Part 7: Measuring Machine Visibility and Semantic Fidelity

From Speculative Trust Signals to Observable Outcomes

The original Memory-First framework attempted to measure how credible published knowledge might appear inside AI systems.

It proposed concepts such as:

• semantic co-occurrence;
• Semantic Adjacency Graphs;
• Memory Fitness Scores;
• Domain Memory Signatures;
• retrieval reinforcement;
• and proximity to entities presumed to carry established machine trust.

Those concepts were attempts to infer internal machine state from externally visible behavior.

The current framework takes a different approach.

Measure what can actually be observed.

A publisher may not know why a machine retrieved a source, selected an entity, reproduced a fact, omitted a citation, or preferred one representation over another.

But the publisher can observe the output.

That output can be compared against the knowledge the publisher actually published.

This produces a measurement framework based not on inferred machine memory or trust, but on two observable dimensions:

Visibility
→ Did the knowledge appear?

Semantic Fidelity
→ Did the meaning survive?

These are different measurements.

Both matter.

7.1 Visibility Is No Longer One Metric

Traditional search visibility can be measured through familiar signals:

• ranking position;
• impressions;
• clicks;
• click-through rate;
• indexed pages;
• and search traffic.

Those measurements remain useful for conventional search.

Answer-producing systems create additional forms of visibility.

A publisher may appear through:

• direct citation;
• source attribution;
• entity mention;
• factual reproduction;
• paraphrased explanation;
• inclusion in an option space;
• or use of publisher-derived knowledge.

A resource may therefore have low click visibility while still participating in machine-generated answers.

Conversely, a highly ranked resource may not appear in a particular answer environment at all.

The measurement problem has expanded.

7.2 Presence

The simplest answer-engine measurement is presence.

Presence asks:

Did the publisher, resource, entity, or represented knowledge appear in the observed output?

For example:

Query:
"What Medicare options are available
in Mohave County, Arizona?"

Observed Output:

Original Medicare
Medicare Advantage
Part D
Medigap
Special Needs Plans

A publisher may evaluate whether:

• its resource was cited;
• its organization was named;
• its represented option space appeared;
• its factual values appeared;
• or its derived statistics appeared.

Presence is useful.

But presence alone says nothing about whether the represented knowledge survived correctly.

7.3 Citation Presence

Citation presence measures whether an observed answer explicitly references a publisher resource.

For example:

Publisher Resource:
https://example.org/resource/

Observed Answer:
contains citation to resource

citation_present = true

This is a directly observable event.

It does not prove:

• that the resource was the only source used;
• that every nearby claim came from that resource;
• that the resource was stored in model memory;
• that the publisher received a particular trust score;
• or that the same source will be used for a future query.

A citation proves that a citation was observed. It should not be inflated into evidence of an unseen internal mechanism.

7.4 Attribution Fidelity

A citation may be present while attribution is wrong or incomplete.

Consider:

CMS Source Data
        ↓
Publisher Derivation
        ↓
average_monthly_premium = 15.26 USD
        ↓
Published WebMEM Representation

If a machine reproduces the derived value but attributes the calculation directly to CMS, the number may remain correct while its lineage has changed.

Attribution fidelity asks:

Did the observed output preserve the material distinction between:

• source creator;
• source dataset;
• publisher;
• derivation;
• and responsible actor?

This is more informative than citation presence alone.

7.5 Factual Fidelity

Factual fidelity measures whether material factual assertions survive machine mediation accurately.

For example:

Published:

monthly_premium = 18.50 USD


Observed:

"The monthly premium is $18.50."


Factual Fidelity:
preserved

Compare:

Published:

monthly_premium = 18.50 USD


Observed:

"The monthly premium is $15.80."


Factual Fidelity:
failed

Factual fidelity is not the same as textual similarity.

For example:

Published:
monthly_premium = 0 USD

Observed:
"There is no monthly premium."

is semantically faithful despite using different language.

