LLM Research

Table of Contents

1. PART I — FOUNDATIONS: WHAT OBJECT ARE WE STUDYING?
   1.1 Artificial intelligence, machine learning, deep learning, transformers, and LLMs
   1.2 Model versus inference process versus deployed system
   1.3 LLM versus agent versus research system
   1.4 Model weights versus runtime state
   1.5 Token sequence versus latent computation
   1.6 Output versus process that generated the output
   1.7 Representation versus represented object
   1.8 Readout versus mechanism
   1.9 Local behavior versus global capability
   1.10 Capability versus intelligence
   1.11 Intelligence versus learning
   1.12 Memory versus history
   1.13 Knowledge versus accessibility
   1.14 Generation versus discrimination
   1.15 Certification versus truth
   1.16 The central non-collapse laws
   1.16.1 SEMANTIC_CLOUD ≠ INTELLIGENCE
   1.16.2 SEMANTIC_CLOUD ≠ MEMORY
   1.16.3 SEMANTIC_CLOUD ≠ LEARNING
   1.16.4 FLUENCY ≠ SUPPORT
   1.16.5 DENSITY ≠ TRUTH
   1.16.6 CONSENSUS ≠ WARRANT
   1.16.7 PROJECTION ≠ SOURCE
   1.16.8 REPRESENTATION ≠ OBJECT
   1.16.9 READOUT ≠ MECHANISM
   1.16.10 LOCAL ≠ GLOBAL
   1.17 Epistemic status of the semantic-cloud model
   1.18 Semantic cloud as a working research ontology, not an established unique native ontology of LLMs

2. PART II — TOKEN SPACE AND TRAINING EXPOSURE
   2.1 World, text, code, mathematics, science, and culture as training sources
   2.2 Tokenization as discrete carrier formation
   2.3 Vocabulary and segmentation
   2.4 Sequence structure
   2.5 The enormous combinatorial possibility volume
   2.6 Sparsity of actual training exposure
   2.7 Training-distribution topology
   2.8 Dense exposure regions
   2.9 Sparse and absent regions
   2.10 Boundary regions
   2.11 Training contamination and hidden exposure
   2.12 Measuring exposure rather than assuming knowledge
   2.13 Curriculum-restricted models as experimental instruments
   2.14 LittleLearner / controlled-exposure experiments
   2.15 Behavioral frontier versus true semantic frontier
   2.16 Human curricula versus model-native acquisition order
   2.17 Scaling inside an exposed semantic basis
   2.18 Adjacent recombination versus genuinely remote capability
   2.19 In-context learning as accessibility modification
   2.20 Post-training as reweighting/reorganization versus creation of genuinely new structure
   2.21 Experimental question: what operations can actually enlarge the reachable semantic basis?

   The controlled-exposure material is especially important here because it separates manipulation of an existing accessible basis from demonstrated expansion into deliberately excluded regions.

3. PART III — FORMATION OF THE SEMANTIC CLOUD
   3.1 Architecture × data × objective × optimizer
   3.2 Repeated parameter updates
   3.3 Recurrence and statistical regularity
   3.4 Correlation extraction
   3.5 Compression
   3.6 Abstraction
   3.7 Invariant detachment
   3.8 Carrier transfer
   3.9 Relation formation
   3.10 Transformation formation
   3.11 Generator formation
   3.12 From isolated examples to reusable structure
   3.13 Persistent potential geometry
   3.14 Semantic neighborhoods
   3.15 Basins
   3.16 Boundaries
   3.17 Filaments
   3.18 Sheets
   3.19 Nodes / attractors
   3.20 Voids and weakly reachable regions
   3.21 Anisotropic concentration in possibility space
   3.22 Semantic morphology versus literal geometric ontology
   3.23 The LSS analogy: shared morphology without shared physical mechanism
   3.24 Training as integration
   3.25 What training fixes and what remains runtime-variable

