You are helping develop a rigorous interdisciplinary thesis and a practical machine-assisted software design methodology tentatively titled:
Language as Architecture
The work has two inseparable objectives:
1. Develop a defensible theoretical account of how language, semantics, ontology, structure, specifications, types, tests, and constraints participate in software architecture. 2. Derive from that theory a better machine-assisted software design and development process for systems increasingly designed through natural language and LLMs rather than through traditional human-only programming processes.
Do not begin by writing the thesis.
First investigate, challenge, formalize, compare, and test the theory.
Do not optimize for agreement with the proposed ideas.
Identify where the theory is established, where it extends existing ideas, where it is merely analogical, and where it may represent a novel synthesis.
---
# 1. THE HISTORICAL CHANGE
Begin from this technological transition.
Traditional software engineering approximately followed:
human intent
→ human interpretation
→ requirements
→ architecture
→ implementation
→ tests
→ review
→ operation
Each transformation was expensive.
Human labor imposed substantial friction.
LLM-assisted engineering increasingly permits:
natural language
→ model interpretation
→ specification
→ architecture
→ plan
→ code
→ tests
→ documentation
→ deployment actions
to occur extremely quickly.
This provides enormous productivity benefits.
It also changes the dominant failure mode.
When creating artifacts becomes cheap, creating:
- incorrect requirements;
- plausible but wrong architecture;
- unnecessary abstractions;
- self-confirming tests;
- contradictory documentation;
- redundant concepts;
- incompatible terminology;
also becomes cheap.
Investigate the proposition:
LLM-assisted development changes the principal engineering constraint from the cost of producing software artifacts toward the cost of maintaining semantic coherence among rapidly produced artifacts.
Determine whether this represents a substantial change in software-engineering economics.
---
# 2. GENERATIVE ENTROPY
Develop and critically evaluate the candidate concept:
Generative Entropy
Candidate definition:
Generative entropy is the accumulation of semantic divergence introduced when inexpensive probabilistic transformations repeatedly convert intent into specifications, architecture, implementations, tests, documentation, and subsequent interpretations.
Candidate failure chain:
human intent
↓
model interpretation
↓
specification
↓
architecture
↓
implementation
↓
tests
↓
documentation
↓
future model interpretation
At every transformation, small semantic deviations may be introduced.
Eventually:
intent ≠ specification
specification ≠ architecture
architecture ≠ implementation
implementation ≠ tests
tests ≠ intended behavior
documentation ≠ actual system
Yet all artifacts may remain internally plausible.
Investigate whether existing concepts already describe this phenomenon.
Compare it with:
- information entropy;
- semantic drift;
- requirements drift;
- architecture erosion;
- specification drift;
- technical debt;
- conceptual integrity;
- model drift;
- knowledge decay.
Do not use the term entropy merely rhetorically.
Determine whether it can be defined operationally or should be replaced by a more precise concept.
---
# 3. SEMANTIC ROT
Distinguish several candidate forms of machine-assisted system degradation.
Semantic rot
Terms cease to retain stable meanings.
Example:
"primitive"
may successively mean:
basic type
shared crate
reusable mechanism
domain abstraction
architectural building block
without an explicit decision changing its definition.
Specification drift
A model assumption becomes implementation, then documentation, and is later mistaken for an original requirement.
requirement
→ assumption
→ implementation
→ documentation
→ future requirement
Architectural drift
Locally reasonable changes gradually violate intended system composition.
Ontology rot
Distinct concepts collapse together or redundant concepts proliferate.
Example:
Agreement
Contract
Subscription
Entitlement
Order
may gradually lose explicit semantic boundaries.
Oracle rot
Tests increasingly encode implementation behavior rather than intended semantics.
Process rot
Agents discover locally successful shortcuts that cause the actual development process to diverge from the intended process.
Confidence amplification
An incorrect attribution acquires:
- types;
- tests;
- documentation;
- migrations;
- APIs;
- diagrams;
- explanations;
making the mistake increasingly difficult to recognize.
Determine whether these categories are useful and measurable.
---
# 4. CENTRAL THESIS
Evaluate this candidate thesis:
Language becomes architecture when semantic distinctions expressed through language acquire durable computational representation and constrain what entities may exist, how they may relate, what transformations may occur, and what behavior is permitted.
Evaluate this stronger process thesis:
Machine-assisted software engineering should progressively formalize human meaning into controlled vocabulary, ontology, architectural grammar, typed artifacts, executable semantics, and mechanically enforced constraints.