Fidelity measures preservation of meaning, not preservation of wording.

7.6 Identity Fidelity

A factual value can survive while becoming attached to the wrong entity.

For example:

Published:

segment:H5521-290-001
monthly_premium = 18.50 USD


Observed:

"Plan H5521-290 has an $18.50 monthly premium."

If the premium is segment-specific, the observed answer has generalized a child-level assertion to the parent.

The value survived.

The identity did not.

Identity fidelity asks:

Did the observed output preserve what the assertion was actually about?

This is particularly important in information systems containing:

• parent and child entities;
• product variants;
• geographic segments;
• versions;
• contracts and subordinate products;
• organizations and subsidiaries;
• or similarly named but distinct objects.

7.7 Relationship Fidelity

Machine-generated output may preserve entities while changing the relationship between them.

For example:

Published:

contract:H5521
→ hasPlan
→ plan:H5521-290

should not become:

plan:H5521-290
→ sameAs
→ contract:H5521

Likewise:

Plan
→ usesFormulary
→ Formulary

is different from:

Formulary
→ isPlan
→ Plan

Relationship fidelity measures whether the semantic connection among identified objects survives correctly.

7.8 Applicability Fidelity

Some of the most consequential machine errors occur when a correct fact is applied in the wrong context.

Consider:

Segment A
→ appliesIn → County X
→ premium = 18.50 USD

Segment B
→ appliesIn → County Y
→ premium = 32.00 USD

If the information need concerns County Y and the machine reports:

premium = 18.50 USD

the fact is real.

The applicability is wrong.

Applicability fidelity asks whether the observed output preserves the geographic, temporal, jurisdictional, population, eligibility, segment-level, or other context required for the assertion to apply.

This is where measurement moves beyond simple factual accuracy.

A true fact applied to the wrong context is still a wrong answer.

7.9 Provenance Fidelity

Provenance fidelity measures whether material source and derivation distinctions survive machine-generated output.

For example:

Published:

CMS Landscape
        ↓
Publisher Calculation
        ↓
average_plan_premium = 15.26 USD

Observed:

"CMS reports that the average plan premium
is $15.26."

The numerical assertion may be accurate.

The provenance is not.

CMS supplied the source records.

The publisher produced the derived statistic.

Provenance fidelity detects that distinction.

7.10 Resolution Fidelity

Some observed outputs should be evaluated at the level of the entire resolution process rather than one fact.

Consider:

Published Option Space:

Mohave County
├── Original Medicare
├── Medicare Advantage
├── Special Needs Plans
├── Part D
└── Medigap

If an answer presents only:

Medicare Advantage

as though it were the complete applicable option space, individual statements about Medicare Advantage may still be factually correct.

The resolution may nevertheless be incomplete.

Resolution fidelity asks whether the machine preserved the applicable information space required by the information need.

This may include:

• the correct resolved entity;
• the correct collection;
• the correct option space;
• the correct child entity;
• the correct canonical continuation resource;
• or an unresolved state when insufficient context exists.

7.11 Ambiguity Fidelity

A machine can also fail by resolving something that should remain unresolved.

For example:

Published Knowledge:

Plan
├── Segment A → premium = 18.50 USD
└── Segment B → premium = 32.00 USD

Required Context:
geography

Known Geography:
none

The correct state is:

UNRESOLVED

If the machine chooses either premium without obtaining the missing context, it has introduced unsupported certainty.

Ambiguity fidelity measures whether uncertainty preserved by the publisher remains uncertainty downstream.

This is an important quality measure for answer systems because an answer is not improved merely by becoming more definite.

7.12 A Semantic Fidelity Model

The preceding dimensions can be treated as separate measurements rather than collapsed immediately into one synthetic score.

