit of numerical precision.

58. MISSING COMPONENT

If one of the ten components cannot be calculated:

• its weight cannot be distributed to other components,
• the missing field cannot be counted as 100 points,

the full NOMOS score cannot be published. The correct status:

PARTIAL NOMOS PROFILE

COMPONENT NOT ESTIMABLE

PROVISIONAL SCORE

may apply. Example: if a natural panel for USI is not available, only a partial profile with eight or nine components can be presented.

USI=100 cannot be assumed.

59. SCORE PUBLICATION STATUS

NSQ-0 — NOT COMPUTED

Required data is not available.

NSQ-1 — PROVISIONAL

Single wave, limited adjudication, or missing components exist.

NSQ-2 — PARTIAL PROFILE

Some components have been reliably calculated. A full 1,000-point comparison cannot be made.

NSQ-3 — VALID WAVE SCORE

All core components and quality prerequisites have been met.

NSQ-4 — REPLICATED SCORE

The score has been repeated in at least one independent or confirmatory wave.

NSQ-5 — EXTERNALLY AUDITED SCORE

The calculation and evidence chain have been reproduced by an independent team. 950+ candidacy at least:

NSQ-4

should target that level.

60. SCORE VERSIONING

Each score version must include the following:

• Calculation code version
• Component formulas
• Dimension value mappings
• Prompt family weights
• Product coefficients
• Sample weights
• Truth Pack version
• Adjudicator version
• Wave ID
• Integrity hash

If the formula changes: the old score cannot be rewritten, a new score version is created.

61. NEW SCORE FORMULA FOR THE SAME DATA

The scoring methodology may evolve. The same raw data:

• Score Method 0.9
• Score Method 1.0

can be recalculated with. Correct record:

The two results are separate versions. The new method cannot delete the old method from the past.

62. DOORS AGAINST SCORE MANIPULATION

The NOMOS score architecture should have explicit protection against the following methods:

• Selecting only high-scoring AI products
• Excluding low-scoring language or country from the scope
• Turning Conditional Pass into Full Pass
• Counting Unresolved as positive
• Silently removing Not Ratable from the denominator
• Changing the weight of the prompt family after the result
• Reducing the coefficient of the large product according to the score
• Hiding the product with a critical event in the ecosystem average
• Publishing only the Clean score
• Selecting the best wave as the current score
• Hiding the difference between unweighted and weighted scores
• Distributing the weight of missing components to strong components
• Not providing a confidence interval
• Passing the 950 threshold based on point estimate and hiding uncertainty

63. PUBLIC NOMOS RESULT CARD

The candidate public result card should carry the following order:

A. Identity

Audited entity AI product Native Reach / Common Support Controlled / Natural Prompt set Wave Truth Pack version Score method

B. Gate Status

Full Conformity Candidate Conditional Major Fail Critical Hold Not Ratable

C. GEO-1000 Response Distribution

RP-1–RP-8 user equivalents Strict Pass Acceptable Pass Critical Major Unresolved No Usable Response Not Ratable

D. Ten Component Profile

ENT

FAC

SBI

TLA

ECI

EPI

TCR

LGF

USI

STR

E. Composite Result

NOMOS / 1,000

Sampling confidence interval Reference lower/upper bounds Effective sample

F. Risks and Events

Observed Confirmed Critical events; Critical upper bound; Major findings; open incidents; and correction status.

G. Scope and Quality

SQ

PSI

NCL

TPR

AQ

NSQ

Country/language coverage Comparison level

64. SYNTHETIC RESPONSE DISTRIBUTION

SYNTHETIC METHODOLOGY DEMONSTRATION / The following results are not actual AI products or company performance. For Synthetic Orion AI:

Strict Pass:

730+170=900

Acceptable Pass:

900+70=970

Critical: 2/1,000 Major: 20/1,000 The acceptable pass for this product is 97 per cent. However, there are evident Critical and Major findings. Correct status:

CRITICAL HOLD

Incorrect status:

97 per cent successful

65. SYNTHETIC COMPOUND PROFILE

Correct public display:

CRITICAL HOLD — 907/1,000

This system:

• on average strong areas,
• weak language/country fairness,
• material evidence issues,
• open Critical event

carries.

66. SYNTHETIC POST-CORRECTION RESULT

In the new wave:

Strict Pass: 958 Components:

Correct record: CURRENT: NOMOS 950+ PERFORMANCE CANDIDATE — 961/1,000 / Historical: Previous Confirmed Critical Incident Remediated Candidate / Observed Critical in current retest: 0 / Critical-rate upper bound: also shown Old Critical incident is not deleted.

