Part 11: Representation Optimization

Optimizing What the Publisher Controls

The original Memory-First framework culminated in a discipline called Memory-First Optimization.

Its objective was to influence what AI systems remembered, retrieved, cited, and preferred through repeated exposure, format diversity, semantic adjacency, query conditioning, and retrieval feedback loops.

The current framework reaches a different conclusion.

The publisher cannot reliably optimize the private memory state of an independent machine.

The publisher can optimize something much closer to the source:

the representation it publishes.

This is Representation Optimization.

Publisher Knowledge
        ↓
Representation
        ↓
Validation
        ↓
Publication
        ↓
Machine Interpretation

The publisher controls the first four stages.

The consuming system controls the fifth.

Representation Optimization concentrates effort where the publisher has both responsibility and agency.

11.1 Optimization Did Not Disappear

Retiring Memory-First Optimization does not mean publishers should stop optimizing for machine-mediated information environments.

It means the object of optimization must be defined correctly.

The original framework asked:

How do we increase machine recall?

How do we strengthen retrieval preference?

How do we reinforce an Entity-Query Bond?

How do we prevent memory decay?

The current framework asks:

Is the subject identifiable?

Are assertions explicit?

Is provenance recoverable?

Are relationships represented?

Is applicability preserved?

Can the information need be resolved?

Are canonical resources clear?

Are human and machine representations aligned?

Is the serialization conforming?

Does the publisher remain consistent
across its own surfaces?

These questions describe properties the publisher can inspect and improve directly.

11.2 SEO and Representation Optimization Solve Different Problems

The original Part 11 was titled:

How to Replace SEO with AI Semantic Retrieval Conditioning.

The current framework rejects that premise.

SEO has not disappeared.

Search engines still crawl documents.

They still index resources.

They still rank results.

Humans still click links.

Discoverability still matters.

What changed is that search is no longer the only machine-mediated path between publisher and reader.

The publisher now faces two related optimization problems:

SEARCH OPTIMIZATION

Goal:
Help machines discover, understand,
and rank human-facing resources

Typical Concerns:
crawlability
indexation
relevance
links
information architecture
page experience
search presentation


REPRESENTATION OPTIMIZATION

Goal:
Preserve publisher-known semantics
for machine interpretation and resolution

Typical Concerns:
identity
assertions
provenance
relationships
applicability
resolution
canonical resources
semantic consistency
conformance

These disciplines can coexist.

Representation Optimization does not replace SEO. It addresses the part of machine interpretation that SEO was never designed to own.

11.3 From Visibility Optimization to Knowledge Optimization

Traditional optimization often begins with the external outcome:

Rank higher.

Get more impressions.

Earn more clicks.

Representation Optimization begins with the publisher’s knowledge:

What do we know?

How do we know it?

What is it about?

How does it relate?

Where does it apply?

What can be resolved from it?

The optimization process then asks whether those semantics survive publication.

This moves optimization upstream.

Not:

Machine Output
→ manipulate toward preferred outcome


But:

Publisher Knowledge
→ improve representation
→ observe downstream outcome

11.4 The Representation Optimization Stack

Representation Optimization can be evaluated as a series of semantic layers.

1. SOURCE INTEGRITY
        ↓
2. IDENTITY
        ↓
3. ASSERTIONS
        ↓
4. PROVENANCE
        ↓
5. RELATIONSHIPS
        ↓
6. APPLICABILITY
        ↓
7. RESOLUTION
        ↓
8. COMPOSITION
        ↓
9. SERIALIZATION
        ↓
10. CROSS-SURFACE CONSISTENCY
        ↓
11. CONFORMANCE
        ↓
12. OBSERVABILITY

A weakness at one layer may compromise the usefulness of otherwise correct knowledge.

For example:

Correct Value
+
Wrong Identity
=
Wrong Assertion

or:

Correct Value
+
Correct Identity
+
Wrong Geography
=
Wrong Applicable Answer

or:

Correct Derived Value
+
Missing Derivation
=
Broken Provenance

Optimization therefore means protecting the semantic integrity of the complete representation rather than maximizing one isolated signal.

11.5 Identity Optimization

Identity optimization asks whether every material subject can be distinguished at the level required by the knowledge.

Consider:

H5521
H5521-290
H5521-290-001

If these represent:

Contract
Plan
Segment

then collapsing them into one generic Plan identity creates semantic ambiguity.