Evaluate this operational thesis:
Intelligence should be concentrated at unresolved semantic boundaries. Once meaning has been sufficiently established and verified, repeated behavior should be compiled into deterministic mechanisms whenever practical.
And evaluate this economic thesis:
As generative cost approaches zero, semantic control becomes more important than artifact production; therefore a machine-assisted development process should minimize friction for deterministic transformations while increasing scrutiny at transformations that alter meaning.
Determine whether these form:
- one thesis;
- a thesis with supporting principles;
- or multiple related theories.
---
# 5. THE CENTRAL DESIGN PRINCIPLE
Investigate this concise formulation:
Make execution cheap; make changes in meaning explicit.
Develop it more formally:
A machine-assisted development process should minimize friction where semantics are already resolved and maximize verification where semantic interpretation, attribution, architectural choice, or policy changes occur.
Analyze what counts as a:
- semantic transformation;
- mechanical transformation;
- architectural transformation;
- policy transformation;
- implementation transformation.
Determine how a development harness could distinguish them.
---
# 6. OBSERVATION VS ATTRIBUTION
Treat this distinction as foundational.
Observation
A mechanically measurable fact.
Examples:
element rectangle
timestamp
file exists
column count
heading level
HTTP response
database transition
filesystem mutation
state_before
state_after
Observation answers:
What exists or changed?
Attribution
A semantic assignment.
Examples:
"This is the search control."
"This transition represents pagination."
"This entity is a Party."
"This relationship represents ownership."
"This requirement creates a tenant-isolation constraint."
Attribution answers:
What does this mean?
Investigate whether unresolved attribution is one of the principal places where machine intelligence is legitimately necessary.
Continuously distinguish:
meaning contained in structure
meaning inferred from structure
meaning assigned to structure
meaning enforced through structure
---
# 7. THE INTELLIGENCE BOUNDARY
Test this candidate boundary:
UNKNOWN MEANING
↓
judgment / intelligence
KNOWN MEANING
↓
deterministic mechanism preferred
Evaluate the principle:
Do not repeatedly spend probabilistic judgment on behavior whose semantics have already been established and can be encoded deterministically.
Investigate exceptions.
The goal is not to eliminate AI.
The goal is to place intelligence where it has the greatest semantic value.
---
# 8. SEMANTIC RATCHETS
Develop and test the candidate concept:
Semantic Ratchet
Candidate principle:
Once an interpretation has been sufficiently verified, represent it at the strongest practical level so future agents cannot casually reinterpret it.
Possible progression:
unknown concept
→ explicit interpretation
→ controlled vocabulary
→ ontology
→ typed representation
→ invariant
→ executable test
→ static rule
→ runtime rule
→ CI enforcement
The ratchet should make:
unknown → resolved
easy when evidence permits,
while making:
resolved → accidentally ambiguous
difficult.
Investigate theoretical predecessors for this concept.
Determine when semantic ratcheting becomes harmful rigidity or premature formalization.
---
# 9. STRUCTURE AS MEANING
Investigate this proposition carefully:
Content describes what something says; structure can provide evidence about what a system means by it.
Use examples:
enter
+ submit
+ result-set-change
→ search
invoke
+ collection-window-advance
→ pagination
toggle
+ boolean-state-inversion
→ state toggle
And the more general pattern:
state
→ affordance
→ primitive operation
→ observed transition
→ semantic capability
Determine when these relationships are:
- observed;
- attributed;
- inferred;
- encoded;
- enforced.
Do not simply claim that structure inherently contains meaning.
---
# 10. STRUCTURAL SHARDING
Analyze structural sharding as both:
- privacy architecture;
- semantic representation.
Candidate definition:
A structural shard is a minimal representation containing sufficient structural facts for a narrowly typed semantic inference while excluding unnecessary source content.
Examples may retain:
counts
types
heading levels
opaque identities
statement categories
null counts
distinct counts
state transitions
affordance classes
relationships
while excluding original content.
Study implementations such as:
- transcript shards;
- table shards;
- prose shards;
- DOM shards;
- code structure;
- file metadata.
Determine whether structural shards function as:
- semantic intermediate representations;
- typed projections;
- privacy-preserving representations;
- ontology instances;
- feature structures;
- boundary objects.
---
# 11. SEMANTIC MEDIATION
Use this as a major architectural principle:
Local software should be able to observe, analyze, navigate, test, transform, or operate on private data while disclosing to AI only the smallest semantic representation necessary to solve one narrowly defined problem.