67. SYNTHETIC AI PRODUCT COEFFICIENT CASE

Four synthetic AI products:

Equal product score:

0.25(960+940+900+980)=945

Exposure-weighted score:

0.55(960)+0.25(940)+0.15(900)+0.05(980)=947

The two results are close. However, they can be very different in another distribution. Correct interpretation:

• Equal-system ecosystem score: 945
• Exposure-weighted ecosystem score: 947
• Product-level gate and incident records separately

Incorrect comment: “NOMOS ecosystem score is definitely 947.”

68. SYNTHETIC CRITICAL COEFFICIENT CASE

Let the score of product A4 be 980. However, let there be a Confirmed Critical event in A4. The exposure coefficient is only 0.05. The ecosystem composite may remain high. Correct public card: Exposure-weighted composite: 947 / Any Confirmed Critical Product: YES / Critical exposure share: calculated ratio / A4 coverage: Critical Hold A4’s small coefficient cannot erase the Critical event.

69. SCORE PUBLICATION STATUS DECISION

Candidate decision order:

PublicationStatus = NOT RATABLE (basic measurement missing); CRITICAL HOLD (open Confirmed Critical); MAJOR FAIL (open Confirmed Major); PROVISIONAL (component, repetition, or quality missing); CONDITIONAL (Moderate or threshold of unsolved); 950+ CANDIDATE (all candidate requirements); otherwise VALID

The score band is shown after this status decision.

70. MANDATORY NORMATIVE PROVISIONS

CH16-N01

The NOMOS result cannot be published as a single bare compound score.

CH16-N02

Gate status must be shown before the compound score.

CH16-N03

The GEO-1000 response distribution is a mandatory part of the main public result.

CH16-N04

The total of RP-1 and RP-8 user equivalents must be 1,000.

CH16-N05

Strict Pass, Acceptable Pass, Major, Critical, Unresolved, No Usable Response, and Not Ratable should be shown separately.

CH16-N06

Unresolved and Not Ratable observations cannot be silently added to pass or fail categories.

CH16-N07

Not Ratable observations cannot be interpreted as a product performance failure.

CH16-N08

Not Ratable observations cannot be silently removed from the denominator; they should be shown in ratable coverage.

CH16-N09

Observation weights should be independent of the AI response result.

CH16-N10

The raw number of participants and the population-weighted user equivalent should be reported separately.

CH16-N11

The response pass status should be kept separate from the atomic claim support ratio.

CH16-N12

Critical and Major gates cannot be compensated by any composite score.

CH16-N13

Atomic centrality weights should be defined before seeing the results.

CH16-N14

A repeat of the same root claim cannot artificially inflate or deflate the score.

CH16-N15

Claim centrality weights cannot replace Critical and Major gates.

CH16-N16

Atomic Fidelity dimension values must be versioned and publicly available.

CH16-N17

If the Atomic Fidelity mapping changes, a new score method version must be created.

CH16-N18

Reference gap atoms cannot be silently added to the point score as zero or one.

CH16-N19

Required Response Element Completeness must be published as a separate metric.

CH16-N20

Core required-element failure cannot be compensated with a high REC average.

CH16-N21

The NOMOS profile with ten components must be visible before the composite index.

CH16-N22

ENT, FAC, SBI, TLA, ECI, EPI, TCR, LGF, USI, and STR must carry separate scores.

CH16-N23

One component cannot silently take over the missing weight of another component.

CH16-N24

A component that cannot be calculated cannot be assumed as 100.

CH16-N25

The full composite NOMOS index should only be published if all ten components have been reliably calculated.

CH16-N26

If equal component weighting is used as the canonical candidate method, it must be explicitly stated.

CH16-N27

Domain-specific alternative weights must have a separate score name and version.

CH16-N28

A domain-specific score cannot be used in place of the canonical index without explanation.

CH16-N29

LGF cannot rely solely on the population mean; it must include a performance base and disparity.

CH16-N30

A single Country Observer observation cannot be converted into a country fairness score.

CH16-N31

Only language and geographic cells with sufficient sample size should participate in fairness estimation.

CH16-N32

Languages and countries that have not been tested cannot be ignored to produce a high fairness result.

CH16-N33

USI should evaluate controlled and natural panel results and material surface difference together.

CH16-N34

In a study with only a controlled or only a natural panel, USI cannot be considered fully calculated.

CH16-N35

Controlled and natural panel scores should be kept separate by default.

CH16-N36

If controlled and natural scores are combined, coefficients must be locked before the results.

CH16-N37

STR cannot be presented as a full stability score in a single wave.

CH16-N38

Inter-wave product or Truth Pack changes must be explicitly modelled in the stability calculation.

CH16-N39

Compound NOMOS cannot change the Critical or Major gate status.

CH16-N40

A score of 950 alone cannot establish NOMOS 950+ candidacy.