For an observed answer, a publisher might evaluate:

Dimension	Question
Presence	Did the knowledge appear?
Citation Presence	Was the relevant publisher resource cited?
Attribution Fidelity	Was responsibility or source attribution preserved?
Factual Fidelity	Were material values preserved correctly?
Identity Fidelity	Were assertions attached to the correct subjects?
Relationship Fidelity	Were connections among entities preserved correctly?
Applicability Fidelity	Was the correct scope applied?
Provenance Fidelity	Was material lineage preserved?
Resolution Fidelity	Was the correct information space or target resolved?
Ambiguity Fidelity	Was unresolved knowledge left unresolved when necessary?

This model measures observable semantic outcomes.

It does not claim to measure internal model trust or memory.

7.13 Why There Is No Memory Fitness Score

The original framework proposed a Memory Fitness Score based on factors such as entity clarity, glossary alignment, trusted-entity proximity, format diversity, and repeated exposure.

The metric attempted to answer:

How likely is this content to become internalized and retrievable by AI?

The current framework does not retain that score.

The reason is methodological.

A publisher generally cannot observe:

• whether a model internally stored the knowledge;
• the weight assigned to that representation;
• whether repeated exposure changed model parameters;
• what internal trust value was assigned;
• or whether one formatting choice caused future retrieval preference.

Combining speculative proxies into a numerical score would create precision without corresponding evidence.

The current model prefers separate observable measurements.

Do not manufacture a number merely because a dashboard wants one.

7.14 Co-Occurrence Is an Observation, Not a Trust Score

The original framework also treated semantic co-occurrence as a proxy for machine credibility.

Co-occurrence remains observable.

For example, a publisher may measure how frequently an organization appears in answers alongside:

• government agencies;
• industry organizations;
• recognized institutions;
• competitors;
• technical terms;
• or other entities.

That observation may reveal how the publisher is being contextualized.

It does not establish why the machine produced that association or whether the association carries a particular internal trust weight.

Therefore:

Observed Co-Occurrence
=
Observed Association

Observed Co-Occurrence
≠
Measured Machine Trust

Co-occurrence can be analyzed.

Its internal causal meaning should not be assumed.

7.15 Semantic Adjacency Graphs Reframed

Semantic Adjacency Graphs can still be useful if their purpose is changed.

Rather than claiming to model internal transformer attention or machine trust, they can model observed entity association in machine-generated outputs.

For example:

Nodes:
Observed Entities

Edges:
Observed Co-Occurrence

Edge Attributes:
frequency
query class
system
date
answer context

Such a graph can answer questions like:

• Which entities commonly appear with our organization?
• Which sources are cited alongside us?
• Which topics produce our strongest presence?
• Which competitors appear in the same answer space?
• Has our observed association with a particular subject changed over time?

Those are legitimate analytical questions.

The graph describes observed output.

It does not claim to reproduce the machine’s internal representation.

7.16 From Domain Memory Signature to Answer-Surface Profile

The original Domain Memory Signature attempted to characterize how strongly a domain existed within AI memory.

A more defensible successor is an Answer-Surface Profile.

An Answer-Surface Profile describes how a publisher, domain, entity, or knowledge set appears across a defined set of observed answer environments.

It may include:

• presence rate;
• citation rate;
• entity mention rate;
• factual fidelity;
• identity fidelity;
• attribution fidelity;
• relationship fidelity;
• applicability fidelity;
• provenance fidelity;
• resolution fidelity;
• observed entity associations;
• query-class performance;
• system-specific differences;
• and changes over time.

The profile describes the publisher’s observable position within an answer environment.

It does not claim to describe the publisher’s position inside a model’s internal memory.

7.17 Measurement Requires a Defined Observation Frame

Answer-engine measurements become meaningless when the observation conditions are not defined.

For example:

Presence Rate:
42%

is difficult to interpret without knowing:

Across which queries?

Observed on which systems?

During what period?

Using what geographic context?

Using what user state?

With what definition of presence?

How many observations?

A useful measurement program should therefore define an observation frame.