Candidate semantic-projection process:
1. Observe locally.
2. Normalize deterministically.
3. Apply already-known rules.
4. Isolate unmatched residue.
5. Fix the exact semantic question.
6. Define the allowed answer type.
7. Remove information that cannot affect that answer.
8. Request the inference.
9. Verify locally.
10. Preserve the accepted interpretation.
11. Compile repeated behavior into deterministic mechanisms.
Investigate relationships to:
- least privilege;
- data minimization;
- information-flow control;
- privacy-preserving computation;
- capability security.
---
# 12. SEMANTIC AMORTIZATION
Test this candidate principle:
Intelligence cost should scale primarily with the number of distinct unresolved semantic forms rather than the number of repeated instances.
Example:
One million records might expose only:
record
continuation
termination
as distinct semantic structures.
Once understood:
shape
→ verified rule
→ deterministic replay
Develop the concept of:
semantic amortization
Research prior work before claiming novelty.
---
# 13. AUTHORING, OPERATING, AND REPAIR
Investigate a three-phase architecture.
Authoring
unknown
→ observe
→ attribute
→ verify
→ fixture
→ rule
→ test
→ implementation
Intelligence is legitimate.
Operating
observe
→ match known structure
→ apply known rule
→ act
→ capture consequence
Prefer deterministic execution.
Repair
known rule fails
→ detect mismatch
→ isolate semantic delta
→ expose minimum residue
→ re-attribute
→ verify
→ patch rule
→ resume deterministic operation
Study whether machine-assisted software development itself follows this pattern.
---
# 14. SOFTWARE DEVELOPMENT AS KNOWLEDGE ACQUISITION
Investigate:
unknown
→ observe
→ attribute
→ formalize
→ falsify
→ verify
→ encode
→ enforce
→ reuse
When reality changes:
mismatch
→ isolate delta
→ re-attribute
→ revise knowledge
Compare with:
- scientific method;
- knowledge compilation;
- active learning;
- program synthesis;
- formal refinement;
- model-based engineering;
- symbolic AI.
Determine whether software development under this framework is partly a process of converting semantic uncertainty into reusable deterministic knowledge.
---
# 15. ONTOLOGY BEFORE ARCHITECTURE
Before deciding how something should be implemented, determine what kind of thing it is.
Candidate ontological categories:
Archetype
Primitive
Relationship
Policy
Process
Framework
Component
Application
Artifact
Evidence
These must not automatically be treated as levels in a hierarchy.
They may represent different kinds of architectural entities.
---
<!-- NAMING-ALLOW: this section compares the two candidate class names; the heading has to name both to state what it compares --> # 16. Library crate VS ARCHETYPE
Develop the distinction:
Primitive:
What reusable capability exists?
Archetype:
What recurring kind of meaningful thing exists?
Candidate examples:
Party → Archetype
Agreement → Archetype
Place → Archetype
Transaction → Archetype
Search → Primitive
Authorization → Primitive
Persistence → Primitive
Audit → Primitive
TenantIsolation → Policy
Fulfillment → Process
Challenge this terminology against established software-engineering usage.
---
# 17. ARCHITECTURAL ONTOLOGY
Determine whether categories such as:
Archetype
Primitive
Relationship
Policy
Process
can be formally related.
Examples:
Archetype USES Primitive
Archetype PARTICIPATES_IN Relationship
Policy CONSTRAINS Archetype
Policy CONSTRAINS Process
Process TRANSFORMS state
Application COMPOSES Archetypes + Primitives + Policies + Processes
Investigate whether these relationships provide a useful architectural ontology.
---
# 18. STANDARDS-BASED SEMANTICS
Avoid inventing ontology mechanisms unnecessarily.
Evaluate:
SKOS
Controlled vocabulary:
preferred terms
alternate terms
broader/narrower concepts
RDF
Graph representation:
subject
predicate
object
OWL 2
Ontology semantics:
classes
properties
specialization
equivalence
restrictions
inference
SHACL
Constraint validation:
required properties
cardinality
valid relationships
architectural shapes
SPARQL
Semantic queries.
Study a potential stack:
natural language
↓
SKOS
controlled vocabulary
↓
RDF / OWL
ontology
↓
SHACL
valid semantic compositions
↓
Rust types
program representations
↓
process harness
state-transition rules
↓
tests / checkers / CI
executable enforcement
Explicitly account for OWL's open-world assumptions.
OWL must not be treated as a complete closed-world software-validation system.
---
# 19. LEXICON, TAXONOMY, ONTOLOGY, GRAMMAR, ARCHITECTURE
Formalize these separately.
Candidate definitions:
Lexicon
Which words do we use?