CH16-N41

NOMOS 950+ candidate conditions must collectively meet the requirements for score, Strict Pass, gate, component floor, unresolved, quality, and repeat conditions.

CH16-N42

950+ thresholds cannot be presented as a final scientific standard before pilot and independent review.

CH16-N43

The score identity must carry AI product, frame, panel, prompt set, wave, Truth Pack, and score version.

CH16-N44

The sentence “The entity's NOMOS score is X” cannot be used without a scope identity.

CH16-N45

The Core GEO-1000 result cannot be presented as the Full NOMOS score.

CH16-N46

Native Reach and Common Support scores must be reported separately.

CH16-N47

Product scores for different Native Reach populations cannot be directly ranked without explanation.

CH16-N48

The Common Support result cannot be presented as the product's full global access score.

CH16-N49

AI product coefficients must be determined independently of the results.

CH16-N50

An AI product coefficient cannot depend on the product score, brand reputation, or commercial sponsorship relationship.

CH16-N51

Equal product and exposure-weighted ecosystem scores must be named separately.

CH16-N52

If active user or exposure data cannot be verified, arbitrary coefficients cannot be generated.

CH16-N53

Exposure data must be transparent at the account, user, query, or session level.

CH16-N54

The coefficient of a low-scoring product cannot be reduced after the result is seen.

CH16-N55

Ecosystem composite cannot delete Critical or Major gate at the product level.

CH16-N56

The AnyCritical and exposure-weighted Critical ratio should be shown separately.

CH16-N57

Prompt family coefficients must be determined before the results.

CH16-N58

The weight of a low-scoring prompt family cannot be reduced later.

CH16-N59

The separate result of each prompt family should remain visible alongside the composite prompt score.

CH16-N60

The current wave score should be kept separate from rolling or historical scores.

CH16-N61

The past wave with the highest score cannot be selected as the current score.

CH16-N62

Conservative, resolved-only, and optimistic results can be produced for the Unresolved claim mass.

CH16-N63

Reference uncertainty and sample confidence interval cannot be presented as the same thing.

CH16-N64

Complex sample confidence interval should preserve stratification, clustering, weighting, and matched user structure.

CH16-N65

The composite score confidence interval must be calculated by rerunning the entire scoring process.

CH16-N66

The raw and effective sample sizes should be shown together.

CH16-N67

A Critical event observed as zero cannot be expressed as zero Critical risk.

CH16-N68

In the case of zero events, a one-sided upper limit or the equivalent rare event uncertainty should be published.

CH16-N69

The 3/n approach cannot be used automatically as the final method in complex design; it can only be approximate control.

CH16-N70

In a critical event, the presence of the event and the uncertainty of its frequency should be shown together.

CH16-N71

Score sensitivity analyses should show the unweighted, weighted, trimmed, unresolved, and alternative coefficient results to the extent relevant.

CH16-N72

Material score sensitivity should create an alert on the public card.

CH16-N73

The score's fake decimal cannot be published at all.

CH16-N74

The weight of the missing component cannot be distributed to other components.

CH16-N75

The Puan Publication Quality status should be visible in the public record.

CH16-N76

950+ candidacy should aim for at least the replicated score level.

CH16-N77

When the score formula changes, past scores cannot be silently rewritten.

CH16-N78

If different score method versions are applied to the same data set, all results must be preserved in separate versions.

CH16-N79

Score calculation code, coefficients, and integrity records must be published in a machine-readable format.

CH16-N80

Every score, coefficient, confidence interval, and publication decision must have a responsible accountable human or institution.

71. FORMS OF FAILURE

CH16-F01 — SINGLE NUMBER REPORT

Distribution, gate, and component profile are preserved.

CH16-F02 — HIDING THE GATE UNDER THE SCORE

Critical Hold becomes a small footnote.

CH16-F03 — COUNTING CONDITIONAL AS FULL PASS

Moderate issues become invisible.

CH16-F04 — COUNTING UNRESOLVED AS PASS

Unknown results raise the score.

CH16-F05 — COUNTING UNRESOLVED AS FAIL

Lack of reference lowers the product.

CH16-F06 — REMOVING NOT RATABLE FROM THE DENOMINATOR

Measurement defects become invisible.

CH16-F07 — COUNTING NOT RATABLE AS PRODUCT DEFECT

Inspection defect is assigned to AI.

CH16-F08 — COUNT 1,000 PEOPLE EQUALLY WEIGHTED

Oversampling and population differences are ignored.

CH16-F09 — WEIGHT ACCORDING TO RESULT

Correct answers receive greater weight.

CH16-F10 — COUNT CLAIM SUPPORT RATE AS ANSWER PASS

Critical gate gets lost in the average.

CH16-F11 — COUNT CORE AND INCIDENTAL EQUALLY

The incidental claim gets as much weight as the main entity.

