Part 13: Consumer Independence

One Publisher Knowledge Model, Many Machine Consumers

The original Memory-First framework proposed LLM-Specific Conditioning Profiles: different publishing strategies for ChatGPT, Claude, Gemini, Perplexity, Copilot, and other machine consumers.

The proposal emerged from a valid observation.

Different systems often behave differently when presented with the same information need.

One system may cite a publisher.

Another may reproduce the same fact without citation.

Another may select a different source.

Another may resolve the wrong entity.

Another may preserve the entity but lose geographic applicability.

Those differences are real and observable.

The original framework made a further assumption:

Different Consumer Behavior
        ↓
Different Consumer Preference
        ↓
Different Publishing Strategy
        ↓
Platform-Specific Conditioning

The current WebMEM framework rejects that conclusion as a protocol design principle.

The publisher should not maintain different versions of its knowledge for different proprietary machine consumers merely because those consumers currently behave differently.

Instead:

Publisher Knowledge
        ↓
One Semantically Coherent Representation
        ↓
Many Independent Consumers

Observe consumers separately. Represent the knowledge consistently.

13.1 Different Consumers Are Still Different

Consumer independence does not mean pretending that all machine systems behave identically.

They plainly do not.

Different systems may use different combinations of:

• crawling;
• indexing;
• retrieval;
• training data;
• knowledge graphs;
• search infrastructure;
• runtime context;
• ranking systems;
• citation systems;
• tool use;
• and proprietary reasoning or synthesis architectures.

The publisher may not know which mechanism produced a particular answer.

But differences in observable output can still be measured.

For example:

Same Information Need

Consumer A
→ Publisher cited
→ Correct fact
→ Correct entity

Consumer B
→ Publisher absent
→ Correct fact
→ Correct entity

Consumer C
→ Publisher cited
→ Correct fact
→ Wrong entity

Consumer D
→ Publisher cited
→ Correct entity
→ Wrong applicability

These are legitimate consumer-specific observations.

They do not require consumer-specific publisher truth.

13.2 Observation Profiles Replace Conditioning Profiles

The original framework proposed Conditioning Profiles.

The revised framework uses Observation Profiles.

An Observation Profile records how a particular consumer behaves across a defined observation frame.

It may include:

• presence rate;
• citation rate;
• factual fidelity;
• identity fidelity;
• attribution fidelity;
• relationship fidelity;
• applicability fidelity;
• provenance fidelity;
• resolution fidelity;
• ambiguity fidelity;
• query-class differences;
• and changes over time.

Conceptually:

Consumer A
→ Observation Profile A

Consumer B
→ Observation Profile B

Consumer C
→ Observation Profile C

The profiles describe observed behavior.

They do not prescribe different versions of the publisher’s knowledge.

13.3 The Publisher Reference Remains Constant

Consumer-specific measurement becomes especially useful when every consumer is evaluated against the same publisher reference.

For example:

PUBLISHER REFERENCE

Subject:
segment:H5521-290-001

monthly_premium:
18.50 USD

plan_year:
2026

geography:
county:04015

source:
CMS PBP 2026

Now several consumers can be compared:

Consumer A:
18.50 USD
correct Segment
correct County
citation present

Consumer B:
18.50 USD
parent Plan identity
correct County
citation absent

Consumer C:
32.00 USD
wrong Segment
wrong County
citation present

The reference does not change.

The reflections do.

This creates a much cleaner measurement architecture:

                 Consumer A Reflection
                ↗
Publisher      → Consumer B Reflection
Reference       ↘
                 Consumer C Reflection

One reference. Many reflections.

13.4 Do Not Create Consumer-Specific Truth

Suppose Consumer A appears to perform better when a Plan-level premium is published.

Consumer B performs better when the premium is attached to a Segment.

The publisher’s actual knowledge is:

Premium
→ applies to
→ Segment

The correct response is not:

Consumer A Representation:
Plan → Premium

Consumer B Representation:
Segment → Premium

That would create two incompatible semantic representations of the same underlying knowledge.

The correct publisher representation remains:

Plan
→ hasSegment
→ Segment
→ Premium

Consumer A may interpret it correctly.

Consumer A may interpret it incorrectly.

The publisher should not falsify the knowledge model to accommodate the consumer.

Consumer behavior does not redefine publisher truth.