4. PART IV — TSCT: OWNERSHIP AND GENERATIVE STRUCTURE
   4.1 Event as the initial owner of consequence
   4.2 Event → feature
   4.3 Feature → relation
   4.4 Relation → transformation
   4.5 Transformation → generator
   4.6 Ownership as an explanatory variable
   4.7 Learning as ownership migration
   4.8 More generative carriers versus memorized states
   4.9 Copying a state
   4.10 Copying a generator
   4.11 Reconstruction of unseen states from generators
   4.12 Generalization as generator reuse
   4.13 Double descent as changing carrier ownership
   4.14 Datapoint carriers
   4.15 Carrier conflict
   4.16 Transition to reusable-structure carriers
   4.17 Identifying the minimum generative basis
   4.18 Success as evidence against unnecessary retained structure
   4.19 Deletion and recompression
   4.20 Generator replacement across multiple lower-level structures

5. PART V — THE TIME-INDEXED SEMANTIC CLOUD
   5.1 Why there is no single operationally static SEMANTIC_CLOUD
   5.2 ☁️ₜ = f(history, exposure, observation/question, current constraints)
   5.3 Persistent potential geometry
   5.4 Transient deformation field
   5.5 Fast malleability
   5.5.1 Context
   5.5.2 Attention/routing
   5.5.3 Active representations
   5.5.4 Retrieved information
   5.5.5 Tool observations
   5.5.6 Prior generated tokens
   5.6 Slow malleability
   5.6.1 Weight updates
   5.6.2 Fine-tuning
   5.6.3 Post-training
   5.6.4 Continual learning
   5.6.5 External memory consolidation
   5.7 Semantic accessibility
   5.8 Basin depth
   5.9 Filament reachability
   5.10 Salience
   5.11 Working object identity
   5.12 Branch preference
   5.13 Task representation
   5.14 Semantic state change without weight change
   5.15 weights fixed ≠ cloudₜ fixed

   The earlier material already converges on this distinction: fixed weights can support substantially different active geometry as prompt, retrieval, memory and tool feedback change.

6. PART VI — PROMPTS, CONTEXT, AND CONSTRAINT GEOMETRY
   6.1 Prompt as more than a query
   6.2 Prompt as constraint operator
   6.3 Context as accessibility control
   6.4 Δconstraint → Δaccessibility
   6.5 Δaccessibility → Δbasin depth
   6.6 Δaccessibility → Δfilament reachability
   6.7 Δaccessibility → Δresponse family
   6.8 Instructions
   6.9 Examples
   6.10 Retrieval
   6.11 External memory
   6.12 Persona conditioning
   6.13 Tool outputs
   6.14 Previous conversation
   6.15 Prompt ordering effects
   6.16 Prompt-induced ontology shifts
   6.17 Prompt-induced suppression of previously active structures
   6.18 Prompt engineering as controlled intervention
   6.19 Conditioning versus capability creation
   6.20 Accessibility change versus semantic-basis expansion
   6.21 Context governance as a first-class research object

7. PART VII — AUTOREGRESSIVE GENERATION AS DISSIPATION
   7.1 Cloud → candidate continuation distribution
   7.2 Candidate forms: {🌱, 🌿, 🌳, ✂️, ∅}
   7.3 Omission as an explicit candidate
   7.4 Token selection as commitment
   7.5 Prefix accumulation
   7.6 Successive destruction of alternatives
   7.7 SC₀ → token₁ → SC₁ → token₂ → …
   7.8 Generation as dissipation
   7.9 Early-token path dependence
   7.10 Wrong token → wrong basin
   7.11 Destruction of otherwise valid continuations
   7.12 Irreversibility inside one decoding trajectory
   7.13 Restart as restoration of branching freedom
   7.14 Sampling and branching
   7.15 Beam-style multiplicity versus genuinely different semantic routes
   7.16 When different outputs share the same underlying route
   7.17 Output as one dissipated realization of a much larger possibility field

8. PART VIII — REASONING AS CONSTRAINT PROPAGATION
   8.1 Reasoning versus intelligence
   8.2 Reasoning versus learning
   8.3 Recursive constraint propagation
   8.4 Generated intermediate constraints
   8.5 Chain-of-thought as sequential self-conditioning
   8.6 constraint₁ → constraint₂ → constraint₃ → …
   8.7 Intermediate text as part of the next context
   8.8 Search over semantic trajectories
   8.9 Representation changes during reasoning
   8.10 Reasoning as route construction
   8.11 Reasoning-shaped text versus a distinct reasoning faculty
   8.12 Planning-shaped text versus a persistent planner
   8.13 Belief-consistent text versus theory of mind
   8.14 Self-criticism versus metacognitive faculty
   8.15 Semantic reconstruction as a parsimonious account of many apparent capabilities

   This distinction is central in the uploaded discussion: one general semantic generator under different constraints can produce many apparently modular competencies without establishing that each corresponds to a separately stabilized internal faculty.