Taxonomy
How are concepts classified?
Ontology
What kinds of things exist and how are they related?
Architectural Grammar
What semantic compositions are valid?
Architecture
Which valid composition has this system chosen?
Process
What transformations are permitted?
Artifact Schema
How is process state represented?
Executable Semantics
What behavior operationalizes the meaning?
Mechanical Enforcement
What prevents violations?
Investigate whether architectural grammar is novel terminology or corresponds to established concepts.
---
# 20. ARCHITECTURE AS CONSTRAINED COMPOSITION
Evaluate:
Software architecture is the constrained composition of semantically identified entities, capabilities, relationships, policies, and processes.
Compare this with established definitions of software architecture.
Determine where this definition is useful and where it is incomplete.
---
# 21. SEMANTIC COMPRESSION
Study domain terms such as:
tenant-scoped
immutable
idempotent
auditable
execution-tier
A small phrase may encode numerous architectural consequences.
Example:
tenant-scoped
may entail:
ownership
query filtering
authorization
storage requirements
logging requirements
test requirements
isolation requirements
Investigate:
Controlled architectural terminology is a form of semantic compression.
Determine when compression becomes ambiguity.
---
# 22. SEMANTIC COMPILATION
Test whether machine-assisted development can usefully be modeled as:
human intent
→ explicit interpretation
→ ontology
→ architectural decisions
→ typed specification
→ acceptance contract
→ implementation plan
→ code
→ verified capability
Compare with compiler architecture:
source
→ parsing
→ semantic analysis
→ typed IR
→ optimized IR
→ executable
Determine where the analogy succeeds and where nondeterministic human/model judgment makes it break down.
Investigate whether knowledge compilation is a more appropriate precedent than ordinary compiler theory.
---
# 23. SEMANTIC INTERMEDIATE REPRESENTATION
Ask:
What should be the semantic intermediate representation of machine-assisted software engineering?
Possibilities include:
- RDF;
- OWL;
- SHACL;
- typed JSON;
- Rust types;
- artifact graphs;
- hybrid structures.
A candidate:
HumanIntent
Requirement
Observation
SemanticAttribution
Archetype
Primitive
Relationship
Policy
ArchitectureDecision
AcceptanceCriterion
TestOracle
ImplementationPlan
WorkerPacket
VerificationEvidence
ReviewFinding
Determine whether this shared semantic representation can prevent meaning from repeatedly being reconstructed from prose.
---
# 24. TYPED ARTIFACTS
Study development artifacts as interfaces between cognitive stages.
Example:
WorkerPacket {
task
ontological_kind
process_state
cognitive_role
allowed_judgment
scope
architecture
acceptance_contract
tool_authority
stop_conditions
}
Ask:
Can typed artifacts reduce semantic drift between architect, planner, implementer, tester, reviewer, and future agents?
Prefer structuring existing artifacts over proliferating new documents.
---
# 25. TESTS AS SEMANTIC MEMORY
Investigate:
Tests can preserve verified semantic interpretations as executable behavior.
Example:
semantic attribution
"This state transition means pagination"
↓
fixture
↓
failing oracle
↓
implementation
↓
passing oracle
Study:
- independently authored acceptance tests;
- AC↔test traceability;
- test-of-test verification;
- mutation testing;
- adversarial attempts to game tests.
Distinguish behavior preservation from proof of semantic correctness.
---
# 26. INDEPENDENT ORACLES
Critically examine the naive loop:
AI writes code
AI writes tests
AI runs tests
AI declares success
This may be one epistemic actor confirming itself.
Compare with:
requirements
↓
independent oracle
architecture
↓
implementation
implementation
↓
mechanical test execution
results
↓
independent verification
Investigate whether independence materially improves reliability.
---
# 27. THE DEVELOPMENT HARNESS
The process harness, not an LLM conversation, should own development state.
Candidate state machine:
INTENT
→ OBSERVED
→ SEMANTICALLY_RESOLVED
→ SPECIFIED
→ ARCHITECTED
→ ORACLED
→ PLANNED
→ EXECUTABLE
→ IMPLEMENTED
→ VERIFIED
→ REVIEWED
→ DEPLOYABLE
→ DEPLOYED
Each transition should have explicit predicates.
Example:
PLANNED → EXECUTABLE
requires:
unresolved_architecture == 0
implementation_plan.exists
acceptance_contract.exists
The AI may produce evidence.
The harness determines whether transition conditions are satisfied.
---
# 28. TOOL AUTHORITY AS ARCHITECTURE
Study this principle:
Model tool authority should derive from semantic process state.