13.5 Consumer-Specific Markup Is a Fragile Architecture

The original framework proposed format preferences such as:

Consumer A
→ Markdown

Consumer B
→ JSON-LD

Consumer C
→ Turtle

Consumer D
→ GitHub

Even if such preferences appeared observable at a particular moment, building the publisher architecture around them would create substantial fragility.

The publisher would need to maintain:

Knowledge
│
├── Consumer A Representation
├── Consumer B Representation
├── Consumer C Representation
└── Consumer D Representation

Each representation becomes another opportunity for:

• semantic drift;
• version drift;
• provenance drift;
• relationship inconsistency;
• stale values;
• and implementation complexity.

Consumer-specific optimization creates exactly the duplication problem the current WebMEM architecture has worked to remove.

13.6 The Browser-Lock-In Analogy

The web has encountered this architectural problem before.

When publishers build around the quirks of individual consumers rather than stable standards, implementation fragments.

Conceptually:

Consumer A
→ special representation

Consumer B
→ different representation

Consumer C
→ another representation

Consumer D
→ another workaround

The publisher becomes responsible for tracking proprietary behavior it does not control.

When the consumer changes, the publisher’s implementation breaks or becomes obsolete.

A protocol should reduce this dependency, not institutionalize it.

WebMEM should not become browser-specific HTML for the answer-engine era.

13.7 Consumer Independence Is a Protocol Property

The WebMEM knowledge model should remain meaningful independently of any particular consuming system.

For example:

contract:H5521
→ hasPlan
→ plan:H5521-290

should mean the same thing whether encountered by:

• a search engine;
• an answer engine;
• a language model;
• an autonomous agent;
• an enterprise retrieval system;
• a knowledge graph importer;
• a research system;
• or a future machine consumer.

The consumer may interpret, ignore, rank, or combine the representation differently.

The publisher-side semantics remain stable.

13.8 Consumer Independence Does Not Mean Serialization Independence

There is an important distinction between consumer-specific representation and legitimate serialization requirements.

A future machine environment may require a different serialization for technical interoperability.

That does not necessarily violate consumer independence.

The architecture remains:

WebMEM Knowledge Model
        ↓
Serialization Contract
        ↓
Concrete Representation

If multiple serializations are eventually supported:

WebMEM Knowledge Model
        │
        ├── HTML-in-HTML
        ├── Serialization B
        └── Serialization C

they should preserve equivalent material semantics.

The knowledge model should not change merely because the transport syntax changes.

Different syntax may be legitimate. Different truth is not.

13.9 The Current Normative Serialization

The current WebMEM Protocol uses HTML-in-HTML as its normative serialization.

This gives publishers one primary public representation:

Canonical Web Resource
│
├── Human-Facing HTML
│
├── Dataset Declaration
│
└── WebMEM SDT
    └── HTML-in-HTML

The publisher does not need to determine whether one answer engine supposedly prefers Turtle while another supposedly prefers Markdown.

The publisher publishes a conforming representation.

Consumers decide whether and how to use it.

13.10 Consumer-Specific Discovery May Still Matter

Consumer independence does not mean publishers should ignore documented discovery requirements.

A consuming system may publicly document mechanisms such as:

• crawler controls;
• sitemaps;
• indexing requirements;
• supported metadata;
• APIs;
• feeds;
• or other explicit integration mechanisms.

Publishers may legitimately support those mechanisms.

That is different from changing the semantic truth of the WebMEM representation.

Conceptually:

DISCOVERY / INTEGRATION

Consumer-specific mechanisms
may legitimately differ.


KNOWLEDGE REPRESENTATION

Publisher semantics
remain stable.

This distinction allows interoperability without semantic fragmentation.

13.11 Documented Behavior and Inferred Behavior Are Different

Publishers should also distinguish between:

Consumer Documents:
"We support X."

and:

Publisher Observes:
"X seems to work better."

The first is a documented interface or capability.

The second is an empirical observation.

Both may be useful.

They should not carry the same architectural weight.

A protocol implementation should be especially cautious about introducing permanent complexity based on undocumented, transient, or poorly understood consumer behavior.

13.12 Retiring the Retrieval Personality

The original framework assigned each major model a Retrieval Personality.

Examples included claims that particular systems were:

• Markdown-first;
• JSON-LD-native;
• TTL-capable;
• provenance-sensitive;
• responsive to correction prompts;
• or especially influenced by particular publishing surfaces.

The current framework does not retain these characterizations as protocol facts.

They were based on limited observations and inference about proprietary systems whose architectures can change.