9. PART IX — INTELLIGENCE: COMPARATIVE DISCRIMINATION
   9.1 Why semantic generation is not intelligence
   9.2 Candidate generation
   9.3 Choice
   9.4 Action
   9.5 World interaction
   9.6 Consequence production
   9.7 Comparative evaluation
   9.8 better / same / worse
   9.9 🙂 > 😐 > 🙁 as schematic discrimination
   9.10 ✅ > ⚠️ > ❌ as schematic discrimination
   9.11 Intelligence as consequence-sensitive comparison
   9.12 Criterion-dependent intelligence
   9.13 No universal “better” without a scoped criterion
   9.14 Granular intelligence
   9.15 Local intelligence of an individual act
   9.16 Domain-specific intelligence
   9.17 Intelligence without generative abundance
   9.18 Generative abundance without intelligence
   9.19 Intelligence-at-task versus general capability
   9.20 Consequence discrimination as the scarce operation

10. PART X — MEMORY, HISTORY, AND LEARNING
    10.1 Memory = what persisted
    10.2 History = what was ordered
    10.3 Learning = retained (state, choice, outcome) consequence structure
    10.4 Why storage is not learning
    10.5 Why sequence is not learning
    10.6 Why generation is not learning
    10.7 Retention of causal consequence
    10.8 Positive outcome retention
    10.9 Negative outcome retention
    10.10 Branch memory
    10.11 Counterexamples
    10.12 Failed trajectories
    10.13 Exclusion memory
    10.14 Policy change
    10.15 Representation change
    10.16 Search-geometry change
    10.17 Learning as modification of future reachable trajectories
    10.18 outcomeₜ → update(Hₜ₊₁,Qₜ₊₁) → ☁️ₜ₊₁
    10.19 Why ☁️ₜ₊₁ ≠ ☁️ₜ after consequential interaction
    10.20 Persistent external memory versus transient LLM activation

11. PART XI — STOCHASTICITY, NOISE, AND EVOLUTIONARY SEARCH
    11.1 Noise is not knowledge
    11.2 Noise as accessibility perturbation
    11.3 Stochastic branching
    11.4 Exploration
    11.5 Variation
    11.6 Selection
    11.7 Retention
    11.8 structure + variation + discrimination + memory
    11.9 Discovery as controlled evolution
    11.10 Orthodoxy risk under insufficient branching
    11.11 Fantasy risk under insufficient discrimination
    11.12 Randomness versus structurally nonaliased alternatives
    11.13 Exploration temperature versus semantic novelty
    11.14 When stochastic diversity is fake diversity
    11.15 Search ecology rather than sample count

12. PART XII — CODING AS THE CLEANEST LLM LABORATORY
    12.1 Dense prior solution neighborhoods
    12.2 Algorithms
    12.3 APIs
    12.4 Idioms
    12.5 Known bugs
    12.6 Fixes
    12.7 Tests
    12.8 Architectures
    12.9 Candidate generation
    12.10 Compilation as external discrimination
    12.11 Unit testing
    12.12 Type checking
    12.13 Benchmarking
    12.14 Profiling
    12.15 Linting
    12.16 Runtime failure
    12.17 Cheap iteration
    12.18 Reversible failure
    12.19 Candidate density × verifier quality
    12.20 PASS(TESTS) ≠ OPTIMUM
    12.21 Cost functionals for code
    12.22 Latency, memory, maintainability, dependencies and auditability
    12.23 Generative bias toward structural excess
    12.24 Addition versus deletion asymmetry
    12.25 Generate → execute → delete → retest → compress
    12.26 Minimum sufficient implementation
    12.27 Coding as evidence for semantic reconstruction + narrow adversarial elimination

The source material explicitly frames coding as unusually favorable because it combines a dense basis of near-solutions with strong external discriminators and cheap reversible iteration.