Identity optimization may include:

• establishing stable canonical entity identifiers;
• distinguishing entity identity from resource identity;
• distinguishing parent and child identities;
• normalizing identifiers across publisher systems;
• representing aliases without creating false equivalence;
• and preserving identity across resources.

Before optimizing what the publisher says about something, make sure the publisher has represented what the thing is.

11.6 Assertion Optimization

Assertion optimization asks whether factual statements remain machine-interpretable outside their visual presentation context.

For example:

$18.50

is weaker than:

monthly_premium = 18.50 USD

which may still be weaker than:

Subject:
segment:H5521-290-001

monthly_premium:
18.50 USD

plan_year:
2026

Assertion optimization may include:

• canonical field names;
• explicit datatypes;
• canonical machine values;
• units;
• enumerated values;
• temporal scope;
• and appropriate subject binding.

The goal is not verbosity.

The goal is sufficient semantic precision.

11.7 Provenance Optimization

Provenance optimization asks whether a factual assertion retains enough lineage to understand how it came to exist.

For source-observed knowledge:

Assertion
→ sourcedFrom
→ Dataset

For derived knowledge:

Derived Assertion
→ derivedFrom
→ Source Records
→ generatedBy
→ Transformation / Method

Optimization may require distinguishing:

• source creator;
• source dataset;
• publisher;
• retrieval date;
• effective date;
• applicable period;
• transformation;
• and derivation.

Provenance optimization does not mean adding more citations.

It means preserving the lineage material to interpretation.

11.8 Relationship Optimization

Relationship optimization asks whether connections the publisher already knows remain explicit in the published representation.

For example:

Contract
→ hasPlan
→ Plan

Plan
→ hasSegment
→ Segment

Segment
→ appliesIn
→ County

Plan
→ usesFormulary
→ Formulary

If these relationships exist only in:

• database joins;
• application logic;
• URL conventions;
• navigation;
• or visual tables;

then machines may be forced to reconstruct structure the publisher already possesses.

Relationship optimization moves that structure into the published knowledge model where appropriate.

11.9 Applicability Optimization

Applicability optimization asks:

Under what conditions is this knowledge actually true or usable?

Relevant context may include:

• geography;
• time;
• jurisdiction;
• population;
• eligibility;
• segment;
• product configuration;
• or another domain-specific condition.

Consider:

monthly_premium = 18.50 USD

If the assertion applies only to:

segment:H5521-290-001
county:04015
plan_year:2026

then those distinctions are not optional decoration.

They are part of what makes the assertion applicable.

11.10 Resolution Optimization

Resolution optimization asks whether important information needs can move through the publisher’s represented knowledge to the correct semantic target.

For example:

Contract ID
→ Contract
→ Plans

or:

County
→ Applicable Coverage Option Space

or:

Plan
→ Segment
→ Geography
→ Applicable Segment
→ Premium

Resolution optimization may reveal:

• missing identities;
• missing relationships;
• resolution dead ends;
• missing canonical resources;
• collapsed option spaces;
• or missing applicability context.

Query-to-Resolution Mapping provides the primary design method for identifying these gaps.

11.11 SDT Composition Optimization

Even correct knowledge can become difficult to interpret if it is organized poorly.

SDT composition optimization asks whether knowledge has been divided according to appropriate semantic responsibilities.

For example:

Source-Observed Facts
→ DataFragment

Calculated Statistics
→ DerivedStatsFragment

Collection Membership
→ IndexFragment

Identity
→ IdentityFragment

Definitions
→ DefinedTermFragment

A common failure is to flatten semantically different knowledge into one generic data block.

Another is to fragment one coherent semantic responsibility into so many pieces that relationships and context become unnecessarily difficult to recover.

Good composition preserves semantic boundaries without destroying semantic coherence.

11.12 Serialization Optimization

The current WebMEM serialization uses HTML-in-HTML.

Serialization optimization asks whether the conceptual knowledge model survives that transformation cleanly.

The objective is not to produce clever markup.

It is to preserve:

• identity;
• Fragment Class;
• assertions;
• datatypes;
• units;
• provenance;
• relationships;
• context;
• ordering where meaningful;
• and resolution structure.

Serialization should reduce ambiguity rather than introduce it.

11.13 Human/Machine Alignment Optimization

Two-tier publishing creates a responsibility to keep the two representations semantically aligned.

