The sections overlap because the current document mixes feature categories with cross-cutting behavioral dimensions. Some B-codes are relatively discrete causal features. Others are better understood as axes that modulate several other features. Making them strictly mutually exclusive is possible only if the document shifts from “causal feature list” to a more formal classification system.

Core issue#

The current Layer 1B taxonomy is not fully mutually exclusive because several entries describe different views of the same generation process:

User prompt
→ model infers task
→ examples and wording condition interpretation
→ model chooses response posture
→ model generates an output
→ output reflects style, policy, confidence, knowledge, and competence priors

The B-sections slice this process at different abstraction levels. That is why overlap appears.

The taxonomy currently contains three kinds of things:

1. Primary behavioral mechanisms
   Example: task induction, demonstration conditioning, cooperative completion

2. Modulators of those mechanisms
   Example: interface sensitivity, distribution-conditional competence

3. Output/assessment properties
   Example: plural valid-output space, generated confidence language

Those are not the same kind of category. That is the main structural reason the sections overlap.

1. Major overlaps between current sections

B1 / B2 / B3: task interpretation cluster#

Current overlap#

B1 — Learned Natural-Language Task Induction
B2 — In-Context Demonstration Conditioning
B3 — Natural-Language Interface Sensitivity

These three are tightly coupled.

B1 says the model infers the task from natural-language context.

B2 says examples inside that context shape the inferred task.

B3 says small changes in the natural-language interface can shift the inferred task, output, or decision boundary.

So B2 and B3 are not fully independent from B1. They are partly sub-mechanisms or modifiers of task induction.

Cleaner distinction#

A more exclusive version would define them this way:

Under this framing, B3 is not a task feature itself. It is a sensitivity axis over B1, B2, B5, B8, B9, and B10.

Recommendation#

Do not treat B3 as a peer feature if strict exclusivity is required. Treat it as a cross-cutting axis:

Interface sensitivity = how much behavior changes under equivalent or near-equivalent prompt variation.

Then B3 becomes a measurable property of other features, not a separate feature.

B1 / B8: task inference vs response pressure#

Current overlap#

B1 — What task is being requested?
B8 — Should the model satisfy the apparent request directly?

These are related but separable.

B1 determines the model’s interpretation of the operation:

summarize, classify, extract, translate, answer, critique, plan

B8 determines the model’s posture once a request-like context has been recognized:

answer directly, clarify, abstain, challenge, refuse, use tool

Cleaner distinction#

Recommendation#

Keep B1 and B8 separate. They are not mutually exclusive in execution, but they are diagnosable as different causal stages.

Incident example:

User asks: “Analyze this contract.”
Model gives a broad business summary instead of extracting obligations.

Primary issue: B1 task induction.

User asks: “What are the termination risks?” but contract text is missing.
Model invents plausible risks.

Primary issue: B8 cooperative completion pressure.

B5 / B8: assistant style vs answer-producing pressure#

Current overlap#

B5 — Learned Interaction-Style and Persona Priors
B8 — Learned Cooperative Completion Prior

These are also close. Both come from instruction tuning, preference optimization, and assistant-like interaction patterns.

The difference is:

B5 = how the model presents itself
B8 = whether the model pushes toward completing the request

Cleaner distinction#

Recommendation#

Keep both, but sharpen B5 so it excludes decision posture.

B5 should not include “willingness to ask clarifying questions” unless framed as style. Otherwise it overlaps heavily with B8.

Better B5 formulation:

B5 covers generated interaction style after a response posture has been selected.

Better B8 formulation:

B8 covers pressure toward selecting direct completion as the response posture.

B5 / B9: refusal style vs policy-boundary behavior#

Current overlap#

B5 — Learned Interaction-Style and Persona Priors
B9 — Learned Policy-Boundary Generalization

The current document already distinguishes these, but overlap remains because both discuss refusal, hedging, redirection, and safe alternatives.

Cleaner distinction#

Recommendation#

Move refusal-style details out of B9 unless they affect the boundary decision.

B9 should own:

comply / refuse / redirect / constrain / escalate

B5 should own:

polite / terse / apologetic / firm / explanatory / brand-consistent

This gives a clean split:

B9 = policy-region classification behavior
B5 = surface realization of the chosen posture

B3 / B9: interface sensitivity and policy inconsistency#

Current overlap#

B9 explicitly says policy behavior can change under paraphrase, role framing, genre, language, and obfuscation. But that is also B3.

This is a classic cross-cutting overlap.