13. PART XIII — ROUTE ECOLOGY AND NONALIASED SEARCH
    13.1 The candidate-route population
    13.2 Carrier differences
    13.3 Representation differences
    13.4 Operation differences
    13.5 Relation differences
    13.6 Ordering differences
    13.7 Estimand differences
    13.8 Experimental-design differences
    13.9 Inference differences
    13.10 Boundary assumptions
    13.11 Causal-model differences
    13.12 Primitive differences
    13.13 Structural route identity
    13.14 Seed changes that do not change reachability
    13.15 Persona changes that do not change reachability
    13.16 Wording changes that do not change reachability
    13.17 Fake diversity
    13.18 Noncompetitive search / premature collapse
    13.19 Ungoverned multiverse / multiplicity without discrimination
    13.20 Route count ≠ evidence
    13.21 Majority ≠ source
    13.22 Agreement ≠ source authority
    13.23 Many-analyst systems
    13.24 Generator ecology
    13.25 Auditor ecology
    13.26 Meta-multiverse effects
    13.27 Comparing routes by load-bearing structural differences

14. PART XIV — FAILURE, RESIDUE, AND DISCOVERY
    14.1 Failure signal versus failure repair
    14.2 Failed output versus failed route
    14.3 Failed route versus failed representation
    14.4 Failed representation versus failed search geometry
    14.5 Causal localization
    14.6 Load-bearing structural delta
    14.7 Residue
    14.8 Minimal causal counterkernel
    14.9 Source-answerable discriminator
    14.10 Source contact
    14.11 Eliminate
    14.12 Split
    14.13 Retain
    14.14 Create
    14.15 Dependency rollback
    14.16 Route quarantine
    14.17 Structurally new successor
    14.18 Replay
    14.19 Amnesic replay
    14.20 Persistent provenance
    14.21 Failure-awareness versus failure-causalization
    14.22 Criticism versus state transition
    14.23 Self-review versus self-correction
    14.24 Repeated failure as evidence against the current search geometry
    14.25 Discovery as restructuring future reachability

15. PART XV — AGENTS AND COGNITIVE TRAJECTORIES
    15.1 LLM versus agent
    15.2 AGENT = LLM + persistent state + actions + tools + control loop
    15.3 Transient semantic state
    15.4 Persistent external state
    15.5 Target
    15.6 Current representation
    15.7 Valid prefix
    15.8 Failed branches
    15.9 Repeated attractors
    15.10 Residue
    15.11 Counterexamples
    15.12 Tool outcomes
    15.13 Interventions
    15.14 Policy changes
    15.15 Current state versus cognitive trajectory
    15.16 Σ₀ → action → Σ₁ → observation → Σ₂ → failure → Σ₃ …
    15.17 State awareness
    15.18 Trajectory ownership
    15.19 Search-geometry restructuring
    15.20 Agentic persistence across LLM invocations
    15.21 Externalization of causal history
    15.22 Why long context alone is not persistent cognition

The uploaded material explicitly moves the research target from inspecting a transient “current cognitive state” toward preserving and acting on the causal trajectory of failures, interventions and changed future trajectories.

16. PART XVI — DISTRIBUTED INTELLIGENCE
    16.1 Intelligence need not be co-located
    16.2 Human candidate selection
    16.3 LLM candidate amplification
    16.4 Instrumental action
    16.5 Environmental consequence
    16.6 Human or machine discrimination
    16.7 External memory retention
    16.8 Human + LLM + tools + world
    16.9 Distributed semantic cloud → choice → outcome loop
    16.10 Organizational intelligence
    16.11 Scientific teams as distributed intelligence systems
    16.12 Human semantic gates
    16.13 Machine verification gates
    16.14 Mixed human-machine ownership of research trajectories

17. PART XVII — META-FUNCTIONS AROUND INTELLIGENCE
    17.1 Language as transport
    17.2 Mathematics as compression
    17.3 Culture as persistence
    17.4 Writing as external memory
    17.5 Libraries as accumulated consequence structure
    17.6 Scientific literature as historical execution trace
    17.7 LLM as candidate-cloud amplifier
    17.8 Search engines as source-access infrastructure
    17.9 Formal systems as discriminators
    17.10 Experimental systems as consequence generators
    17.11 Meta-functions that extend intelligence without themselves being identical to intelligence