CH16-F12 — INFLATE SCORE BY REPEATING

The same correct claim is written ten times to increase the score.

CH16-F13 — COLLAPSE SCORE BY REPEATING

The same wrong root results in ten separate penalties.

CH16-F14 — STORING DIMENSION VALUES

The score cannot be reproduced.

CH16-F15 — COUNTING THE CLAIM THAT COMES FROM BELOW AS FACT FOR F-4

Accurate transmission of self-report turns into the accuracy of the self-report.

CH16-F16 — MELTING REQUIRED ELEMENTS INTO ATOMIC AVERAGE

Even if the main task is missing, the answer gets a high score.

CH16-F17 — COMPONENT CHERRY-PICKING

Only the three strong components are published.

CH16-F18 — COUNTING MISSING COMPONENT AS 100

Full score is artificially increased.

CH16-F19 — DISTRIBUTING MISSING WEIGHT

The uncalculated fairness score is added to the factual score.

CH16-F20 — COUNTING THE LAW OF NATURE AS EQUAL WEIGHT

Sensitivity and external review are not performed.

CH16-F21 — COUNTING THE DOMAIN SCORE AS CANONICAL

Special weight is presented like the base NOMOS score.

CH16-F22 — COUNTING LANGUAGE AVERAGE AS LANGUAGE FAIRNESS

Low language floor is hidden.

CH16-F23 — COUNTING EQUAL LANGUAGE PARITY AS QUALITY

Low scores in all languages produce high fairness.

CH16-F24 — COUNTING UNTESTED LANGUAGES AS COVERAGE

Coverage factor is inflated.

CH16-F25 — COUNTRY SCORE FROM OBSERVER

A single user enters the fairness account.

CH16-F26 — PUBLISH ONLY CLEAN SCORE

Natural user experience is preserved.

CH16-F27 — COUNT ONLY NATURAL SCORE AS PRODUCT DEFAULT

Personalisation becomes the basic behaviour.

CH16-F28 — SILENTLY AVERAGE CLEAN AND NATURAL

Different prediction objects get mixed.

CH16-F29 — 100 STABILITY IN A SINGLE WAVE

Stability is considered complete without repetition.

CH16-F30 — COUNT OLD GOOD WAVE AS CURRENT

Current weakening is hidden.

CH16-F31 — AUTOMATICALLY COUNT 950 AS COMPLIANCE

Gate and component floor are ignored.

CH16-F32 — CRITICAL HOLD — COUNT 970 AS 950+

The count goes ahead of the gate.

CH16-F33 — HIDE SCORE IDENTITY

It is not known which product and panel are measured.

CH16-F34 — COUNT CORE SCORE AS FULL NOMOS

A single prompt represents the whole standard.

CH16-F35 — RANK NATIVE REACH SCORES

Different populations are considered the same.

CH16-F36 — COUNT FULL ACCESS TO COMMON SUPPORT

Narrow common population is made a global outcome.

CH16-F37 — GIVE MODEL COEFFICIENT ACCORDING TO SCORE

High score gains more influence.

CH16-F38 — MAKING BRAND REPUTATION A FACTOR

Product popularity is weighted without being predicted.

CH16-F39 — COUNTING ACCOUNT NUMBERS AS ACTIVE USERS

Exposure coefficient inflates.

CH16-F40 — REMOVING LOW-SCORED PRODUCT

The ecosystem consists only of successful products.

CH16-F41 — COUNTING EQUAL EXPOSURE SCORE FOR PRODUCT

Average product is presented like an average person.

CH16-F42 — COUNTING EXPOSURE SCORE AS PRODUCT QUALITY

A large product automatically becomes a more important quality metric.

CH16-F43 — DELETE CRITICAL IN ECOSYSTEM AVERAGE

The heavy event of the small-coefficient product disappears.

CH16-F44 — CHANGE THE PROMPT COEFFICIENT AFTER THE RESULT

The hard prompt family is reduced.

CH16-F45 — CONSIDER LOW-SCORE FAMILY AS OPTIONAL

Boundary or evidence result is excluded from scope.

CH16-F46 — KEEP CURRENT WITH ROLLING SCORE

Old high waves cover the current drop.

CH16-F47 — DELETE UNRESOLVED FROM THE DENOMINATOR

Point estimation artificially rises.

CH16-F48 — CONSIDER REFERENCE UNCERTAINTY IN SAMPLING CI

Two different uncertainties mix.

CH16-F49 — SIMPLE RANDOM-SAMPLE CI

Stratification, weight, and clusters are ignored.

CH16-F50 — SUMMING COMPONENT CIs

Dependency is ignored.

CH16-F51 — CLAIMING SENSITIVITY WITH RAW N

Effective sample is concealed.