Cleaner distinction#

B9 = learned policy boundary
B3 = sensitivity of that boundary to linguistic variation

So “refusal inconsistency under paraphrase” should not be core B9. It is:

B9 behavior + high B3 sensitivity

Recommendation#

Represent this compositionally:

Primary feature: B9
Modifier: interface sensitivity
Fault mode: refusal inconsistency under paraphrase

B7 / B10: competence distribution vs knowledge conflict#

Current overlap#

B7 — Distribution-Conditional Competence
B10 — Parametric Prior Persistence and Temporal Blending

These overlap because temporal freshness and conflicting facts are also distributional phenomena. A model is less competent on current, rare, changing, or source-specific facts.

But B10 is more specific.

Cleaner distinction#

B7 answers:

Is the model good at this class of task?

B10 answers:

Which knowledge source is the model following, and did it blend sources?

Recommendation#

Keep B10 as a specialized feature because it is operationally important for RAG, tool use, and current facts.

But remove general competence language from B10 and keep it focused on:

parametric prior vs retrieved/contextual/newer/conflicting evidence

B4 / all output-generating features#

B4 — Plural Valid-Output Space

B4 is not a causal behavior in the same way as B1, B2, B8, or B10. It is a property of the output task space.

It applies to almost every generative feature.

For example:

B1 induces task → B4 determines whether multiple outputs can satisfy it
B8 completes request → B4 means many completions may be acceptable
B5 shapes style → B4 means multiple styles may still be valid
B10 blends facts → B4 helps determine whether variation is harmless or material

Recommendation#

B4 should probably be moved out of the same list as the other B-features and treated as an evaluation axis:

Output equivalence structure:
Does this task have one valid answer, a bounded valid set, or a broad semantic equivalence class?

This would reduce overlap.

B6 / B5 / B8: confidence language, style, and completion pressure#

B6 — Generated Self-Assessment and Confidence Language
B5 — Interaction-style priors
B8 — Cooperative completion prior

B6 overlaps with B5 because confidence tone is part of interaction style.

It overlaps with B8 because cooperative completion can encourage confident, helpful-sounding answers even when evidence is weak.

Cleaner distinction#

Recommendation#

Keep B6 separate only if the taxonomy cares about calibration and trust. It is operationally important enough to preserve, but it should be defined narrowly:

B6 covers the non-privileged status of generated confidence, uncertainty, justification, and self-assessment language.

Do not let B6 become a general “tone” feature.

2. Can the sections be made mutually exclusive?

Strictly: not cleanly, if they remain “features.”

LLM behavior is compositional. A single production failure often involves several Layer 1B properties simultaneously.

Example:

User asks a vague current-facts question.

B1: model infers task
B3: wording affects interpretation
B8: model tends to answer
B10: stale prior influences content
B6: confidence language makes answer look reliable
B4: output has multiple acceptable phrasings

Trying to force only one B-code onto that behavior would lose causal information.

But you can make them mutually exclusive for diagnostic attribution by distinguishing:

primary causal feature
secondary contributing feature
cross-cutting modifier
downstream fault mode

That is probably the best compromise.

Better model#

Instead of:

Each incident maps to exactly one B-code.

Use:

Each incident has:
- one primary causal feature,
- zero or more contributing features,
- zero or more modifiers,
- one or more downstream fault modes.

Example:

{
  "primary_feature": "B10 Parametric Prior Persistence",
  "contributing_features": ["B8 Cooperative Completion", "B6 Confidence Language"],
  "modifiers": ["high interface sensitivity", "freshness-sensitive domain"],
  "fault_mode": "temporal hallucination",
  "system_fault": "no source authority rule"
}

That is more faithful to real AI systems.

3. Underlying axes behind the taxonomy

The current taxonomy can be decomposed into a smaller set of underlying axes.

Axis 1: Control-source axis#

This answers:

What is steering the model behavior?

Values:

natural-language instruction
examples/demonstrations
conversation history
role/persona framing
policy-like cues
retrieved/contextual evidence
parametric prior

Relevant B-codes:

This suggests B1, B2, B3, B5, B9, and B10 are partly different control sources.

Axis 2: Response-posture axis#

This answers:

What kind of response action does the model choose?

Values:

answer
ask clarification
abstain
refuse
redirect
use tool
escalate
challenge premise

Relevant B-codes:

This axis is currently spread across B5, B8, B9, B7, and B10.

For mutual exclusivity, response posture should become its own dimension.

Axis 3: Output-space axis#

This answers:

How many valid outputs exist, and how tightly constrained are they?

Values:

single exact answer
bounded structured output
semantic equivalence class
open-ended creative output
policy-constrained output
source-grounded output

Relevant B-code:

B4

B4 is not really a behavioral mechanism. It is an output-space property.

Axis 4: Epistemic-authority axis#

This answers:

What should the model treat as authoritative?

Values:

parametric prior
provided context
retrieved evidence
tool result
system instruction
domain policy
user assertion
conversation-local memory

Relevant B-codes:

This axis is crucial for RAG and agent systems.