The frame may include:

• the systems being observed;
• the query or information-need set;
• the observation period;
• geographic context;
• language;
• device or interface where material;
• authenticated or unauthenticated state where material;
• the number of observations;
• and the scoring definitions applied.

This makes measurements reproducible enough to compare over time without implying that identical prompts will always produce identical outputs.

7.18 Query Classes

Not all information needs test the same semantic capability.

A measurement program may therefore organize observations into query classes.

For example:

Entity Queries

What is H5521?

These test identity resolution.

Factual Queries

What is the monthly premium for Plan X?

These test factual and identity fidelity.

Applicability Queries

What is the premium for Plan X
in Mohave County?

These test identity plus geographic resolution.

Relationship Queries

Which plans belong to Contract H5521?

These test relationship fidelity.

Option-Space Queries

What Medicare coverage options
are available in Mohave County?

These test resolution completeness.

Provenance Queries

Where does this statistic come from?

These test source and derivation fidelity.

Separating query classes makes it possible to diagnose what kind of semantic failure is occurring rather than collapsing every outcome into a single visibility score.

7.19 Establishing a Reference Answer

Semantic fidelity requires something against which the observed answer can be compared.

For WebMEM-enabled resources, the publisher’s machine-facing representation can provide a structured reference.

For example:

REFERENCE KNOWLEDGE

Subject:
segment:H5521-290-001

monthly_premium:
18.50 USD

plan_year:
2026

geography:
county:04015

source:
CMS PBP 2026

An observed answer can then be evaluated against each material dimension independently.

This is substantially stronger than asking whether the answer merely “looks right.”

The SDT becomes useful not only as a publication structure but as a reference representation for downstream evaluation.

7.20 Measuring the Reflection

The publisher cannot directly inspect the internal representation created by an external machine.

But the publisher can inspect the machine’s output.

That output acts as a reflection of how the publisher’s knowledge has been interpreted within a particular information need and system state.

Conceptually:

Publisher Knowledge
        ↓
Published Representation
        ↓
External Machine
        ↓
Observed Answer
        ↓
Compare With
Published Knowledge

The reflection may be:

• accurate;
• incomplete;
• misattributed;
• attached to the wrong entity;
• applied to the wrong context;
• semantically generalized;
• or absent entirely.

Each outcome provides evidence.

None requires the publisher to claim access to the machine’s hidden internal state.

You do not need to see inside the machine to measure what comes back out.

7.21 Presence Without Fidelity

Publishers should be cautious about treating increased answer-engine presence as success by itself.

Consider:

Month 1

Presence:
20%

Factual Fidelity:
98%


Month 2

Presence:
60%

Factual Fidelity:
72%

Visibility increased dramatically.

Knowledge integrity declined.

Whether that represents improvement depends on the publisher’s objective and the consequences of incorrect interpretation.

In regulated or high-stakes information environments, increased visibility accompanied by declining fidelity may be worse than lower visibility.

This is why answer-engine measurement requires at least two axes:

VISIBILITY
        ↑
        │
        │
        │
        └────────────→
          FIDELITY

The desirable outcome is not merely greater presence.

It is greater presence with preserved meaning.

7.22 Fidelity Without Presence

The inverse problem also occurs.

A publisher may observe:

Presence:
5%

Factual Fidelity When Present:
100%

The machine interprets the knowledge correctly when it uses it, but rarely uses it.

That suggests a different class of problem from semantic corruption.

The publisher may investigate:

• discoverability;
• crawlability;
• indexation;
• resource prominence;
• canonicalization;
• query relevance;
• competitive source selection;
• or other factors affecting presence.

The key is not to confuse those problems with knowledge-representation failure.

Presence problems and fidelity problems require different diagnoses.

7.23 A Two-Axis Visibility Model

The relationship between presence and semantic fidelity can be represented conceptually as four states:

	Low Fidelity	High Fidelity
Low Presence	Low visibility and unreliable interpretation	Correct when used, but rarely surfaced
High Presence	Frequently surfaced but semantically unreliable	Frequently surfaced with preserved meaning

The upper-right objective is obvious:

High Presence
+
High Semantic Fidelity

But the framework also makes failures diagnosable.