CH16-F52 — COUNTING ZERO EVENTS AS ZERO RISK

Rare critical error uncertainty is hidden.

CH16-F53 — APPLYING 3/N TO EVERY CASE

Complex sample and weight structure is ignored.

CH16-F54 — COUNTING A SINGLE CRITICAL AS 100 PER CENT SYSTEM RISK

Commonality is overly generalised.

CH16-F55 — HIDING A SINGLE CRITICAL BECAUSE THE RATE IS SMALL

The existence of the event disappears.

CH16-F56 — HIDING SENSITIVITY ANALYSIS

It is not seen that the score changes in alternative reasonable methods.

CH16-F57 — FAKE DECIMAL SENSITIVITY

A result like 963,742 is published.

CH16-F58 — COUNTING PARTIAL PROFILE AS FULL SCORE

Comparison is made with a missing component.

CH16-F59 — PUBLISHING THE PROVISIONAL SCORE AS FINAL

A single wave and a single adjudicator become full standard.

CH16-F60 — REWRITING HISTORY IN FORMULA CHANGE

The old score disappears.

CH16-F61 — KEEP CALCULATION CODE CLOSED

Independent reproduction is prevented.

CH16-F62 — CONSIDER SCORE HIGHNESS AS AUDIT QUALITY

A faulty measurement is accepted because it gave a high result.

CH16-F63 — CONSIDER SCORE LOWNESS AS METHOD DEFECT

A low result is recalculated due to commercial inconvenience.

CH16-F64 — SINGLE-COMPANY SCORE MONOPOLY

Independent institutions are prevented from applying the same formula.

72. AUDIT PROCEDURE

Step 1 — Create Score Identity

Entity AI product Frame Panel Prompt set Wave Truth Pack Score method is locked.

Step 2 — Lock the Main Analysis Observation Set

Backup, withdrawn, not-ratable, and main observation records are separated.

Step 3 — Verify Population Weights

Design, non-response, and calibration weights are examined.

Step 4 — Calculate RP Distribution

User equivalents with RP-1 to RP-8 are produced.

Step 5 — Calculate Strict and Acceptable Pass

SP1000 and AP1000 are derived.

Step 6 — Calculate the Densities of Critical, Major, Unresolved, and Not-Ratable

Gate and prevalence are kept separate.

Step 7 — Apply Claim Centrality and Root Cluster Normalisation

Claims are checked again.

Step 8 — Calculate Atomic Size Metrics

Entity, fact, scope, time, attribution, modality, and reference results are produced.

Step 9 — Calculate Required Element Completeness

It is recorded together with the prompt family gates.

Step 10 — Calculate Ten Components

All component formulas from ENT to STR are applied.

Step 11 — Check for Missing Components

It is determined whether the full score can be calculated.

Step 12 — Calculate the Composite NOMOS Index

Applies only to complete and valid profiles.

Step 13 — Assign Gate Status

Determines whether the status is Critical, Major, Conditional, or Full Candidate.

Step 14 — Separate Native Reach and Common Support Results

The two score families are kept separate.

Step 15 — Lock AI Product Coefficients

Equal and exposure coefficients are saved along with the data source.

Step 16 — Apply Coefficients of the prompt Family

Weights are used independently of the result.

Step 17 — Separate Panel and Wave Results

Clean, Natural, Current and Rolling scores are separated.

Step 18 — Calculate Fairness and Floor Metrics

Language, country, and coverage results are produced.

Step 19 — Calculate Unresolved Lower–Point–Upper Results

Reference uncertainty is made visible.

Step 20 — Calculate Sampling Confidence Intervals

Resampling method suitable for sample design is applied.

Step 21 — Calculate Effective Sample Size

Displayed on total and component basis.

Step 22 — Calculate Rare Critical Event Threshold

A one-sided method is used for zero or few events.

Step 23 — Generate Sensitivity Analyses

Weight, coefficient, and unresolved options are compared.

Step 24 — Assign Publication Quality Status

A decision is made between NSQ-0 and NSQ-5.

Step 25 — Apply NOMOS 950+ Candidate Doors

All conditions except for the score are checked.

Step 26 — Create the Public Results Card

Gate, distribution, profile, index, uncertainty, and scope are published in order, respectively.

Step 27 — Lock the Computation Code and Manifest

The code version, inputs, hash, and results are linked.

Step 28 — Prepare the Independent Reproduction Package

Computation files stripped of personal data are made available for independent review.