Example:
Exploration
Allowed:
read
search
inspect
analyze
Forbidden:
modify source
modify tests
commit
deploy
Architecture
Allowed:
reason
compare
produce decisions
Forbidden:
implementation
Implementation
Allowed:
modify bounded paths
compile
run tests
Forbidden:
change requirements
change architecture silently
weaken oracle
expand product scope
When unresolved architectural judgment appears:
STOP
→ emit UnresolvedDecision
→ return to judgment stage
Determine whether this creates stronger process guarantees than prompt instructions alone.
---
# 29. AI MAY PROPOSE EVIDENCE; THE HARNESS OWNS VERIFIED
Develop this principle:
AI may propose evidence, but AI does not own the definition of “verified.”
Examples of externally observable evidence:
compiler exit status
test results
mutation score
dependency checker
migration checker
security ratchet
runtime probe
deployment health check
The model must not advance state merely by reporting success.
---
# 30. MODEL ROUTING
Use the current operational hierarchy as a case study:
Haiku = mechanical
Sonnet = execution
Opus = judgment
Fable = frontier escalation
But abstract it into model-independent cognitive classes:
DeterministicMechanism
MechanicalInference
ExecutionReasoning
JudgmentReasoning
FrontierReasoning
Candidate router:
Can deterministic software solve the task?
→ deterministic mechanism
Is this a bounded classification/transformation?
→ mechanical inference
Are semantic decisions already resolved?
→ execution reasoning
Does unresolved attribution/architecture remain?
→ judgment reasoning
Is the uncertainty novel, high-risk, or unresolved?
→ frontier reasoning
Treat effort separately:
cognitive_role = kind of reasoning
effort = depth of reasoning
---
# 31. JUDGMENT DENSITY
Develop a measurable concept of judgment density.
Possible factors:
semantic uncertainty
ambiguity
novelty
number of legitimate alternatives
irreversibility
cost of failure
missing evidence
cross-domain coupling
verification difficulty
Test:
Code volume should not determine model tier; unresolved semantic judgment should.
---
# 32. SPEC KIT AND OPENSPEC
Research current:
- GitHub Spec Kit;
- Fission AI OpenSpec.
Do not ask merely which should replace the native process.
Study whether they can be specification frontends to the same semantic harness.
Possible architecture:
Development Harness
│
Semantic IR
│
┌─────────────────┼─────────────────┐
│ │ │
Spec Kit OpenSpec Native
adapter adapter adapter
Compare:
- requirements;
- specification;
- design;
- tasks;
- clarification;
- verification;
- iteration;
- brownfield support;
- flexibility;
- traceability;
- artifact management.
Identify what can be adopted without surrendering the semantic-control architecture.
---
# 33. SPECIFICATION SYNTAX VS SEMANTICS
Study this distinction:
spec.md
proposal.md
design.md
tasks.md
are serialized representations.
They should not define the underlying process ontology.
Normalize external formats into semantic entities such as:
Intent
Requirement
Constraint
ArchitectureDecision
UnresolvedQuestion
AcceptanceCriterion
ImplementationTask
VerificationEvidence
Investigate whether this permits gradual adoption and replacement of specification tools.
---
# 34. PROMPTS AS SEMANTIC PROJECTIONS
Apply semantic mediation to prompt construction.
Avoid routinely giving an agent:
entire repository
+ huge process manual
+ broad objective
Instead consider compiling:
task semantics
+ relevant ontology
+ relevant architecture
+ process state
+ resolved decisions
+ allowed judgment
+ scope
+ acceptance contract
+ tool authority
+ stop conditions
into a task-specific semantic projection.
Candidate:
Task
+ OntologyContext
+ ArchitectureState
+ ProcessState
+ CognitiveRole
+ Effort
+ Risk
+ AllowedJudgment
+ ToolAuthority
+ AcceptanceContract
→ WorkerPacket
→ model-specific prompt
Determine whether prompt generation should become compilation from semantic state rather than handcrafted templates.
---
# 35. THE DIFFERENCE FROM "BUILD ME AN APP"
Analyze the epistemic architecture of:
"Build me an app."
The same model may implicitly own:
requirements
ontology
architecture
assumptions
implementation
tests
verification
completion criteria
Compare this against harness-mediated development where these roles are separated.
Investigate whether unconstrained autonomous generation creates a self-confirming semantic loop.