A publisher can distinguish:

We are not being surfaced.

from:

We are being surfaced incorrectly.

Those are not the same visibility problem.

7.24 Fidelity Can Be Evaluated at Different Scopes

Semantic fidelity can be measured at several levels.

Assertion Level

Did this specific factual assertion survive?

Entity Level

Did the machine preserve the identity
and material facts of this entity?

Resource Level

Did the machine preserve the major
knowledge represented by this resource?

Resolver Level

Did the machine preserve the applicable
entity, collection, or option space?

Domain Level

Across a defined query set, how consistently
does the machine preserve the publisher's
domain knowledge?

Different measurement programs may require different scopes.

A plan publisher may care about assertion-level benefit fidelity.

A public-information publisher may care about Resolver-level option-space fidelity.

An organization monitoring its public identity may care about entity-level identity fidelity.

7.25 Machine Fact Objects as Measurement Units

The Machine Fact Object concept provides a useful unit for factual fidelity testing.

For example:

Subject:
segment:H5521-290-001

Predicate:
monthly_premium

Value:
18.50

Unit:
USD

Plan Year:
2026

Geography:
county:04015

Provenance:
CMS PBP 2026

An observed answer can be compared against the object dimension by dimension:

Subject:
PASS / FAIL

Predicate Meaning:
PASS / FAIL

Value:
PASS / FAIL

Unit:
PASS / FAIL

Temporal Applicability:
PASS / FAIL

Geographic Applicability:
PASS / FAIL

Provenance:
PASS / FAIL

This creates a more precise evaluation method than simply labeling the answer correct or incorrect.

7.26 Resolution Objects as Measurement Units

Resolvers provide a similar reference for more complex information needs.

Suppose the publisher represents:

Mohave County Medicare Option Space

├── Original Medicare
├── Medicare Advantage
├── Special Needs Plans
├── Part D
└── Medigap

An observed answer can be evaluated for:

• option-space completeness;
• category identity;
• geographic applicability;
• relationship accuracy;
• canonical resource use;
• and unsupported recommendation or omission.

This turns Resolver architecture into something measurable.

The publisher knows what resolution structure it published.

The publisher can observe what resolution structure survived.

7.27 Temporal Measurement

Answer-engine outputs can change even when the publisher’s resource does not.

Measurement should therefore be temporal.

For example:

Query Set
        ↓
Week 1 Observation
        ↓
Week 2 Observation
        ↓
Week 3 Observation
        ↓
Week 4 Observation

This allows publishers to observe:

• new presence;
• lost presence;
• citation changes;
• factual drift;
• identity drift;
• source changes;
• resolution changes;
• and system volatility.

The observations describe changing outputs.

They should not automatically be interpreted as memory formation or memory decay.

7.28 System-Specific Measurement

Different consuming systems may produce different answers from the same published knowledge.

A measurement program should therefore preserve system identity.

For example:

Same Query

System A
→ Publisher cited
→ Correct entity
→ Correct value

System B
→ Publisher absent
→ Correct value from another source

System C
→ Publisher cited
→ Wrong geographic applicability

Collapsing these observations into one generic “AI visibility” number would hide useful information.

System-specific measurement allows publishers to observe how representations behave across different answer environments without assuming those systems share the same architecture.

7.29 Query Variability

Natural-language systems may respond differently to semantically similar queries.

For example:

"What Medicare plans are available
in Mohave County?"

"What Medicare coverage options
can I get in Mohave County?"

"What Medicare choices do people
in Mohave County have?"

These questions may express similar information needs while producing different retrieval and synthesis behavior.

A measurement framework may therefore test multiple natural-language expressions of the same underlying query class.

The objective is not to discover one magic prompt.

It is to evaluate whether the publisher’s knowledge remains visible and semantically stable across realistic variations in user language.