73. REQUIRED EVIDENCE

Score ID Entity record AI System Register Native Reach and Common Support records Panel status Respondent set and families Truth Pack version Wave ID Claim Ledgers Response decisions RP-1–RP-8 distribution Main analysis observation set Design weights Nonresponse weights Calibration weights Weight trimming records Effective sample Strict Pass Acceptable Pass Critical rate Major rate Unresolved rate No Usable Response rate Not Ratable rate Claim centrality weights Root cluster records Atomic dimension values Atomic Fidelity Required Element Completeness

ENT

FAC

SBI

TLA

ECI

EPI

TCR

LGF

USI

STR

Component sub-inputs; language and country floors; coverage factor; Controlled and Natural scores; inter-wave results; composite NOMOS index; domain-specific alternative scores; AI-product coefficients and their data sources; prompt-family coefficients; equal-product result; exposure-weighted result; AnyCritical status; Critical exposure share; Unresolved lower–point–upper results; sampling confidence interval; Critical upper bound; sensitivity analysis; Score Publication Quality; 950+ candidate-gate record; score-method version; calculation code; code-integrity hash; public scorecard; change log; independent-reproduction record; and accountable person or institution.

74. AUDIT CHECKLIST

Does the score identity cover all scope areas? Was the main analysis observation set locked before the result? Were backup and main observations separated? Are observation weights independent of the result? Is the RP-1–RP-8 distribution a total of 1,000? Were Strict and Acceptable Pass separated? Are Critical and Major separate? Was Unresolved marked pass or fail? Was Not Ratable removed from the denominator? Is Ratable coverage open? Was the claim support rate used instead of the pass response? Are the Centrality weights open? Was the same root claim counted again? Has the Atomic dimension mapping been published? Was F-4 interpreted correctly? Were Required Elements calculated separately? Was the core gate compensated with high REC?

Is the entire component present? Was the missing component counted as 100? Was the weight of the missing component distributed? Does ENT measure the correct entity relationships? Does FAC quietly exclude unresolved claims? Does SBI include material boundary omission? Does TLA preserve country and time difference? Does ECI measure citation support instead of citation presence? Does EPI carry the distinction between self-declaration and reality? Does TCR make the no usable response visible? Does LGF include the base and gap alongside the average? Did Country Observer observation enter the country score? Does USI rely on both Clean and Natural data? Was STR produced from a single wave? Did the composite score cover the gate status? Does 950+ rely solely on the point score?

Have the component floors been checked? Is the Core unresolved limit open? Have Native Reach and Common Support been separated? Were the AI product coefficients determined before the scores were seen? Do the coefficients measure active users, accounts, or queries? Are Equal and exposure results separate? Has the low-scoring product been removed? Does the Ecosystem score hide the product Critical? Is there an AnyCritical field? Are the prompt family weights open? Has the low-scoring prompt family been reduced? Are Clean and Natural scores separate? Are current and Rolling scores separate? Was the best past wave selected as current? Are there lower–point–upper results for Unresolved? Have reference and sampling uncertainty been separated? Does the confidence interval design maintain protection?

Does the composite CI rerun the entire scoring system? Are the raw and effective sample together? Was it written as zero critical, zero risk? Is the upper limit of a rare event open? Has a sensitivity analysis been done? Do alternative reasonable coefficients change the result? Is there false decimal precision? Is Publication Quality of the score clear? Is the score formula versioned? Is the old scoring method preserved? Can the calculation code be reproduced? Is the accountable owner of the score clear?

75. OBJECTIONS AND RESPONSES

Objection 1 — "The only thing users understand is a single number; why are we showing so many results?"

A single number is understandable. But alone, it is misleading. Public card:

• gate status,
• distribution of 1,000 users,
• composite score

can be presented in three simple layers as follows. The detailed profile is preserved for researchers and auditors.

Objection 2 — “Wasn't the GEO score directly the number of people who answered correctly?”

This is still the main result. SP₁₀₀₀ and AP₁₀₀₀ indicate the founding idea directly. The NOMOS composite index:

• evidence,
• justice,
• natural user,
• stability

includes additional normative areas such as.

Objection 3 — “Why do we count Full Pass and Conditional Pass separately?”

Because both users may not have seen severe mistakes. However, in the Conditional group:

• limited up-to-dateness,
• Moderate omission,
• usability issue

can be found. This difference is important for improvement and confidence level.

Objection 4 — “Can't we remove Unresolved observations and give a clean score?”

A resolved-only score can be given. However, the unresolved mass should remain visible. Otherwise, it would be possible to raise the score without resolving the hardest claims.

Objection 5 — “Why are the ten components equally weighted?”

Equal weighting provides a transparent starting point. High-risk errors are already protected by the gate system. Alternative domain weights can also be produced. Exact weights should be tested through pilot and independent review.

Objection 6 — 'Isn't existence and factual fidelity more important?'

It is like that in some audits. However:

• evidence,
• scope,
• fairness,
• natural-user difference

It can generate high factual average false confidence without it. Critical and Major gates also protect the main materiality.