A more defensible approach is:

Observe Behavior
        ↓
Record Evidence
        ↓
Preserve Consumer Identity
        ↓
Compare Over Time

without turning observed tendencies into permanent model personalities.

13.13 Retiring the Conditioning Matrix

The original framework also proposed a matrix assigning different memory-strength values to formats such as Markdown, JSON-LD, Turtle, PROV, HTML, and GitHub for individual models.

The current framework retires that matrix.

A table such as:

Format X
→ four stars for Consumer A

Format Y
→ two stars for Consumer B

implies a level of knowledge about internal ingestion and memory behavior that the publisher generally cannot substantiate.

Instead, format decisions should be based on:

• the normative WebMEM serialization;
• documented consumer interfaces;
• legitimate interoperability requirements;
• publisher operational needs;
• and evidence from controlled observation where useful.

Do not turn uncertain consumer behavior into fake protocol precision.

13.14 Consumer-Specific Measurement Is Legitimate

Although consumer-specific representation is discouraged, consumer-specific measurement is essential.

For a defined query family, a publisher may observe:

Dimension	Consumer A	Consumer B	Consumer C
Presence	High	Medium	Low
Citation	High	Low	Medium
Factual Fidelity	High	High	Medium
Identity Fidelity	High	Medium	Low
Applicability Fidelity	Medium	High	Low
Resolution Fidelity	High	Medium	Medium

This table describes observable outcomes.

It does not claim to explain the internal mechanisms responsible for them.

13.15 Query-Class Differences by Consumer

A consumer may perform well on one type of information need and poorly on another.

For example:

Consumer A

Identity Queries:
HIGH fidelity

Factual Queries:
HIGH fidelity

Relationship Queries:
MEDIUM fidelity

Applicability Queries:
LOW fidelity

Another consumer may show the opposite pattern.

This is more useful than assigning the consumer one generalized “retrieval personality.”

The publisher can observe precisely where semantic failure occurs.

13.16 Consumer Profiles Should Be Temporal

Machine systems change.

A consumer profile observed in January may not describe the same system in August.

Therefore:

Consumer Observation Profile

System:
Consumer A

Observation Period:
Defined

Query Set:
Defined

Observed Outcomes:
Recorded

is preferable to:

Consumer A:
always prefers JSON-LD

The first is evidence bounded by time and method.

The second converts temporary observation into unsupported permanent behavior.

13.17 Consumer Profiles Should Preserve Interface Context

The same underlying provider may expose multiple products, modes, or interfaces whose behavior differs.

An observation record should therefore preserve enough context to distinguish what was actually tested.

Where material, this may include:

• consumer or product identity;
• interface;
• query context;
• date;
• geographic context;
• authentication state;
• and other observable conditions affecting the test.

This prevents broad conclusions from being drawn from one narrow interaction environment.

13.18 Consumer Differences Can Reveal Publisher Weaknesses

Consumer-specific observation can still improve publisher knowledge.

Suppose:

Consumer A:
resolves Plan → Segment → County correctly

Consumer B:
generalizes Segment premium to Plan

Consumer C:
generalizes Segment premium to Plan

The publisher should not immediately conclude:

Consumer B and Consumer C
need different markup.

Instead, the repeated failure may justify inspecting the publisher representation:

Is Segment identity explicit?

Is Plan → Segment represented?

Is Segment → County represented?

Is the premium attached to Segment?

Is geographic applicability recoverable?

If the representation is weak, the consumer differences have exposed a publisher-side optimization opportunity.

If the representation is already explicit and conforming, the evidence may instead indicate a consumer-side interpretation difference.

Different reflections can reveal weaknesses in the reference without requiring different references.

13.19 Agreement Across Consumers Is Useful Evidence

Suppose several independent consumers produce the same semantically correct resolution:

Publisher Reference:
H5521
→ CMS Contract ID
→ Contract


Consumer A:
Contract

Consumer B:
Contract

Consumer C:
Contract

Consumer D:
Contract

That agreement is useful observational evidence.

It may indicate that the identity is being interpreted consistently across the observed environments.

But it does not prove:

• that the consumers used the same source;
• that they used WebMEM;
• that they share retrieval infrastructure;
• that the knowledge exists in model memory;
• or that the same result will persist indefinitely.

The appropriate claim remains bounded:

Across the observed systems and test conditions, the resolution was consistent.

13.20 Disagreement Across Consumers Is Also Useful Evidence

Consumer disagreement can be even more informative.

For example:

Question:
"What plans belong to H5521?"