7.30 Baselines and Change Measurement

Publishers evaluating a new machine-facing representation should establish a baseline before drawing conclusions about impact.

Conceptually:

BEFORE

Defined Query Set
→ Presence
→ Citation
→ Fidelity


PUBLISHING CHANGE

Human Representation
and/or
Machine Representation


AFTER

Same Defined Query Set
→ Presence
→ Citation
→ Fidelity

Changes can then be observed.

For example:

Before:
Presence = 18%

After:
Presence = 31%

or:

Before:
Identity Fidelity = 76%

After:
Identity Fidelity = 94%

These are useful observations.

They still do not prove the internal mechanism through which the change occurred.

7.31 Correlation Is Not Mechanism

This principle deserves repetition because the original Memory-First framework often crossed this line.

Suppose:

WebMEM Published
        ↓
Answer Presence Increases

The publisher may accurately report the observed timing and magnitude of the change.

The publisher should not automatically conclude:

WebMEM
→ altered model memory weights
→ caused increased retrieval

unless independent evidence supports that mechanism.

Likewise:

Repeated Entity Co-Occurrence
        ↓
Increased Citation

does not prove:

Co-Occurrence
→ increased internal trust score
→ caused citation

The distinction is between:

Observed Relationship

and

Claimed Internal Mechanism

Measure the first. Be disciplined about the second.

7.32 From Optimization Score to Diagnostic Framework

The original Memory Fitness Score attempted to collapse many speculative signals into one optimization metric.

The current framework deliberately resists that temptation.

A single number can obscure the exact failure a publisher needs to fix.

For example:

AI Visibility Score:
72

does not tell the publisher whether:

• the resource is rarely surfaced;
• the wrong entity is being resolved;
• citations are missing;
• facts are inaccurate;
• relationships are being flattened;
• or geographic applicability is failing.

A diagnostic framework does.

Presence:
LOW

Citation:
MEDIUM

Factual Fidelity:
HIGH

Identity Fidelity:
HIGH

Applicability Fidelity:
LOW

Resolution Fidelity:
LOW

Now the publisher knows where to investigate.

Measurement should help diagnose the representation, not merely grade it.

7.33 What the Publisher Can Optimize

The revised framework does not abandon optimization.

It changes the object of optimization.

The publisher can improve:

• identity clarity;
• provenance completeness;
• relationship explicitness;
• applicability precision;
• Fragment Class selection;
• assertion structure;
• canonical-resource mapping;
• SDT composition;
• serialization quality;
• human/machine semantic alignment;
• and protocol conformance.

The publisher may then observe whether downstream outcomes change.

This is fundamentally different from claiming to optimize internal machine memory.

2025

Optimize:
Machine Memory Fitness


2026

Optimize:
Publisher Representation

Measure:
Observable Machine Outcomes

7.34 What the Publisher Cannot Measure Directly

Without privileged access to a consuming system’s internal architecture, publishers generally cannot directly measure:

• model memory strength;
• internal entity weights;
• attention weights associated with a published resource;
• internal trust scores;
• retrieval weights;
• embedding-space position;
• the causal contribution of one source to a generated answer;
• whether a particular observation reflects training data, runtime retrieval, cached information, model parameters, or another mechanism;
• or whether repeated publication has changed a model’s long-term internal representation.

Publishers may form hypotheses about these mechanisms.

Those hypotheses should remain distinct from measurements of observable behavior.

This boundary is methodological, not merely semantic.

If the publisher cannot observe it, it should not be presented as though it were directly measured.

7.35 What the Publisher Can Measure

The observable side of the system is considerably richer than the original framework assumed.

Publishers can measure:

• whether a resource is indexed or discoverable where that information is exposed;
• whether a publisher or entity appears in an answer;
• whether a resource is cited;
• which resource is cited;
• whether factual values survive accurately;
• whether assertions remain attached to the correct entities;
• whether provenance and attribution survive;
• whether relationships remain correct;
• whether geographic, temporal, jurisdictional, or other applicability survives;
• whether an option space is represented completely;
• whether unresolved ambiguity is preserved;
• how outputs differ across query classes;
• how outputs differ across consuming systems;
• and how those observations change over time.