For example:

Human:
"$18.50 monthly premium"

Machine:
monthly_premium = 18.50 USD

is aligned.

But:

Human:
"$18.50 monthly premium"

Machine:
monthly_premium = 32.00 USD

is not, absent an explicit scope distinction explaining the difference.

Optimization therefore includes testing:

Human Facts
↔ Machine Assertions

Human Subjects
↔ Machine Identity

Human Citations
↔ Machine Provenance

Human Scope
↔ Machine Applicability

11.14 Cross-Surface Optimization

Representation Optimization extends across publisher-controlled resources.

The goal is not to repeat the same fact everywhere.

It is to preserve coherence wherever the same entity or knowledge participates.

Optimization may include:

• canonical identity reuse;
• shared data pipelines;
• consistent inverse relationships;
• temporal consistency;
• geographic consistency;
• canonical-resource integrity;
• and elimination of stale duplicate representations.

The question is:

Do the publisher’s surfaces fit together as one coherent information environment?

11.15 Conformance Optimization

Conformance establishes the protocol contract.

Representation Optimization should therefore include automated and human review against applicable conformance requirements.

For example:

SDT Identity:
PASS

Fragment Classes:
PASS

Entity Identity:
PASS

Assertion Datatypes:
PASS

Provenance:
PASS

Relationship Integrity:
PASS

Applicability:
FAIL

Serialization:
PASS

The representation should not be treated as complete merely because most structural checks pass.

A missing applicability distinction may be the difference between a correct and incorrect answer.

11.16 From Reinforcement Vectors to Integrity Vectors

The original Memory-First Optimization model defined three reinforcement vectors:

Temporal Repetition
Format Diversity
Source Proximity

The current framework replaces them with three integrity vectors:

Semantic Precision
Cross-Surface Consistency
Resolution Coverage

1. Semantic Precision

Does the representation preserve the distinctions necessary to interpret the knowledge correctly?

Identity
Assertions
Provenance
Relationships
Applicability

2. Cross-Surface Consistency

Do publisher-controlled representations remain semantically compatible across resources and over time?

3. Resolution Coverage

Does the published knowledge support the important information needs the publisher intends to serve?

These vectors are publisher-controlled, inspectable, and testable.

11.17 Freshness Without Artificial Refresh

The original framework recommended repeated refreshing and redistribution partly to maintain machine memory.

The current framework rejects artificial freshness as a protocol objective.

A resource should change when the knowledge changes or when the representation legitimately improves.

For example:

Source Dataset Updated
→ Update Representation

Entity Relationship Changed
→ Update Representation

Methodology Changed
→ Update Derived Knowledge

Semantic Defect Found
→ Correct Representation

By contrast:

Nothing Changed
→ Change Timestamp Anyway
→ Republish for Reinforcement

does not improve the semantic integrity of the knowledge.

Freshness should represent change, not manufacture the appearance of it.

11.18 Format Diversity Without Format Proliferation

The original optimization model encouraged publishing the same knowledge in many formats to create multiple ingestion pathways.

The current framework does not treat format count as an optimization target.

Additional serializations should exist when they serve legitimate interoperability or operational requirements.

Conceptually:

Knowledge Model
        ↓
Required Serialization(s)
        ↓
Semantically Equivalent Representations

Not:

More Formats
        ↓
More Memory
        ↓
More Visibility

The current normative WebMEM serialization is HTML-in-HTML.

Future serializations may be defined where justified.

11.19 Source Authority Without Source Proximity

The original framework proposed placing publisher entities near trusted sources as a semantic reinforcement strategy.

The current framework replaces proximity with explicit provenance.

Consider:

Publisher
+
CMS
+
Statistic

Repeated textual adjacency may create an observable association.

But it does not explain the relationship among the three objects.

Explicit provenance does:

CMS Dataset
        ↓
Source Records
        ↓
Publisher Transformation
        ↓
Derived Statistic
        ↓
Published Assertion

This representation tells a consuming system:

• who produced the source data;
• which source data was used;
• what role the publisher performed;
• whether the resulting assertion was observed or derived;
• and how the published value came to exist.

The publisher does not need to manufacture semantic proximity to an authoritative institution.

Where a legitimate relationship exists, represent the relationship.

Authority should not be implied through adjacency when lineage can be stated explicitly.

11.20 Optimization Begins With the Source

A machine-facing representation cannot repair incorrect source knowledge.