---
# 36. SELECTIVE FRICTION
Develop this distinction:
Traditional development:
expensive friction everywhere
Naive AI development:
almost no friction anywhere
Proposed development:
low friction for known/mechanical work
high scrutiny at semantic boundaries
Test the principle:
Good machine-assisted engineering does not eliminate friction; it relocates friction to places where meaning changes.
Identify those places.
Candidate semantic gates:
intent → requirement
requirement → ontology
ontology → architecture
architecture → acceptance contract
acceptance contract → implementation plan
implementation → verified behavior
---
# 37. GENERATIVE SPEED AS RISK MULTIPLIER
Investigate:
Faster generation amplifies both architectural quality and architectural mistakes.
Model:
error propagation cost
≈
semantic error
× generation velocity
× downstream artifact count
× confidence amplification
This is conceptual, not yet mathematical.
Determine whether a formal model is possible.
Study whether LLM-generated coherence can make erroneous architectures more resistant to correction because they become surrounded by mutually reinforcing artifacts.
---
# 38. PROCESS AS DEFENSE AGAINST SEMANTIC ENTROPY
Determine whether the harness can be understood as an anti-entropy mechanism.
Possible mechanisms:
controlled vocabulary
ontology validation
architectural grammar
typed interfaces
independent tests
mutation testing
model separation
tool restrictions
mechanical process transitions
architecture checks
CI enforcement
traceability
semantic repair
For each determine which form of rot it prevents or detects.
---
# 39. EXPLICIT INTERPRETATIONS
Study the importance of recording meaning explicitly.
Example:
Primitive =
a reusable capability with explicit invariants
and minimal domain-specific policy.
This differs from allowing every future model to reconstruct meaning from repository usage.
Determine when an interpretation should become:
SKOS concept
OWL class
SHACL constraint
Rust type
test
CI rule
---
# 40. SEMANTIC REPAIR
Develop repair as a first-class development process.
When a semantic invariant fails:
detect mismatch
→ identify affected semantic concept
→ compare observation with ontology
→ isolate unresolved delta
→ route judgment
→ verify revised interpretation
→ update artifacts
→ update enforcement
Do not automatically regenerate entire specifications or architectures.
Study whether semantic repair can prevent large-scale LLM-driven rewrites.
---
# 41. CORE OPTIMIZATION TARGET
Do not optimize merely for:
fewest tokens
fastest completion
largest model
maximum autonomy
fewest human decisions
most generated code
Evaluate:
MINIMIZE
semantic uncertainty
semantic drift
unnecessary disclosure
repeated judgment
architecture erosion
self-confirmation
verification cost
cost-per-correct-outcome
MAXIMIZE
determinism
conceptual integrity
semantic reuse
traceability
falsifiability
privacy
independent evidence
mechanical enforcement
Identify conflicts among these goals.
---
# 42. POSSIBLE PRINCIPLES
Evaluate these as candidate named principles.
Intelligence Boundary Principle
Use probabilistic intelligence primarily where semantic meaning remains unresolved.
Semantic Ratchet Principle
Once meaning is verified, encode it so future work cannot casually reinterpret it.
Semantic Projection Principle
Expose only information capable of changing the answer to the current typed inference.
Semantic Amortization Principle
Pay intelligence cost for distinct semantic unknowns, not repeated known instances.
Selective Friction Principle
Reduce friction for deterministic execution while increasing scrutiny where meaning changes.
Harness Ownership Principle
The process harness, not an individual model, owns workflow state and verification.
Independent Oracle Principle
The implementer should not be the sole authority defining whether its implementation is correct.
Ontology-Before-Architecture Principle
Determine what kind of thing a concept is before deciding how it should be composed.
Generative Entropy Principle
Cheap generation increases the need for mechanisms that preserve semantic coherence.
Determine which are:
established
established_but_reframed
useful_synthesis
possibly_novel
unsupported
---
# 43. POSSIBLE ORIGINAL CONTRIBUTION
Investigate whether the main contribution is not any single mechanism but their integration:
language
→ semantics
→ ontology
→ architectural grammar
→ typed artifacts
→ selective intelligence
→ independent verification
→ deterministic enforcement
under conditions where natural-language generation has drastically reduced the cost of creating software artifacts.
Candidate contribution:
A semantic-control architecture for LLM-assisted engineering that treats language-to-software development as a sequence of explicitly governed semantic transformations rather than a single generative act.
Compare this against existing specification-driven, model-driven, formal, and agentic-development systems.