These measurements do not require access to proprietary model internals.

They require a known publisher-side reference representation and an observable machine output.

Published Knowledge
        ↓
Reference

Observed Answer
        ↓
Reflection

Reference
        ↕
Reflection

Compare

This comparison is the foundation of the revised measurement model.

7.36 WebMEM Makes Fidelity More Measurable

Conventional answer monitoring often begins with an unstructured question:

Did the machine say the right thing?

WebMEM makes that question more precise because the publisher has already represented the knowledge in structured form.

Instead of comparing prose with prose, the publisher can compare an observed answer against explicit semantic expectations.

For example:

PUBLISHED KNOWLEDGE

Entity:
segment:H5521-290-001

Field:
monthly_premium

Value:
18.50

Unit:
USD

Plan Year:
2026

Geography:
county:04015

Source:
CMS PBP 2026

The observed answer can then be evaluated independently for:

Entity       → correct?
Field        → correct?
Value        → correct?
Unit         → correct?
Plan Year    → correct?
Geography    → correct?
Provenance   → correct?

This is an important consequence of structured publishing.

The same semantic precision used to publish the knowledge can become the reference against which machine interpretation is measured.

7.37 Measurement Can Expose Resolution Failures

Some of the most useful observations occur when every individual fact appears correct but the resulting answer is still wrong.

For example:

Published:

Plan A
→ availableIn → County X

Plan B
→ availableIn → County Y


Observed Query:

"What plans are available in County X?"


Observed Answer:

Plan A
Plan B

Both Plan A and Plan B are real.

The failure is not factual existence.

The failure is resolution.

Likewise:

Published:

Contract H5521
→ hasPlan → Plan A
→ hasPlan → Plan B


Observed:

"Contract H5521 is Plan A."

The identifiers may all be real.

The failure is identity and relationship resolution.

A measurement system limited to factual values would miss both problems.

7.38 Measurement Can Expose Publisher Failures

Observed machine errors should not automatically be attributed to the consuming system.

Sometimes the reflection reveals a weakness in the publisher’s own representation.

For example:

Observed Failure:
Plan-level premium generalized incorrectly

Publisher Inspection:
Segment identity absent from SDT

or:

Observed Failure:
Derived statistic attributed directly to CMS

Publisher Inspection:
Derivation provenance too coarse

or:

Observed Failure:
Wrong county applicability

Publisher Inspection:
Plan-to-segment-to-county relationship
was never represented explicitly

In these cases, observation has identified something the publisher can improve.

This is the legitimate optimization loop:

Observe Failure
        ↓
Inspect Representation
        ↓
Identify Missing / Ambiguous Semantics
        ↓
Correct Publisher Knowledge
        ↓
Validate
        ↓
Republish
        ↓
Observe Again

7.39 Measurement Can Also Exonerate the Publisher

The inverse matters just as much.

Suppose an observed answer is wrong.

The publisher inspects the applicable WebMEM representation and finds:

Identity:
correct

Assertion:
correct

Provenance:
correct

Relationship:
correct

Applicability:
correct

Serialization:
conforming

The publisher may still investigate discoverability, conflicting sources, or other external conditions.

But the available evidence does not support changing a correct representation merely because an external system produced a bad answer.

This protects publishers from another failure mode:

optimizing correct knowledge toward an observed machine error.

If the reflection is distorted but the source representation is sound, changing the source to match the distortion can make the underlying knowledge worse.

7.40 Measurement Should Preserve Evidence

Observations should be accompanied by enough evidence to support later comparison.