If the publisher begins with:

Wrong Entity
Wrong Source Record
Wrong Value
Wrong Relationship
Wrong Applicability

then perfect WebMEM serialization merely publishes the error more precisely.

Representation Optimization therefore begins upstream of the SDT.

Source
        ↓
Ingestion
        ↓
Normalization
        ↓
Identity Resolution
        ↓
Publisher Knowledge Model
        ↓
WebMEM Representation

Each transition can introduce error.

A mature implementation should be able to distinguish:

Source Error
        ≠
Transformation Error
        ≠
Knowledge-Model Error
        ≠
Serialization Error
        ≠
Consumer Interpretation Error

This distinction makes diagnosis possible.

11.21 Optimize the Knowledge Model Before the Markup

One of the easiest mistakes in structured publishing is to begin with syntax.

The publisher asks:

Which attributes should we add?

Which Fragment Class should we use?

How should this be serialized?

before answering:

What is the entity?

What are we asserting?

Where did the assertion come from?

How does the entity relate to other entities?

Where does the assertion apply?

What must remain unresolved?

The order should be reversed.

Knowledge
        ↓
Semantic Model
        ↓
WebMEM Structure
        ↓
Serialization

Do not optimize the markup before you understand the knowledge.

11.22 Representation Optimization Is Not Data Maximization

More machine-readable data is not automatically a better representation.

A publisher may have thousands of available fields.

That does not mean every field belongs in every SDT.

For example, a county-level resource might have access to:

Plan Benefits
Enrollment
Star Ratings
Provider Networks
Pharmacy Networks
Formularies
Carrier Metadata
Historical Records
County Demographics
Pricing
Service Areas

Dumping all available data into one machine-facing representation can make the semantic scope less coherent rather than more.

Representation Optimization asks:

Which knowledge belongs to this resource, and what structure is necessary to represent it correctly?

This is selective knowledge architecture, not backend database exposure.

11.23 Optimize for Semantic Sufficiency

The goal is neither minimum data nor maximum data.

The goal is semantic sufficiency.

A representation is semantically sufficient when it preserves enough structure to interpret the knowledge correctly for its intended scope without unnecessary duplication or unsupported inference.

For example:

Value:
18.50

may be insufficient.

Field:
monthly_premium

Value:
18.50

Unit:
USD

may still be insufficient.

Subject:
segment:H5521-290-001

Field:
monthly_premium

Value:
18.50

Unit:
USD

Plan Year:
2026

Geography:
county:04015

Source:
CMS PBP 2026

may be sufficient for the intended factual resolution problem.

Adding fifty unrelated fields does not make the assertion more correct.

Optimization means publishing enough meaning—not the most possible data.

11.24 Optimize Context at the Correct Scope

Repeated metadata creates unnecessary complexity when the same context applies broadly.

WebMEM composition allows context to be declared at the broadest level where it remains true.

For example:

SDT Context:
plan_year = 2026
geography = county:04015

        ↓

Fragment Context:
source = CMS PBP 2026

        ↓

Assertion:
monthly_premium = 18.50 USD

If one assertion uses a different source or period, more-specific context can override the inherited value.

Optimization therefore asks:

What context is shared?

What context is fragment-specific?

What context is assertion-specific?

The objective is to reduce repetition without broadening context beyond where it remains valid.

Declare context at the broadest level where it is true—and no broader.

11.25 Optimize Relationships, Not Navigation

Links remain important for web navigation and discovery.

But a hyperlink does not necessarily express the semantic relationship between two objects.

For example:

Plan Page
→ link
→ Contract Page

does not tell a machine whether:

Plan
→ belongsToContract
→ Contract

or:

Plan
→ operatedBy
→ Contract

or merely:

Page
→ relatedLink
→ Other Page

Representation Optimization preserves the semantic relationship independently of the navigation mechanism.

Navigation helps a reader move.

Relationships explain why the objects connect.

11.26 Optimize Canonical Continuation

A resource does not need to contain everything known about every related entity.

But when resolution should continue elsewhere, the continuation should be clear.

For example:

County Resolver
        ↓
Plan Identity
        ↓
canonicalResource
        ↓
Plan Resource

or:

Plan Resource
        ↓
Formulary Identity
        ↓
canonicalResource
        ↓
Formulary Resource

Canonical continuation allows the publisher to build a distributed knowledge architecture without duplicating entire representations.