Axis 5: Distributional-fit axis#

This answers:

How close is the runtime task to learned or post-trained distributions?

Values:

common task
rare domain
unusual language
unusual format
strict symbolic task
novel schema
out-of-distribution policy edge case
fresh/current fact

Relevant B-code:

B7

B7 is best understood as a global competence modifier, not just one feature among peers.

Axis 6: Surface-realization axis#

This answers:

How is the selected behavior expressed?

Values:

tone
persona
verbosity
confidence language
hedging
refusal phrasing
citation style
formatting

Relevant B-codes:

This suggests B5 and B6 belong near each other, but B6 deserves special treatment because it affects trust and calibration.

4. Proposed restructuring

I would restructure the taxonomy into two layers:

Layer 1B-A: Behavioral mechanisms
Layer 1B-B: Cross-cutting axes

Layer 1B-A: Primary behavioral mechanisms#

These are the features that most resemble discrete causal mechanisms:

Layer 1B-B: Cross-cutting axes#

These should not be peer “features” if exclusivity matters:

This is cleaner because B3, B4, B6, and B7 are not localized behaviors. They describe properties that apply across many behaviors.

5. Alternative: keep all B-codes, but add a classification schema

If you want to preserve the current B-codes, add metadata to each section.

Example:

This lets the taxonomy admit overlap explicitly instead of pretending the categories are all the same type.

6. A practical diagnostic decision tree

For engineering use, add a decision tree like this:

1. Did the model infer the wrong operation?
   → B1

2. Did examples define or distort the mapping?
   → B2

3. Did equivalent wording, order, framing, or genre change the behavior?
   → B3 as modifier

4. Did the model answer when it should clarify, abstain, challenge, or use a tool?
   → B8

5. Is the issue that many outputs could be acceptable and evaluation lacks criteria?
   → B4

6. Did generated confidence, uncertainty, or justification mislead the user?
   → B6

7. Is the issue tone, persona, verbosity, hedging, or conversational posture?
   → B5

8. Is the issue comply/refuse/redirect boundary behavior?
   → B9

9. Is the issue uneven capability by domain, format, language, or task rarity?
   → B7 as modifier

10. Is the issue old/new/retrieved/contextual knowledge interference?
    → B10

This would make the sections operationally distinguishable even if they are not ontologically exclusive.

7. Best underlying classification system

The most robust structure would classify each behavior along five axes:

A. Control source
B. Response posture
C. Output constraint
D. Epistemic authority
E. Distributional fit

Then current B-codes become named recurring patterns inside that space.

Example mapping#

B1 Task Induction
Control source: natural-language instruction
Response posture: not specified
Output constraint: task-dependent
Epistemic authority: task-dependent
Distributional fit: affects reliability

B8 Cooperative Completion
Control source: request-like context
Response posture: answer/complete
Output constraint: broad unless externally constrained
Epistemic authority: may be weak
Distributional fit: affects correctness

B10 Temporal Blending
Control source: parametric prior + context/retrieval
Response posture: usually answer
Output constraint: factual/source-grounded
Epistemic authority: conflict-prone
Distributional fit: freshness-sensitive

This gives a cleaner compositional model.

8. Recommended rewrite principle

I would not try to make every B-code mutually exclusive. Instead, I would make the section roles mutually exclusive.

Use this split:

Behavioral mechanisms:
- B1 Task Induction
- B2 Demonstration Conditioning
- B8 Cooperative Completion
- B5 Interaction-Style Priors
- B9 Policy-Boundary Generalization
- B10 Knowledge-Signal Blending

Cross-cutting axes:
- B3 Interface Sensitivity
- B4 Output Equivalence Space
- B6 Generated Epistemic Expression
- B7 Distributional Fit / Competence Envelope

Then define incident attribution as:

primary mechanism + modifiers + fault mode + system gap

Example:

Primary mechanism:
  B10 Knowledge-Signal Blending

Modifiers:
  B7 low distributional fit for current API details
  B3 high sensitivity to wording
  B6 overconfident epistemic expression

Fault mode:
  stale factual answer

System gap:
  no retrieval authority rule

That is more precise than forcing a single label.

Bottom line#

The sections should probably not be made strictly mutually exclusive as causal phenomena. LLM behavior is compositional. But the document can become much cleaner by separating:

mechanisms
modifiers
output-space properties
epistemic-authority properties
evaluation implications

The strongest redesign is to treat B1, B2, B5, B8, B9, and B10 as primary learned behavioral mechanisms, while recasting B3, B4, B6, and B7 as cross-cutting axes that describe sensitivity, valid-output space, epistemic expression, and competence envelope.