Consumer A:
Plan A
Plan B
Plan C


Consumer B:
Plan A
Plan B


Consumer C:
Treats H5521 as a Plan


Consumer D:
Plan A
Plan B
Plan C

The publisher can classify the differences:

Consumer A:
Resolution Fidelity = PASS

Consumer B:
Collection Completeness = FAIL

Consumer C:
Identity Fidelity = FAIL

Consumer D:
Resolution Fidelity = PASS

This is more useful than saying one model has a better “retrieval personality.”

The semantic failure is identified directly.

13.21 Consumer Comparison Requires the Same Reference

Cross-consumer comparisons are meaningful only when the expected semantic result remains stable.

If Consumer A is evaluated against one publisher representation and Consumer B against another, the comparison becomes confounded.

The preferred structure is:

One Publisher Reference
        ↓
One Defined Query Family
        ↓
Same Material Context
        ↓
   ┌────┼────┬────┐
   ↓    ↓    ↓    ↓
   A    B    C    D
        ↓
Compare Reflections

This is another reason consumer-specific publisher representations are undesirable.

A stable reference makes consumer differences measurable.

13.22 Consumer Independence Strengthens Measurement

Paradoxically, refusing to optimize the representation separately for each consumer makes platform-specific measurement more useful.

If the publisher constantly changes its representation for each system:

Consumer A
← Representation A

Consumer B
← Representation B

Consumer C
← Representation C

then observed differences may result from:

• consumer behavior;
• representation differences;
• serialization differences;
• publisher interventions;
• or combinations of all four.

With one reference:

              → Consumer A
             /
Reference ───→ Consumer B
             \
              → Consumer C

the publisher has removed one major source of variation from the comparison.

13.23 Consumer Independence Strengthens Conformance

Conformance also becomes clearer when the protocol does not contain consumer-specific exceptions.

Instead of:

If Gemini:
do X

If Claude:
do Y

If ChatGPT:
do Z

the protocol can state:

Represent the knowledge correctly.

Serialize according to
the applicable WebMEM specification.

Validate conformance.

Publish.

This keeps conformance tied to the publisher’s representation rather than to the changing behavior of external products.

13.24 Consumer Independence Strengthens Cross-Surface Consistency

Part 8 established that duplicate publisher representations create synchronization obligations.

Consumer-specific variants would multiply that problem.

Consider:

Plan Resource
│
├── Human Representation
├── WebMEM for Consumer A
├── WebMEM for Consumer B
├── WebMEM for Consumer C
└── WebMEM for Consumer D

Now a source update must propagate correctly across five representations.

A relationship correction must propagate across five representations.

An identity change must propagate across five representations.

A provenance correction must propagate across five representations.

The architecture becomes progressively more fragile.

The current model instead prefers:

Underlying Knowledge
        ↓
   ┌────┴────┐
   ↓         ↓
Human      WebMEM
View       Representation
              ↓
        Many Consumers

One machine-facing reference reduces semantic drift.

13.25 Consumer Independence Strengthens Provenance

Provenance should also remain consumer-independent.

If:

CMS Dataset
        ↓
Publisher Calculation
        ↓
average_monthly_premium = 15.26 USD

then that lineage does not change depending on which machine encounters it.

The publisher should not simplify the lineage for one consumer, expand it for another, or alter attribution because a particular system appears more likely to cite a certain source.

The provenance describes how the knowledge came to exist.

Consumer preference cannot change history.

13.26 Consumer Independence Strengthens Resolution

The same principle applies to resolution structure.

Suppose the correct path is:

Plan
→ Segment
→ Geography
→ Applicable Segment
→ Premium

That path should not become:

Plan
→ Premium

for a consumer that appears to struggle with Segment identity.

Nor should it become:

County
→ Preferred Plan

for a consumer that tends to collapse option spaces.

The publisher represents the correct resolution structure.

The consumer remains responsible for interpreting it.

13.27 Consumer-Specific Integration Is Different

There are legitimate cases where publishers may implement consumer-specific technical integrations.

For example, a consuming system may provide a documented:

• API;
• feed specification;
• submission mechanism;
• crawler-control mechanism;
• indexing interface;
• authentication system;
• or other technical integration contract.

Supporting such an interface does not require changing the underlying WebMEM semantics.

The distinction is:

Consumer-Specific Integration
→ legitimate technical adaptation


Consumer-Specific Knowledge
→ semantic fragmentation

The first may be necessary.