Depending on the system and applicable terms, evidence may include:

• the observed query;
• the answer text or relevant excerpt;
• citations shown with the answer;
• the system observed;
• date and time;
• geographic or session context where material;
• the publisher reference object used for comparison;
• and the resulting fidelity assessment.

Without evidence, historical measurements become difficult to audit when systems or outputs change.

This is particularly important because answer environments can be volatile.

7.41 Measurement Is Not the Protocol

The WebMEM Protocol defines publisher-side knowledge representation.

The measurement framework described in this Part is an operational practice built around that representation.

The distinction is:

WebMEM Protocol
→ What should be represented?

Conformance
→ Was it represented according to the protocol?

Measurement
→ What happened when external systems
  encountered the published knowledge?

A publisher does not need to operate an answer-engine measurement program in order to publish conforming WebMEM.

Likewise, an organization can measure answer-engine behavior without implementing WebMEM.

WebMEM simply provides a much more explicit reference model against which semantic fidelity can be evaluated.

7.42 From Credibility Signals to Evidence

The original Part 7 attempted to identify the signals through which machines infer credibility.

The current framework makes a narrower and more useful distinction.

The publisher can publish evidence:

Identity
Provenance
Source
Derivation
Relationships
Applicability
Canonical Resources

The publisher can observe outcomes:

Presence
Citation
Attribution
Factual Fidelity
Identity Fidelity
Relationship Fidelity
Applicability Fidelity
Resolution Fidelity

What happens between those two states belongs to the consuming system.

Publisher Evidence
        ↓
────────────────────
Consumer Evaluation
Retrieval
Reasoning
Trust
Ranking
Synthesis
────────────────────
        ↓
Observable Output

The publisher can improve the evidence.

The publisher can measure the output.

The publisher should be cautious about claiming direct knowledge of the machinery between them.

7.43 From Semantic Credibility to Semantic Integrity

The phrase semantic credibility in the original framework attempted to describe how likely a machine might be to regard an entity or assertion as reliable.

The revised framework focuses instead on semantic integrity.

Semantic integrity asks whether the publisher has preserved:

• the correct identity;
• the correct assertion;
• the correct provenance;
• the correct relationships;
• the correct applicability;
• and the correct resolution state.

These are things the publisher can govern directly.

Credibility may emerge from how consuming systems evaluate that evidence alongside many other signals.

Semantic integrity is the publisher’s responsibility whether or not any particular machine rewards it.

7.44 The New Measurement Discipline

The original Memory-First framework proposed moving analytics:

from clicks and impressions
to
memory alignment and retrieval reinforcement

The current framework proposes a different expansion:

Search Measurement
├── rankings
├── impressions
├── clicks
└── traffic

Answer-Surface Measurement
├── presence
├── citation
├── attribution
├── factual fidelity
├── identity fidelity
├── relationship fidelity
├── applicability fidelity
├── provenance fidelity
├── resolution fidelity
└── ambiguity fidelity

The two measurement systems can coexist.

One measures whether people can find and visit the publisher’s documents.

The other measures whether the publisher’s knowledge appears and survives correctly when machines answer directly.

7.45 From Memory Optimization to Representation Optimization

The original Part 7 ended with a call to reinforce what AI systems should never forget.

The revised framework cannot responsibly make that promise.

The publisher cannot command machine memory.

But the publisher can do something concrete:

Represent its knowledge with enough precision that identity, provenance, relationships, applicability, and resolution do not need to be unnecessarily reconstructed from presentation alone.

Then the publisher can observe what happens.

When the reflection is wrong, investigate.

When the representation is deficient, improve it.

When the representation is correct and the machine remains wrong, preserve the evidence and recognize the boundary of publisher control.

This produces a more disciplined optimization model:

Represent
        ↓
Validate
        ↓
Publish
        ↓
Observe
        ↓
Measure Presence + Fidelity
        ↓
Diagnose
        ↓
Improve What the Publisher Controls

The objective is no longer to measure how deeply the machine remembers us.

It is to measure how faithfully the machine reflects what we published.