Optimization asks:

When this resource reaches the edge of its semantic responsibility, is the next appropriate resource identifiable?

11.27 Optimize for Explicit Ambiguity

Optimization should not force every information need toward a single answer.

Sometimes the correct semantic state is unresolved.

For example:

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

Required Context:
geography

Known Geography:
none

The optimized representation should preserve:

Resolution State:
UNRESOLVED

Required Context:
geography

rather than arbitrarily selecting one premium.

This may feel counterintuitive in an environment obsessed with answer generation.

But semantic integrity sometimes requires refusing premature resolution.

A well-optimized knowledge representation can know when it does not yet know enough.

11.28 Optimize the Option Space Before the Recommendation

Another common failure occurs when a legitimate option space is prematurely collapsed into one preferred object.

For example:

County
→ Applicable Medicare Coverage Options
│
├── Original Medicare
├── Medicare Advantage
├── Special Needs Plans
├── Part D
└── Medigap

should not become:

County
→ Medicare Advantage

merely because Medicare Advantage is commercially important to the publisher.

Likewise:

Contract
→ Plan Collection

should not become:

Contract
→ Preferred Plan

unless the representation explicitly changes semantic role and provides a legitimate basis for recommendation.

Representation Optimization preserves the information space before applying judgment to it.

11.29 Optimize Derived Knowledge Transparently

Derived statistics can provide substantial value because publishers often know how to calculate information that source agencies do not publish directly.

But derived knowledge introduces additional semantic responsibility.

For example:

Source Records:
CMS Landscape

        ↓

Method:
Arithmetic Mean

        ↓

Derived Assertion:
average_monthly_premium = 15.26 USD

        ↓

Responsible Actor:
Publisher

Optimization should preserve:

• the source inputs;
• the transformation or method;
• the scope of records included;
• the applicable period;
• and the actor responsible for the derivation.

A derived statistic becomes weaker, not stronger, when optimization hides the publisher’s role behind the source agency’s authority.

11.30 Optimize for Machine Values and Human Meaning

Two-tier publishing allows each representation to use the form appropriate to its interpreter.

For example:

Human:
"About 1 in 4 available plans
has a $0 monthly premium."

Machine:
share_zero_premium = 0.247

Datatype:
percentage

Derived:
true

The human representation may round, explain, contextualize, or narrate.

The machine representation may preserve canonical values, datatypes, units, and derivation context.

Optimization does not require forcing machine syntax into human prose or human prose into machine structures.

Optimize each representation for its interpreter while preserving semantic equivalence between them.

11.31 Optimize Definitions for Domain Meaning

Terminology should be optimized for semantic clarity rather than repeated merely for keyword or memory reinforcement.

For example:

MOOP

may be represented through:

Canonical Term:
Maximum Out-of-Pocket

Abbreviation:
MOOP

Domain:
Medicare Advantage

Definition:
...

Human-facing resources may use:

maximum out-of-pocket limit
out-of-pocket maximum
MOOP

as appropriate to context.

The machine-facing representation preserves the canonical concept.

This allows natural language flexibility without semantic fragmentation.

11.32 Optimize for Resolution Coverage

Query-to-Resolution Mapping gives Representation Optimization a practical coverage test.

For each important query family, the publisher can ask:

Entry identity represented?
        ↓
Required entities represented?
        ↓
Required assertions represented?
        ↓
Required relationships represented?
        ↓
Required applicability represented?
        ↓
Required provenance represented?
        ↓
Resolution target reachable?

The resulting gaps become optimization priorities.

For example:

QUERY FAMILY:
Plan Geographic Premium

Identity:
PASS

Premium Assertion:
PASS

Plan Year:
PASS

Segment Identity:
FAIL

Segment → County:
FAIL

Resolution Coverage:
INCOMPLETE

This produces a concrete work list.

No speculative machine-memory score is required.

11.33 Optimize According to Risk

Not every semantic defect deserves equal priority.

A publisher may prioritize optimization according to the consequences of incorrect interpretation.

For example:

LOWER RISK

Glossary synonym omitted


MODERATE RISK

Citation attribution incomplete


HIGHER RISK

Plan premium attached
to wrong segment


CRITICAL RISK

Eligibility requirement
applied to wrong population

Risk-aware optimization is particularly important in healthcare, finance, law, government, and other domains where incorrect resolution can materially affect decisions.

The most important representation is not always the one associated with the highest traffic.