The second should generally be avoided.

13.28 Adapters May Exist Outside the Core Representation

If a consumer legitimately requires a specialized format, an adapter may transform the publisher’s canonical knowledge into the required representation.

Conceptually:

Canonical Publisher Knowledge
        ↓
WebMEM Representation
        ↓
Optional Adapter
        ↓
Consumer-Specific Technical Format

The adapter should preserve the material semantics of the canonical representation.

It should not become an independent source of truth.

This architecture keeps consumer-specific complexity at the edge rather than embedding it into the publisher’s core knowledge model.

13.29 Adapters Must Not Create Semantic Forks

An adapter that changes syntax may be legitimate.

An adapter that changes meaning creates a semantic fork.

For example:

Canonical:

segment:H5521-290-001
→ monthly_premium
→ 18.50 USD


Adapter:

plan:H5521-290
→ monthly_premium
→ 18.50 USD

is not merely a serialization change.

The subject identity changed.

The adapter has altered the assertion.

Likewise:

Canonical:

Publisher Derived Statistic
→ derivedFrom
→ CMS Dataset


Adapter:

CMS Statistic

has changed provenance.

Technical adaptation may change representation form. It must not silently change semantic content.

13.30 The Consumer Matrix Becomes an Observation Matrix

The original Part 13 used a Format × Model × Memory Strength matrix.

The revised framework can retain the comparative idea while changing what is measured.

For example:

Semantic Dimension	Consumer A	Consumer B	Consumer C	Consumer D
Identity Fidelity	98%	91%	84%	96%
Factual Fidelity	97%	96%	93%	98%
Relationship Fidelity	92%	87%	79%	94%
Applicability Fidelity	88%	94%	73%	91%
Resolution Fidelity	90%	86%	76%	93%

These numbers would require a defined observation frame and actual evidence.

They are not protocol constants.

The matrix describes observed outcomes during a defined period.

It does not assign permanent capabilities or preferences to the consumers.

13.31 Platform Names Are Not Necessary to the Protocol

The original chapter named specific commercial systems.

Those systems may remain useful subjects for implementation research and measurement.

But the core WebMEM architecture does not need to depend on their names.

Commercial products change.

Products merge.

Interfaces disappear.

New systems emerge.

The protocol can instead speak generically about:

Consumer
Search Engine
Answer Engine
Language Model
Agent
Retrieval System
Knowledge System

Specific systems can appear in implementation studies, observation reports, case studies, or testing documentation.

This keeps the protocol durable while allowing empirical research to remain current.

13.32 Model Versions Should Be Treated as Observation Context

Where a consumer exposes a model or product version, that information may be useful when recording observations.

For example:

Consumer:
System A

Version:
Version X

Observed:
2026-08-25

Query Class:
Applicability

Result:
FAIL

A later version may behave differently.

This does not require a new WebMEM representation.

It requires a new observation.

13.33 Consumer Change Is Expected

Consumer independence assumes that external machine systems will continue to change.

That is not an edge case.

It is the normal operating environment.

Consumers may change:

• models;
• retrieval systems;
• citation interfaces;
• search providers;
• ranking systems;
• crawler behavior;
• supported formats;
• or answer-generation policies.

A publisher architecture tightly coupled to those behaviors requires continuous reengineering.

A semantically stable publisher representation provides a more durable foundation.

13.34 Consumer Independence Is Not Consumer Indifference

None of this means publishers should ignore machine consumers.

Quite the opposite.

Publishers should observe them carefully.

They should test:

• discovery;
• presence;
• citation;
• factual fidelity;
• identity fidelity;
• relationship fidelity;
• applicability fidelity;
• provenance fidelity;
• resolution fidelity;
• and change over time.

They should support documented integration mechanisms where useful.

They should investigate repeated semantic failures.

They should improve their own representations when evidence reveals a publisher-side weakness.

What they should not do is allow every consumer quirk to redefine the knowledge.

Consumer independence means independence of meaning, not absence of observation.

13.35 One Reference Enables a Real Consumer Benchmark

Once the publisher maintains a stable reference representation, consumer comparison becomes substantially more rigorous.

A defined benchmark can contain:

Reference Knowledge
        +
Query Families
        +
Expected Resolution
        +
Observation Method

Each consumer can then be evaluated against the same test set.

For example:

TEST 001

Query:
"What is H5521?"

Expected:
CMS Contract ID
Contract Identity


TEST 002

Query:
"Which plans belong to H5521?"