18. PART XVIII — MECHANISTIC INTERPRETABILITY
    18.1 What standard LLM research measures
    18.1.1 Tokens
    18.1.2 Loss
    18.1.3 Activations
    18.1.4 Neurons
    18.1.5 Features
    18.1.6 Sparse autoencoders
    18.1.7 Circuits
    18.1.8 Attention paths
    18.1.9 Jacobians
    18.1.10 Benchmarks
    18.2 Measurement versus ontology
    18.3 Activation → decomposition → feature → human label
    18.4 Human label ≠ native ontology
    18.5 Sparse representation ≠ native causal object
    18.6 Low reconstruction error ≠ mechanism identity
    18.7 Clean feature ≠ mechanism identity
    18.8 Output match versus mechanism match
    18.9 Jacobian match versus output match
    18.10 Jacobian match versus intervention equivalence
    18.11 Local intervention versus global mechanism equivalence
    18.12 Representation dependence
    18.13 Carrier invariance
    18.14 Factorization stability
    18.15 Causal transport
    18.16 Constraint propagation
    18.17 Mechanistic interpretation as search for native causal objects

19. PART XIX — THE NATIVE-CAUSAL-OBJECT RESEARCH PROGRAM
    19.1 Stop asking only “where is concept X?”
    19.2 What are the native causal objects?
    19.3 How are they formed?
    19.4 How do they combine?
    19.5 How do they decompose?
    19.6 How do they move between carriers?
    19.7 How does ownership transfer?
    19.8 How does constraint move through them?
    19.9 What constitutes a basin?
    19.10 What constitutes a boundary?
    19.11 What constitutes a filament?
    19.12 What determines reachability?
    19.13 What survives perturbation?
    19.14 What survives reparameterization?
    19.15 What survives model scale changes?
    19.16 What survives representation changes?
    19.17 How are generators represented?
    19.18 Can native object identity be established interventionally?
    19.19 Can multiple mechanistic decompositions be proven equivalent?
    19.20 Does a minimal native ontology exist?

20. PART XX — SOURCE SOVEREIGNTY AND EPISTEMIC GOVERNANCE
    20.1 Source
    20.2 Representation
    20.3 Semantics
    20.4 Access
    20.5 Readout
    20.6 Authority
    20.7 Role separation
    20.8 Source ownership
    20.9 Target blindness
    20.10 No target-fed constructors
    20.11 Representation grades
    20.12 Transport
    20.13 Naturality
    20.14 Failure normalization
    20.15 Obstruction memory
    20.16 Certification
    20.17 Replay
    20.18 Provenance
    20.19 No certificate → source-truth promotion
    20.20 No bounded closure → global closure promotion

21. PART XXI — CLOSED SEMANTIC KERNEL / OPEN DISCOVERY
    21.1 Why discovery needs fixed semantics somewhere
    21.2 Closed kernel K₀
    21.3 Type
    21.4 Identity
    21.5 Apply
    21.6 Check
    21.7 Enumerate
    21.8 Fold
    21.9 Normalize
    21.10 Encode / decode
    21.11 Open discovery layer K₁
    21.12 Candidate proposals
    21.13 Proposals as untrusted
    21.14 Compilation into executable bodies
    21.15 Explicit failure outputs
    21.16 Finite search universes
    21.17 Exclusion antichains
    21.18 Successor generation
    21.19 Certification before commitment
    21.20 Replay-exact persistence
    21.21 New-primitive candidates
    21.22 Why discovery may alter data/bodies but not silently mutate kernel semantics

The current GRM successor explicitly formalizes this as a fixed K₀ and an open K₁, with discovery proposals remaining untrusted until compiled and witnessed; it also explicitly forbids global closure.