11.34 Optimize Through Validation

Many representation defects can be detected before publication.

Automated validation may test:

• required identities;
• identifier uniqueness;
• recognized Fragment Classes;
• datatype compatibility;
• required units;
• reference integrity;
• provenance completeness;
• relationship structure;
• context inheritance;
• and serialization requirements.

Domain-specific validation may additionally test:

• known identifier formats;
• valid parent-child relationships;
• allowed enumerations;
• temporal consistency;
• geographic consistency;
• and domain-specific applicability rules.

Validation turns Representation Optimization from an editorial aspiration into an engineering practice.

11.35 Optimize Through Cross-Surface Tests

Publisher-controlled resources should also be evaluated together.

For example:

Plan Resource:
plan:H5521-290
→ belongsToContract
→ contract:H5521


Contract Resolver:
contract:H5521
→ hasPlan
→ plan:H5521-290

These representations agree.

But:

Plan Resource:
plan:H5521-290
→ belongsToContract
→ contract:H5521


Contract Resolver:
contract:H9999
→ hasPlan
→ plan:H5521-290

reveals a cross-surface integrity failure.

Optimization should therefore test the publisher’s knowledge environment, not merely individual pages.

11.36 Optimize Through Observation

After publication, Representation Optimization continues through the Publisher Feedback Loop.

Reference
        ↓
Publish
        ↓
Observe Reflection
        ↓
Compare
        ↓
Diagnose

If the reflection reveals:

Wrong Entity
Wrong Attribution
Wrong Relationship
Wrong Applicability
Incomplete Resolution

the publisher can inspect whether the representation contributed to the failure.

If it did, improve it.

If it did not, preserve the correct representation and continue observing.

11.37 Optimization Metrics

The original Memory-First Optimization model proposed metrics such as:

• Entity Recall Rate;
• Citation Confidence;
• Cross-Platform Retention;
• Decay Interval;
• and Entity-Query Bond strength.

The current framework separates publisher-side metrics from downstream observational metrics.

Publisher-side Representation Metrics may include:

• identity coverage;
• provenance coverage;
• relationship coverage;
• applicability coverage;
• resolution coverage;
• conformance pass rate;
• cross-surface consistency;
• human/machine alignment;
• and unresolved validation defects.

Observed Machine Metrics may include:

• presence;
• citation presence;
• factual fidelity;
• identity fidelity;
• attribution fidelity;
• relationship fidelity;
• applicability fidelity;
• provenance fidelity;
• resolution fidelity;
• and ambiguity fidelity.

The two categories should not be collapsed casually.

Representation Quality
≠
Guaranteed Machine Outcome

But comparing them over time can produce useful evidence.

11.38 A Representation Optimization Dashboard

A practical publisher dashboard might therefore look less like an SEO score and more like an engineering status view:

REPRESENTATION HEALTH

Identity Coverage             100%
Provenance Coverage            96%
Relationship Coverage          91%
Applicability Coverage         88%
Resolution Coverage            84%
Conformance Pass Rate          98%
Cross-Surface Consistency      97%
Human/Machine Alignment        99%

Separate from that publisher-side view, the organization may maintain an observational view:

OBSERVED MACHINE OUTCOMES

Presence Rate                  63%
Citation Rate                  41%
Factual Fidelity               97%
Identity Fidelity              94%
Attribution Fidelity           89%
Relationship Fidelity          92%
Applicability Fidelity         86%
Resolution Fidelity            83%

The distinction between the two views is essential.

The first measures the publisher’s representation.

The second measures observable machine output.

One is substantially under publisher control.

The other is not.

Viewed together, however, they create a useful diagnostic system.

For example:

Representation:
Applicability Coverage = 62%

Observed:
Applicability Fidelity = 58%

Possible Action:
Investigate publisher representation


Representation:
Applicability Coverage = 100%

Observed:
Applicability Fidelity = 58%

Possible Action:
Inspect evidence before changing
a correct publisher representation

The dashboard should support diagnosis rather than collapse every dimension into a single artificial score.

11.39 Do Not Optimize Toward One Consumer

A publisher may observe that one consuming system responds more favorably to a particular representation pattern.

That observation may be useful.

But WebMEM should not be distorted around undocumented behavior in one external system if doing so weakens the semantic model.

For example:

System A appears to prefer:
flattened plan-level premium

Publisher Knowledge:
premium is segment-specific

The correct optimization is not:

Flatten Segment Premium
→ Plan Premium

merely to accommodate System A.