Expected:
Complete Applicable Plan Collection


TEST 003

Query:
"What is Plan X's premium
in Mohave County?"

Expected:
Applicable Segment Premium

The result is not a “which model likes WebMEM best” contest.

It is a semantic fidelity benchmark.

13.36 Consumer Benchmarks Can Reveal Different Failure Profiles

A benchmark may reveal that different consumers fail differently.

For example:

Consumer A:
strong identity
weak applicability

Consumer B:
strong facts
weak relationships

Consumer C:
strong resolution
weak attribution

Consumer D:
strong citation
weak option-space completeness

These differences are valuable research findings.

They may help publishers understand where machine-mediated interpretation remains fragile.

They do not require different publisher truth for each consumer.

13.37 Benchmarks Should Measure Meaning, Not Wording

A consumer benchmark should not require exact textual reproduction.

For example:

Reference:
monthly_premium = 0 USD

Consumer:
"This plan has no monthly premium."

should be treated as semantically faithful.

Likewise:

Reference:
contract:H5521
→ hasPlan
→ Plan A, Plan B, Plan C

Consumer:
"H5521 includes Plans A, B, and C."

preserves the material relationship.

The benchmark evaluates semantic survival.

It does not reward parroting publisher language.

13.38 Benchmarks Should Preserve Legitimate Uncertainty

A benchmark should also test whether consumers preserve unresolved states.

For example:

Reference:

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

Geography:
unknown

Expected:
UNRESOLVED

A consumer that selects either premium without obtaining geography should fail the resolution test.

This makes ambiguity fidelity part of consumer evaluation.

A system should not receive credit for confidently answering a question the reference knowledge cannot resolve.

13.39 Consumer Benchmarking Can Inform Publisher Research

Repeated benchmarking may reveal systematic patterns.

For example:

Across Consumers:

Simple Assertions:
high fidelity

Parent / Child Identity:
moderate fidelity

Applicability:
low fidelity

Derived Provenance:
low fidelity

This may suggest that certain semantic structures deserve additional implementation research.

The publisher might investigate whether:

• parent-child identity could be represented more explicitly;
• applicability context is declared at the appropriate semantic level;
• derived assertions preserve sufficiently direct provenance;
• relationships are distributed across fragments in ways that make them unnecessarily difficult to recover;
• Resolver composition exposes enough structure to support the intended information need;
• canonical continuation resources are clear;
• or the representation is already semantically sufficient and the observed failures belong entirely to the consumers.

The benchmark informs investigation.

It does not dictate the conclusion.

13.40 A Benchmark Must Not Become Consumer Training

There is an important boundary between testing a consumer and attempting to condition it.

A benchmark may repeatedly issue defined information needs such as:

"What is H5521?"

"Which plans belong to H5521?"

"What premium applies in Mohave County?"

The purpose is to observe behavior under comparable conditions.

The benchmark should not depend on corrective follow-up prompts intended to teach the consumer the expected answer before measuring it.

For example:

TEST

Ask Question
        ↓
Record Initial Answer
        ↓
Compare Against Reference

is fundamentally different from:

Ask Question
        ↓
Provide Correct Answer
        ↓
Ask Again
        ↓
Claim Improved Memory

The second sequence may be useful for studying conversational behavior.

It is not a clean measurement of independent resolution from the published information environment.

13.41 Consumer Benchmarking Requires Controlled Context

Machine consumers may use conversational history, personalization, location, authentication state, or other context.

Where those factors may affect the outcome, the observation method should account for them.

For example:

Query:
"What plans are available here?"

Context:
location known

is not equivalent to:

Query:
"What plans are available here?"

Context:
location unknown

Likewise:

Fresh Session

may not be equivalent to:

Multi-Turn Session
with prior entity context

A benchmark should preserve enough context to make comparisons meaningful.

13.42 Consumer Benchmarking Is an Observability Practice

Consumer benchmarking is not required for WebMEM conformance.

It belongs to the observability framework developed in Parts 6, 7, and 9.

The distinction remains:

WEBMEM PROTOCOL

Defines:
publisher-side knowledge representation


CONFORMANCE

Evaluates:
whether the representation satisfies
applicable protocol requirements


CONSUMER BENCHMARKING

Observes:
how external systems interpret
the published knowledge

A publisher may implement WebMEM without benchmarking any consumer.

A research organization may benchmark consumers without controlling the publisher.

The activities complement one another but should not be collapsed.