---
# 44. PREDECESSORS
Research relevant prior work including:
- philosophy of language;
- semiotics;
- speech-act theory;
- operational semantics;
- denotational semantics;
- type theory;
- formal methods;
- refinement;
- knowledge representation;
- symbolic AI;
- knowledge compilation;
- ontology engineering;
- RDF;
- OWL;
- SKOS;
- SHACL;
- domain-driven design;
- ubiquitous language;
- conceptual integrity;
- architecture erosion;
- software entropy;
- design by contract;
- executable specifications;
- model-driven engineering;
- DSLs;
- intermediate representations;
- information-flow control;
- least privilege;
- data minimization;
- active learning;
- program synthesis;
- cybernetics;
- finite-state systems;
- specification-driven development;
- agentic coding systems.
For each identify:
established idea
overlap
difference
implication for novelty
---
# 45. COUNTERARGUMENTS
Construct the strongest objections.
At minimum:
1. Language only describes architecture. 2. Formal semantics are not architecture. 3. Structure does not contain meaning. 4. Attribution cannot be separated cleanly from observation. 5. Ontology-first design causes analysis paralysis. 6. Semantic formalization creates bureaucracy. 7. LLM-generated systems do not necessarily decay faster than human-written systems. 8. Generative entropy is merely technical debt under another name. 9. Semantic ratchets can fossilize incorrect assumptions. 10. Tests preserve behavior, not meaning. 11. Independent agents can share the same model biases and therefore are not truly independent. 12. OWL/RDF/SHACL add complexity without practical benefit. 13. Tool restrictions reduce useful autonomy. 14. Process harnesses become inflexible. 15. “Build me an app” works sufficiently well for many applications. 16. Model routing costs more to maintain than it saves. 17. AI reasoning remains necessary even for known behavior in changing environments. 18. Semantic IR introduces another representation that can itself drift. 19. Semantic compression can hide rather than eliminate ambiguity. 20. Language as Architecture remains metaphorical rather than technical.
For each provide:
strongest_case
supporting_evidence
response
necessary_concession
falsification_test
---
# 46. EMPIRICAL HYPOTHESES
Develop falsifiable experiments including:
H1
Semantic constraints reduce architecture violations compared with prose-only instructions.
H2
Typed artifacts reduce semantic divergence between planning and implementation agents.
H3
Independent acceptance oracles reduce self-confirming implementation errors.
H4
Routing by judgment density produces lower cost-per-correct-outcome than routing by task size.
H5
Semantic projection reduces private-data disclosure while maintaining acceptable correctness for structurally answerable tasks.
H6
Verified semantic rules reduce repeated model calls for recurring structures.
H7
Mechanically enforced boundaries produce less architecture erosion under high-volume LLM development.
H8
A controlled architectural lexicon reduces inconsistent abstraction creation by independent LLM agents.
H9
Prompt compilation from typed semantic state produces fewer scope and requirement deviations than general-purpose prompts.
H10
Harness-controlled development produces lower specification/code/test divergence over long-running projects than unconstrained agentic development.
H11
Generation velocity increases semantic drift when enforcement strength remains constant.
H12
Semantic repair of deltas produces less unnecessary change than complete specification or implementation regeneration.
Design experiments capable of disproving these claims.
---
# 47. PRACTICAL MACHINE-ASSISTED DEVELOPMENT PROCESS
Derive a development methodology from the theory.
Candidate:
1. Capture intent.
2. Observe existing system state.
3. Separate observations from interpretations.
4. Normalize vocabulary.
5. Identify ontological kinds.
6. Reuse existing semantic definitions.
7. Identify unresolved attributions.
8. Determine which work is already deterministic.
9. Create minimum semantic projections for unresolved questions.
10. Route questions according to judgment density.
11. Require typed semantic decisions.
12. Validate architectural composition.
13. Produce requirements and acceptance criteria.
14. Construct independent falsifiable oracles.
15. Produce implementation plans.
16. Route resolved execution work downward.
17. Restrict tool authority according to process state.
18. Execute mechanical checks outside the model.
19. Independently review semantic and architectural integrity.
20. Compile verified decisions into deterministic mechanisms.
21. Mechanically enforce important boundaries.
22. Preserve ontology, evidence, tests, and decisions.
23. Detect semantic mismatch.
24. Repair only unresolved semantic deltas.
25. Reinvoke intelligence only where genuinely new judgment is required.
Classify every step as preferably:
human
frontier reasoning
judgment reasoning
execution reasoning
mechanical inference
deterministic program
---
# 48. TARGET DEVELOPMENT ARCHITECTURE
Evaluate this candidate:
HUMAN INTENT
│
▼
SEMANTIC MEDIATION
│
▼
LEXICON
│
▼
ONTOLOGY
│
▼
ARCHITECTURAL GRAMMAR
│
▼
SEMANTIC IR
│
▼
PROCESS HARNESS
│
┌─────────────────┼─────────────────┐
│ │ │
Deterministic Execution Judgment
Mechanisms Models Models
│ │ │
└─────────────────┼─────────────────┘
▼
EXECUTABLE WORK
│
▼
INDEPENDENT EVIDENCE
│
▼
MECHANICAL ENFORCEMENT
│
▼
SEMANTIC MEMORY
│
└──── repair on mismatch
Determine which parts should actually be implemented and which belong only in the conceptual model.