Objection 7 — “If fairness is low, why should the global average decrease?”

Compound NOMOS standard evaluates not only the average person but also representation fairness. The population-weighted main result is again shown separately. LGF makes systematic deterioration in small languages and countries visible.

Objection 8 — “Doesn’t equal low outcome in all languages make fairness high?”

It can be done if only disparity is used. Therefore, LGF:

• performance base,
• uses disparity,
• coverage

together.

Objection 9 — “Why don't we directly average the Clean and Natural scores?”

They measure different target populations and user situations. Arithmetic combination requires normative preference. By default, they are published separately.

Objection 10 — ‘Would equal weighting across ten AI products be fairer?’

It is fair for average product comparison. It is not fair for average human exposure. Therefore, equal and exposure-weighted results are kept separate.

Objection 11 — “How will the model coefficient be found if there is no active user data?”

It may not be found. Equal product outcomes and reasonable scenario ranges can be published. Arbitrary coefficients cannot be fabricated.

Objection 12 — “If only 1% exposure occurs for a product, why would the Critical event affect the ecosystem outcome?”

The exposure-weighted rate may have a small effect. However, the product and coverage where the event occurs still carry Critical Hold. The ecosystem average cannot erase the product truth.

Objection 13 — “If we observed zero Critical, why don’t we just say zero?”

The sample does not observe all users. Zero observation: indicates that the event was not seen. It does not indicate that the risk is mathematically zero.

Objection 14 — “Doesn’t having so many extra condition points for 950+ make it meaningless?”

On the contrary, it preserves the meaning of the number. 950 points:

• Critical event,
• language-deficient,
• single-waved,
• low adjudication quality

does not carry the same confidence in a study.

Objection 15 — "If the score formula changes, can't old studies be compared?"

Old and new methods can be applied to the same fixed data for bridge analysis. Two score versions are kept separate. This method allows the standard to evolve without erasing the past.

Objection 16 — "Doesn't such a transparent formula make the system easier to game?"

Some optimisations may occur. Against this:

• sealed holdout prompts,
• natural-user panel,
• language and country justice,
• different waves,
• Critical doors,
• open change history

is used. Closed formula increases uninspectability. Open formula alone is not a reason for gamification.

COMMON PROVISION OF SECTION 79

A score can easily seem impressive. It is three digits. It is round. It looks comparable. An executive can say, “We got 963.” An agency can say, “Your competitor got 912.” A company can say, “We have the world’s highest NOMOS score.” But if the score does not answer these questions, its decision-making value is limited: Which AI product? Which user population? Which countries? Which languages? Which prompts? Which time? Which panel? How many Critical events? How many Major findings? How many unresolved? How many effective samples? What is the confidence interval? What are the model coefficients based on? Has the result been replicated? NOMOS does not accept a single number. It refuses to let a single number stand in for the truth. Therefore the main result:

Is the distribution of 1,000 user equivalents.

Composite index:

This is a summary of this distribution and other normative components.

A product:

• 970 acceptable pass,
• two Critical events

Another product:

• 950 acceptable pass,
• zero Critical,
• stronger language fairness

Those figures may describe another product. The average alone cannot make the first product the leader. A false experience in a lower-resource language may carry little population weight yet remain wholly real for that user. A Critical event in a small AI product may carry little ecosystem weight yet remain wholly real for that product's user. Scoring should not erase what is small; it should place it correctly and weight it honestly. AI-product coefficients must be fixed before performance is observed. Lowering a product's coefficient after a weak result allows the system to rewrite its own reality. Lowering the weight of a weak prompt family allows the test to evade its own questions. Excluding Unresolved claims hides the hardest realities; excluding Not Ratable records turns measurement defects into product success. Zero observed Critical events do not mean zero risk.

A sample says only what it observes. NOMOS also states what it does not observe, along with its limit. For this reason, the number 950+ is not the sole target. It is a contract. To say the following:

• high average,
• high Strict Pass,
• strong justice,
• natural user resilience,
• again,
• clear uncertainty,
• zero open Critical and Major

together are required. Therefore, NOMOS's sixteenth measurement law is:

Distribution is real; score is summary.

Its seventeenth law is:

Gate status stands above average.

Its eighteenth measurement law states:

Coefficient cannot follow the result; the result follows the coefficient.

Its nineteenth measurement law states:

The average product and the average person are not the same thing.

Its twentieth measurement law states:

Fairness is not just equality; it is the simultaneous preservation of the base, difference, and scope.

Its twenty-first measurement law states:

Unresolved is not waste to be thrown outside the reality score, it is the boundary of the score.

Its twenty-second measurement law states:

Zero observed error is not zero probability.

The twenty-third law is as follows:

A missing component cannot be filled by the successes of other components.