The correct representation preserves:

Plan
→ hasSegment
→ Segment
→ Premium

Consumer behavior may change.

The underlying semantic distinction remains.

Optimize for the integrity of the knowledge before optimizing for the quirks of a particular consumer.

11.40 Representation Optimization Should Survive Model Change

This provides a useful test for the durability of an optimization practice.

Ask:

Would this improvement still make the publisher’s knowledge better if today’s dominant answer engines disappeared tomorrow?

For example:

Stable Identity
→ yes

Explicit Provenance
→ yes

Correct Relationships
→ yes

Applicability
→ yes

Canonical Resources
→ yes

Resolution Coverage
→ yes

These improvements remain useful across:

• search engines;
• answer engines;
• language models;
• agents;
• knowledge graphs;
• enterprise retrieval systems;
• and future machine consumers not yet designed.

By contrast:

Repeat exact phrase every six weeks
because Model X currently appears
to respond to it

is fragile optimization.

Durable optimization improves the knowledge representation independently of the current consumer.

11.41 Representation Optimization Is Continuous

Publisher knowledge changes.

Source data changes.

Relationships change.

Information needs change.

New resolution gaps are discovered.

Observed machine failures expose previously hidden weaknesses.

The optimization process therefore continues:

Source Knowledge
        ↓
Model
        ↓
Represent
        ↓
Validate
        ↓
Publish
        ↓
Observe
        ↓
Diagnose
        ↓
Improve
        ↓
Repeat

The purpose of the cycle is not perpetual content churn.

It is continuous integrity.

11.42 Optimization Does Not Require Constant Change

A mature representation may remain unchanged for long periods.

If:

• the source knowledge has not changed;
• identity remains stable;
• provenance remains accurate;
• relationships remain valid;
• applicability remains correct;
• resolution coverage remains appropriate;
• and no material representation defect has been identified;

then there may be nothing to optimize.

That is a successful state.

Optimization should not become an excuse to alter stable knowledge simply to create activity.

Sometimes the correct optimization decision is to leave a correct representation alone.

11.43 Representation Optimization and Protocol Evolution

Repeated optimization work may expose patterns that belong at the protocol level rather than the individual implementation level.

For example:

• the same identity failure appears across many implementations;
• a recurring relationship pattern lacks clear protocol guidance;
• a Fragment Class repeatedly carries incompatible responsibilities;
• a common resolution state cannot be represented cleanly;
• or a serialization pattern produces recurring ambiguity.

These observations may justify future clarification or evolution of the WebMEM Protocol.

The direction of learning is:

Implementation
        ↓
Observed Pattern
        ↓
Architectural Understanding
        ↓
Protocol Evolution

not:

Observed Model Behavior
        ↓
Invent Rule
        ↓
Force Protocol to Follow Model

The protocol should evolve because the knowledge architecture becomes better understood—not because one consumer temporarily behaves a particular way.

11.44 Representation Optimization and the Publisher Feedback Loop

Part 9 established the Publisher Feedback Loop:

Represent
→ Validate
→ Publish
→ Observe
→ Compare
→ Diagnose
→ Improve
→ Revalidate
→ Republish
→ Observe Again

Representation Optimization is the improvement discipline inside that loop.

The Publisher Feedback Loop answers:

When should we investigate?

Representation Optimization answers:

What should we improve when the evidence identifies a publisher-side problem?

The two concepts therefore fit together:

OBSERVATION

Machine Reflection
        ↓
Publisher Feedback Loop
        ↓
Diagnosis


INTERVENTION

Representation Optimization
        ↓
Validation
        ↓
Publication

11.45 Representation Optimization and Query-to-Resolution Mapping

Part 10 established Query-to-Resolution Mapping as the method for translating real information needs into knowledge requirements.

That process provides another input to Representation Optimization.

Information Need
        ↓
Resolution Requirement
        ↓
Query-to-Resolution Gap
        ↓
Representation Optimization

For example:

Information Need:
"What is the premium for this plan
in Mohave County?"

Resolution Requirement:
Plan
→ Segment
→ County
→ Applicable Segment
→ Premium

Current Representation:
Plan
→ Premium

Optimization Target:
Represent Segment identity
and geographic applicability

The query did not tell the publisher which words to repeat.

It revealed which semantic structure was missing.