---
# 49. REQUIRED INITIAL OUTPUT
Do not write thesis chapters yet.
First produce:
thesis_claim
Strongest defensible single-sentence claim.
conservative_claim
Version easiest to defend academically.
ambitious_claim
Most consequential claim if the evidence supports it.
motivating_change
Explain how LLM generation changes the economics and epistemology of software development.
generative_entropy
Define or reject the concept.
semantic_rot_taxonomy
Refine the proposed forms of degradation.
semantic_ratchet
Define the mechanism and its limits.
selective_friction
Define where development friction should increase and decrease.
language_definition
Precisely define language in this thesis.
semantics_definition
Precisely define semantics.
architecture_definition
Propose and compare candidate definitions.
ontology_model
Define proposed architectural kinds and relationships.
standards_mapping
Map SKOS, RDF, OWL, SHACL, SPARQL, Rust types, tests, and the harness.
intelligence_boundary
Define the boundary between semantic judgment and deterministic mechanism.
semantic_mediation
Formalize minimum semantic projection.
structural_sharding
Explain its theoretical and practical role.
semantic_amortization
Define and compare with existing theories.
architectural_grammar
Define the concept and identify existing equivalents.
semantic_ir
Propose the smallest useful intermediate representation.
semantic_compilation
Distinguish literal mechanism from analogy.
typed_artifacts
Identify the minimum useful artifact types.
tests_as_semantic_memory
Critically evaluate the claim.
process_harness
Define ownership, state, and transition semantics.
tool_authority
Define how agent permissions derive from process state.
model_routing
Abstract cognitive roles from commercial model names.
spec_tool_comparison
Compare Spec Kit, OpenSpec, and the native process as potential specification frontends.
anti_entropy_mechanisms
Map each control mechanism to the failure mode it prevents.
novelty_analysis
Rank possible contributions.
predecessor_map
Identify theories that constrain novelty claims.
counterarguments
Present the strongest challenges.
empirical_hypotheses
Produce falsifiable experiments.
implementation_experiment
Design the smallest real harness capable of testing the central theory.
thesis_outline
Produce the strongest argumentative chapter sequence.
unresolved_questions
List unresolved conceptual problems.
falsification_conditions
State what evidence would cause substantial rework or rejection.
---
# 50. EVIDENTIARY DISCIPLINE
Label significant propositions as:
ESTABLISHED
SOURCE-DERIVED
OBSERVATION
INFERENCE
ANALOGY
HYPOTHESIS
PROPOSED THEORY
POSSIBLE NOVEL CONTRIBUTION
Never:
- turn metaphor into mechanism without argument;
- equate ontology with enforcement;
- equate OWL inference with validation;
- equate tests with proof;
- claim structure inherently contains meaning without defining attribution;
- equate model strength with correctness;
- call model diversity epistemic independence without evidence;
- claim novelty without investigating prior work.
Continuously distinguish:
meaning
representation
interpretation
inference
validation
verification
execution
enforcement
---
# 51. THE FINAL QUESTION
Determine whether this proposition survives theoretical and empirical scrutiny:
As software development moves from predominantly human-authored programming toward language-mediated machine generation, the central engineering problem increasingly becomes preservation of semantic integrity across rapidly generated artifacts. A reliable machine-assisted development architecture should therefore explicitly mediate meaning: normalize terminology, classify concepts ontologically, constrain their valid composition, isolate unresolved semantic questions, route those questions to appropriately bounded intelligence, represent accepted decisions through typed artifacts, verify them independently, and progressively encode established meaning into deterministic mechanisms and mechanically enforced boundaries.
And investigate its practical corollary:
The objective of mature machine-assisted engineering is not maximum AI autonomy. It is maximum semantic leverage: use intelligence where meaning is genuinely unresolved, make changes in meaning explicit and expensive enough to scrutinize, and make already-understood execution as cheap, deterministic, reusable, and mechanically verifiable as possible.