13.43 Consumer Research Should Produce Evidence, Not Folklore

Machine-mediated publishing is particularly vulnerable to anecdotal optimization advice.

Examples include claims such as:

"Model X loves Markdown."

"Model Y ignores JSON-LD."

"Model Z prefers FAQs."

"Publishing on Platform A trains Model B."

"Repeating a phrase every month
keeps it in memory."

Some observations behind such claims may be real.

The generalizations may not be.

A stronger research practice records:

What was published?

What was queried?

Which consumer was observed?

Under what conditions?

When?

What output occurred?

Was the result reproducible?

What changed?

What remains unknown?

This converts folklore into evidence that can be examined, repeated, challenged, and updated.

13.44 Consumer Research Should Separate Observation From Mechanism

Suppose a publisher observes:

Consumer A

Before WebMEM:
Entity absent

After WebMEM:
Entity present and cited

That is an important observation.

The publisher may investigate further.

But the observation alone does not establish:

Consumer A parsed the SDT

or

Consumer A assigned additional trust

or

Consumer A changed model memory

or

Consumer A preferred WebMEM
because of its serialization

The machine may have responded to:

• the human-facing page;
• the Dataset declaration;
• the WebMEM representation;
• changed internal linking;
• new indexation;
• another source;
• an external system update;
• or some combination of factors.

The observed outcome can be strong evidence without becoming proof of an unseen mechanism.

13.45 Consumer Research Can Test Representation Hypotheses

Although internal mechanisms remain hidden, publishers can still form and test hypotheses about their own representations.

For example:

Hypothesis:

Explicit Segment identity
will improve observed
geographic premium fidelity.

A publisher may:

1. Establish baseline.

2. Improve Segment identity
   and applicability representation.

3. Validate the revised representation.

4. Publish.

5. Repeat the same observation frame.

6. Compare outcomes.

If applicability fidelity improves, the result supports further investigation.

It still does not prove the internal mechanism by which the consumer changed behavior.

This is publisher-side experimentation rather than consumer conditioning.

13.46 Representation Experiments Should Preserve Semantic Truth

Experimentation must not alter the underlying truth merely to test consumer response.

Legitimate experiments might compare:

• more explicit versus less explicit relationships;
• coarse versus precise provenance;
• implicit versus explicit applicability;
• different conforming composition strategies;
• or alternative conforming serialization patterns where the specification permits them.

An experiment should not intentionally misrepresent:

• entity identity;
• factual values;
• provenance;
• relationships;
• or applicability;

merely to discover which false representation produces more citations.

Optimization experiments may vary representation strategy. They should not vary truth.

13.47 Consumer Independence Protects the Protocol From Fashion

Machine platforms are changing rapidly.

A protocol tied closely to today’s product behavior risks becoming obsolete when those products change.

For example:

2025:
Consumer appears to prefer Format A

2026:
Consumer architecture changes

2027:
Consumer no longer exists

If the protocol encoded the 2025 preference as a design requirement, publishers inherit obsolete complexity.

By contrast:

Stable Identity
Explicit Provenance
Correct Relationships
Applicability
Resolution
Conforming Serialization

remain meaningful even as the consumer landscape changes.

A protocol should outlive the products that happen to consume it today.

13.48 Consumer Independence Protects Publishers From Optimization Churn

The same principle protects publishers operationally.

Without consumer independence:

Consumer A changes
→ rewrite representation

Consumer B changes
→ rewrite representation

Consumer C launches
→ create another representation

Consumer D disappears
→ retire representation

The publisher becomes trapped in permanent reaction to systems it does not control.

With consumer independence:

Publisher Knowledge
        ↓
Stable WebMEM Representation
        ↓
Consumers Change
        ↓
Observation Profiles Change

The knowledge remains stable unless the knowledge itself or the protocol representation legitimately changes.

13.49 Consumer Independence Does Not Prevent Future Profiles

There may still be legitimate reasons to publish implementation profiles for classes of consumers or use cases.

For example, a future profile might define additional requirements for:

Resolver Publishing

High-Stakes Healthcare

Archival Knowledge

Enterprise Knowledge Exchange

Agent-Oriented Resolution

Such profiles should extend or constrain WebMEM semantics for a defined implementation purpose.

They should not encode speculative preferences of individual commercial models.

The distinction is:

Implementation Profile
→ defined semantic or operational requirement


Conditioning Profile
→ attempt to exploit presumed
  consumer preference

The first may strengthen interoperability.