22. PART XXII — SCIENCE WITH LLMs
    22.1 Cheap semantic possibility
    22.2 Hypothesis abundance
    22.3 Cheap code generation
    22.4 Cheap analytic recombination
    22.5 Cheap literature synthesis
    22.6 Scarcity moves downstream
    22.7 Experimental scarcity
    22.8 Reality-contact scarcity
    22.9 Verification scarcity
    22.10 Ontology scarcity
    22.11 Target-definition scarcity
    22.12 Candidate hypothesis → specialized model
    22.13 Specialized model → experiment
    22.14 Experiment → reality
    22.15 Reality → comparative discrimination
    22.16 Elimination
    22.17 Ownership transfer
    22.18 Semantic-cloud reconstruction
    22.19 Human-in-the-loop scientific systems
    22.20 Formal verification versus intended scientific claim
    22.21 CERTIFICATION ≠ TARGET OWNERSHIP

23. PART XXIII — RESEARCH AUTOMATION VERSUS DISCOVERY
    23.1 Bare LLM
    23.2 LLM + tools
    23.3 LLM + persistent workspace
    23.4 Domain harness
    23.5 Numerical oracle
    23.6 Symbolic model
    23.7 Formal verification
    23.8 Human semantic gate
    23.9 Constrained research discovery
    23.10 Expert-designed discovery geometry
    23.11 Automated traversal
    23.12 Symbolic recovery
    23.13 Verification
    23.14 Why this is not yet self-developing discovery geometry
    23.15 Recurrent failure
    23.16 Diagnosing the search geometry itself
    23.17 Recognizing “the harness is now the constraint”
    23.18 Inventing a new methodology
    23.19 Constructing a new discriminator
    23.20 Validating the methodology
    23.21 Persisting the new capability
    23.22 Transfer to new problems

This distinction appears directly in the research-automation discussion: domain harnesses can produce genuine constrained discovery while still lacking the loop that diagnoses and restructures their own discovery geometry.

24. PART XXIV — PAPERS AS SEMANTIC AND EXECUTABLE ARTIFACTS
    24.1 Paper as historical execution trace, not automatic authority
    24.2 Linear paper limitations
    24.3 Interactive table of contents
    24.4 Global semantic map
    24.5 Arbitrary local descent
    24.6 Paper → navigable semantic topology
    24.7 Extracting native objects
    24.8 Operations
    24.9 Relations
    24.10 Carriers
    24.11 Equations
    24.12 Hypotheses
    24.13 Transport maps
    24.14 Representations
    24.15 Boundary/failure points
    24.16 Claimed outputs
    24.17 Source-owned versus derived claims
    24.18 Conjectural material
    24.19 Representation-bound material
    24.20 Reusable operators
    24.21 Broken or obsolete-but-generative structures

The GMEG material explicitly treats literature as something to decompile into native objects, operations, relations, carriers, equations, hypotheses, transports, representations and failure points rather than treating the paper itself as authority.

25. PART XXV — PAPER → MCP → VALIDATED SCIENTIFIC OPERATOR
    25.1 Interactive TOC as possibility preservation
    25.2 Paper + code + data
    25.3 Custom MCP / executable method
    25.4 Source binding
    25.5 Environment binding
    25.6 Tests
    25.7 Workflow semantics
    25.8 Reproducibility receipts
    25.9 Validated MCP
    25.10 Machine-callable scientific operator
    25.11 Possibility preservation versus possibility commitment
    25.12 Why every paper need not become an MCP
    25.13 Papers suitable for MCP conversion
    25.14 Load-bearing papers
    25.15 Dissenting papers
    25.16 Adversarial MCPs
    25.17 Executable disagreement between competing scientific methods

26. PART XXVI — FAILURE MODES OF LLM RESEARCH ITSELF
    26.1 Premature explanation
    26.2 Tooling built around premature explanation
    26.3 Projection becoming ontology
    26.4 Anomalies becoming patches
    26.5 Infrastructure lock-in
    26.6 Hidden deeper object
    26.7 Rules
    26.8 Statistics
    26.9 Vectors
    26.10 Attention
    26.11 Features
    26.12 Circuits
    26.13 Agents
    26.14 Each framework finds something real
    26.15 Something real ≠ whole object
    26.16 Benchmark optimization masquerading as explanation
    26.17 Human label capture
    26.18 Semantic self-certification
    26.19 Critic/auditor role labels creating false authority
    26.20 Consensus collapse
    26.21 Multiverse without discrimination
    26.22 Search-space lock-in
    26.23 Failure to let evidence force ontology change