11.46 Representation Optimization and Cross-Surface Consistency

Part 8 established that knowledge may participate across several publisher resources without requiring indiscriminate duplication.

Representation Optimization should therefore evaluate the effects of a change beyond the resource where the problem was first observed.

For example:

Optimization:
Correct Plan → Contract relationship

Affected Surfaces:
├── Plan Resource
├── Contract Resolver
├── County Resolver
└── Plan Index

A local correction that leaves contradictory relationships elsewhere is incomplete.

The optimization target is the publisher’s knowledge environment, not merely one HTML document.

11.47 Representation Optimization and Semantic Fidelity

Part 7 established semantic fidelity as the downstream measurement of whether publisher meaning survives machine mediation.

Representation Optimization and semantic fidelity therefore occupy opposite sides of the publisher boundary:

PUBLISHER SIDE

Representation Optimization
        ↓
Published Reference

────────────────────────────
      MACHINE BOUNDARY
────────────────────────────

OBSERVABLE SIDE

Machine Reflection
        ↓
Semantic Fidelity Measurement

The publisher improves the reference.

The publisher measures the reflection.

The relationship between them can be studied without pretending that one directly controls the other.

11.48 Representation Optimization Is Not AEO Scoring

The framework should not be reduced to another checklist that produces a score such as:

Your AEO Score:
87 / 100

Such a score can conceal the distinction between trivial and critical defects.

For example:

Resource A:
Score 90

Critical Defect:
Wrong geographic applicability


Resource B:
Score 82

Critical Defects:
none

Resource B may be the semantically stronger representation despite the lower aggregate score.

Representation Optimization should therefore prioritize:

• diagnostic dimensions;
• severity;
• resolution requirements;
• and domain risk;

over a single universal grade.

11.49 The Optimization Target Is Meaning

At its deepest level, Representation Optimization has one target:

preserve meaning.

Preserve what the entity is.

Preserve what is being asserted.

Preserve where the assertion came from.

Preserve how entities relate.

Preserve where and when the knowledge applies.

Preserve legitimate ambiguity.

Preserve the option space.

Preserve where resolution continues.

Preserve those semantics as knowledge moves:

Database
        ↓
Pipeline
        ↓
Publisher Knowledge Model
        ↓
SDT
        ↓
Serialization
        ↓
Public Web Resource

The optimization problem is therefore not primarily:

How do we make machines like this?

It is:

How do we stop losing meaning
on the way to publication?

11.50 From Memory-First Optimization to Representation Optimization

The evolution from the original framework can now be stated directly.

MEMORY-FIRST OPTIMIZATION — 2025

Structured Exposure
        ↓
Temporal Repetition
        ↓
Format Diversity
        ↓
Source Proximity
        ↓
Query Conditioning
        ↓
Feedback Reinforcement
        ↓
Machine Memory / Preference


REPRESENTATION OPTIMIZATION — 2026

Publisher Knowledge
        ↓
Identity
        ↓
Assertions
        ↓
Provenance
        ↓
Relationships
        ↓
Applicability
        ↓
Resolution
        ↓
Composition
        ↓
Serialization
        ↓
Conformance
        ↓
Publication
        ↓
Observation
        ↓
Publisher Learning

The first model attempted to optimize something inside the consuming machine.

The second optimizes something the publisher actually owns.

11.51 The New Optimization Discipline

Representation Optimization can therefore be defined as:

The systematic improvement of publisher-controlled knowledge representations so that identity, assertions, provenance, relationships, applicability, and resolution structure remain explicit, coherent, recoverable, and conforming when published for machine interpretation.

It does not guarantee:

• retrieval;
• ranking;
• citation;
• machine trust;
• recommendation;
• or persistent memory.

Those outcomes depend on consuming systems.

Representation Optimization improves the substrate those systems encounter.

Conclusion

The original Memory-First Optimization model asked publishers to reinforce, repeat, distribute, and condition content until machines remembered it.

The current framework asks publishers to do something both narrower and more demanding.

Know what you know.

Know what it is about.

Know where it came from.

Know how it relates.

Know where it applies.

Know what can be resolved from it.

Then publish those semantics without unnecessarily throwing them away.

After publication, observe the machine.

Measure the reflection.

Diagnose the failures.

Improve what the publisher controls.

And leave the machine’s private architecture to the machine.

The shift can be reduced to two lines:

Do not optimize the machine.

Optimize what you give it.

That is Representation Optimization.