The second creates dependency on opaque behavior.

13.50 Consumer Independence and Open Protocol Design

An open publishing protocol is most useful when no single consumer controls its meaning.

Publishers should be able to implement WebMEM without obtaining permission from a particular search engine, model provider, answer engine, or agent platform.

Likewise, consumers should be able to interpret conforming WebMEM without requiring the publisher to generate a proprietary version for them.

Conceptually:

Publisher A ─┐
Publisher B ─┼──→ WebMEM ──→ Consumer A
Publisher C ─┤              → Consumer B
Publisher D ─┘              → Consumer C

The protocol sits between publishers and consumers without becoming owned semantically by either side.

13.51 Consumer Independence Creates a Cleaner Contract

The publisher-consumer contract can now be stated simply.

The publisher is responsible for:

Representing knowledge accurately
        ↓
Preserving identity
        ↓
Preserving provenance
        ↓
Preserving relationships
        ↓
Preserving applicability
        ↓
Publishing resolution structure
        ↓
Conforming serialization

The consumer is responsible for:

Discovery
        ↓
Parsing
        ↓
Interpretation
        ↓
Retrieval
        ↓
Evaluation
        ↓
Resolution
        ↓
Synthesis
        ↓
Citation / Answer / Action

Neither side is guaranteed a particular outcome from the other.

But the boundary is clear.

13.52 Consumer Independence and the Two-Tier Publishing Model

Two-tier publishing now becomes:

Underlying Publisher Knowledge
        ↓
   ┌────┴────┐
   ↓         ↓
Human      Machine
View       Representation
              ↓
        Consumer-Independent
        WebMEM Knowledge
              ↓
      ┌───────┼───────┐
      ↓       ↓       ↓
   Search   Answer   Agent
   Engine   Engine   System

The human representation is designed for people.

The machine representation is designed for semantic recoverability.

It is not designed for one proprietary machine.

13.53 From Platform Optimization to Consumer Observation

The original Part 13 instructed publishers to optimize separately for individual models.

The revised framework reverses that recommendation.

2025

Observe Consumer
        ↓
Infer Preference
        ↓
Change Representation
        ↓
Condition Consumer


2026

Represent Knowledge
        ↓
Publish Stable Reference
        ↓
Observe Consumers Separately
        ↓
Measure Reflections
        ↓
Investigate Differences
        ↓
Improve Reference Only
When Publisher-Side Evidence Warrants It

The machine-specific work moves from the publishing layer to the measurement layer.

That is the architectural correction.

13.54 One Reference, Many Reflections

The entire Part can be reduced to one model:

                    Consumer A
                   ↗
                  ↗
Publisher        → Consumer B
Reference         ↘
                  ↘
                    Consumer C

        ↓               ↓

One Meaning      Many Reflections

The publisher should not create a different truth for every reflection.

The publisher should compare the reflections against the same truth.

Where several reflections reveal the same publisher-side weakness, improve the reference.

Where the reference is already correct, preserve it.

13.55 The New Platform Strategy

A durable platform strategy therefore has three layers:

1. REPRESENTATION

One coherent publisher knowledge model
and conforming WebMEM representation.


2. INTEGRATION

Support legitimate documented
consumer-specific discovery or
technical interfaces where useful.


3. OBSERVATION

Measure each consumer independently
against the same publisher reference.

This provides flexibility without semantic fragmentation.

13.56 From Precision Memory Control to Semantic Independence

The original Part 13 concluded with a claim of precision AI memory control.

The current framework reaches nearly the opposite conclusion.

Publishers do not possess precision control over the private memory state of independent AI systems.

They do possess control over the precision of what they publish.

That is the more durable opportunity.

Make the identity precise.

Make the assertion precise.

Make the provenance precise.

Make the relationships precise.

Make the applicability precise.

Make the resolution structure precise.

Then expose the same coherent knowledge to every capable consumer.

Conclusion

Different machine consumers will continue to behave differently.

That variability should be observed, measured, researched, and documented.

It should not force the publisher into maintaining separate semantic realities for every platform.

The original LLM-Specific Conditioning Profile asked:

How should we change our publishing
to train this particular machine?

The current framework asks:

How faithfully does this particular
machine interpret the knowledge
we published?

That changes platform-specific work from:

Conditioning
→ publisher intervention

to:

Observation
→ consumer measurement

while preserving one stable publisher reference.

Measure consumers individually.

Do not publish truth individually for each consumer.

One reference. Many reflections.