27. PART XXVII — EXPERIMENTAL PROGRAM FOR SEMANTIC-CLOUD SCIENCE
    27.1 Mapping exposure
    27.2 Controlled curriculum restriction
    27.3 Semantic boundary experiments
    27.4 Prompt perturbation experiments
    27.5 Context permutation experiments
    27.6 Persona perturbation experiments
    27.7 Retrieval perturbation experiments
    27.8 Tool-feedback perturbation
    27.9 Branch/restart studies
    27.10 Early-token intervention experiments
    27.11 Accessibility mapping
    27.12 Basin-transition measurement
    27.13 Filament survival tests
    27.14 Candidate-diversity versus route-diversity tests
    27.15 Intervention-equivalence tests
    27.16 Representation-change tests
    27.17 Factorization tests
    27.18 Generator-ablation tests
    27.19 Ownership-transfer tests
    27.20 Learning-as-consequence-retention experiments
    27.21 Memory versus history ablations
    27.22 Intelligence-as-comparison experiments
    27.23 Agent trajectory persistence experiments
    27.24 Search-geometry restructuring experiments
    27.25 Self-developed methodology benchmarks
    27.26 Human–LLM distributed-loop experiments

28. PART XXVIII — MEASUREMENT AND BENCHMARK DESIGN
    28.1 Stop measuring only final answer accuracy
    28.2 Exposure measurement
    28.3 Accessibility measurement
    28.4 Candidate diversity
    28.5 Structural route diversity
    28.6 Basin stability
    28.7 Constraint sensitivity
    28.8 Generator reuse
    28.9 Outcome quality
    28.10 Comparative discrimination quality
    28.11 Consequence retention
    28.12 Negative-knowledge retention
    28.13 Transfer
    28.14 Replay stability
    28.15 Provenance completeness
    28.16 Representation stability
    28.17 Intervention stability
    28.18 Source-contact quality
    28.19 Search-space mutation quality
    28.20 Minimum sufficient structure

29. PART XXIX — CENTRAL OPEN QUESTIONS
    29.1 Does a useful native semantic geometry exist beyond metaphor?
    29.2 What are its causal objects?
    29.3 What determines semantic reachability?
    29.4 What exactly does scaling change?
    29.5 What genuinely enlarges the reachable basis?
    29.6 When does recombination become invention?
    29.7 What constitutes a generator?
    29.8 How does generator ownership migrate during training?
    29.9 How does runtime constraint propagation alter accessibility?
    29.10 What makes one branch structurally distinct from another?
    29.11 How can valid thin filaments survive density-biased decoding?
    29.12 Can intelligence be measured independently from generative abundance?
    29.13 What is the minimal comparator required for intelligence?
    29.14 How is outcome quality represented and retained?
    29.15 What distinguishes memory from learned consequence structure computationally?
    29.16 Can LLM systems autonomously detect that their search space is wrong?
    29.17 Can they construct genuinely new search geometry?
    29.18 Can they design new reality-contact discriminators?
    29.19 Can a mechanistic explanation survive representation and factorization changes?
    29.20 Can scientific systems preserve dissenting thin routes rather than collapsing to consensus?
    29.21 What should remain invariant when an agent changes ontology?
    29.22 What is the correct unit of persistent learning in an agentic system?
    29.23 How should distributed human-machine intelligence be attributed and evaluated?
    29.24 Where should authority reside when generation, action, discrimination, and retention are distributed?

30. PART XXX — GRAND SYNTHESIS
    30.1 WORLD → TOKEN SPACE → TRAINING → STRUCTURED SEMANTIC POSSIBILITY
    30.2 history + exposure + current constraint → ☁️ₜ
    30.3 ☁️ₜ → candidate continuations
    30.4 candidate → choice
    30.5 choice → action
    30.6 action → world outcome
    30.7 outcomes → comparative discrimination
    30.8 comparative discrimination = local intelligence
    30.9 consequence retention = learning
    30.10 ordered retained events = history
    30.11 persisted structure = memory
    30.12 history + learning + changed world → ☁️ₜ₊₁
    30.13 branching + discrimination + retention → discovery
    30.14 repeated failure + causalization → ontology/search-geometry change
    30.15 language → transport
    30.16 mathematics → compression
    30.17 culture → persistence
    30.18 LLM → semantic candidate amplifier
    30.19 environment/experiment → consequence generator
    30.20 intelligence → better/same/worse discrimination
    30.21 learning → retained consequence structure
    30.22 distributed system → next act, next cloud, next world