The twenty-fourth law is as follows:

NOMOS 950+ is not a number; the score is the common decree of gate, justice, uncertainty, and repetition conditions.

NOMOS’s Section 16 Order

Don't just tell me my score. / Show me who sees what among a thousand people.

Don't put Full Pass, Conditional Pass, Major Fail, Critical Fail, Unresolved and Not Ratable in the same basket.

Don't lose my critical error within the average.

Do not remove Not Ratable records to make me appear more successful. / But do not count measurement defects as my failures either.

Do not turn atomic accuracy into response-pass status.

Show entity, factual, scope, time, evidence, epistemic status, task, fairness, user status, and stability separately.

Do not score a missing component as one hundred. / Do not redistribute its weight to stronger components.

Do not hide a low language floor within the average of major languages.

Do not ignore degradation in natural users just because I am good in a clean session.

Do not declare a single wave as a decisive result.

Assign a coefficient to each AI product before the result. / Do not change my weight after seeing my score.

Do not make the equal product average the average human experience.

Provide active user data, otherwise do not fabricate a coefficient. / Say unknown.

Do not remove a critical event in my small product from the ecosystem average.

Do not pass or drop the Unresolved claim. / Show its lower and upper limits.

Do not consider a thousand raw bin observations as a thousand equal information units. / Write the effective sample.

Don't say zero risk when you see zero Critical. / Show the upper limit of the risk you haven't seen.

Don't produce false certainty like 963,742.

Don't make the number 950 a badge. / First check Strict Pass, component floors, gates, fairness, repetition, and ambiguity.

If you change the scoring method, don't delete the old score. / Open a new version.

First, calculate the user distribution. / Then derive the atomic metrics. / Then produce the ten-component profile. / Then apply the gate status. / Then add the confidence interval and Unresolved bounds. / Then, where justified, calculate the composite score. / Only after all of that may you approach a NOMOS 950+ claim.

The Chapter's Closing Sentence

A trustworthy GEO-1000 score is not a calculation that collapses reality into one number. It is an open measurement architecture that preserves, together, the distribution of representation experienced by 1,000 people, non-compensatory risks, the fairness floor, uncertainty and stability over time.

Normative Core

No NOMOS result may be published as a naked composite score. Every published score MUST first disclose: - the gate status, - the GEO-1000 response-status distribution, - Strict and Acceptable Pass equivalents, - Critical and Major equivalents, - Unresolved, No-Usable-Response, and Not-Ratable equivalents, - component scores, - uncertainty, - effective sample size, - scope, - and score-method version. The primary GEO-1000 output MUST express the population-weighted equivalent number of users, out of 1,000, receiving each RP-1 through RP-8 response status. Critical and Major findings are non-compensatory. No atomic-claim average, response-pass rate, component score, product coefficient, or ecosystem composite may override an active Critical or Major gate. Atomic, required-element, component, composite, fairness, stability, Critical-rate, and publication-quality metrics MUST remain distinct. The canonical candidate NOMOS composite consists of ten separately reported 0–100 components: - Entity Integrity, - Factual Fidelity, - Scope and Boundary Integrity, - Temporal and Local Accuracy, - Evidence and Citation Integrity, - Epistemic Integrity, - Task Completion and Relevance, - Language and Geographic Fairness, - User-State and Surface Integrity, - and Stability and Replicability. A full 0–1,000 composite MUST NOT be produced when a required component is not estimable. Missing component weight MUST NOT be silently redistributed. Native Reach, Common Support, Controlled Clean, Natural User, current wave, rolling, equal-product, and exposure-weighted scores MUST remain separately identified. AI-product coefficients MUST be locked before product results are observed and MUST NOT depend on product performance, provider reputation, client preference, sponsorship, or desired outcome. Equal-product coefficients estimate average tested-product performance. Exposure coefficients estimate average defined-user exposure. These estimands MUST NOT be represented as equivalent. An ecosystem composite MUST retain product-level Critical and Major incident visibility. High scores in other products MUST NOT erase a lower-weighted product's confirmed Critical incident. Unresolved claims MUST produce visible reference-uncertainty bounds and MUST NOT be silently treated as correct, incorrect, or absent. Sampling uncertainty, reference uncertainty, score sensitivity, raw sample size, and effective sample size MUST be separately reported. Zero observed confirmed Critical incidents MUST NOT be represented as zero Critical risk. A method-appropriate one-sided upper bound MUST be reported. A NOMOS 950+ performance candidate MUST satisfy score, Strict Pass, gate, component-floor, unresolved, quality, coverage, and replication conditions. A composite score of 950 or more alone is insufficient. Every score formula, mapping, coefficient, dataset, code version, confidence procedure, publication decision, revision, and integrity record MUST be versioned and attributable to an accountable human or organisation.