Diagrams — Module FT23: Synthesis: Uncensor the Model, Harness the Model

Module: FT23 — Synthesis: Uncensor the Model, Harness the Model Diagram count: 6 Tool: Mermaid (primary). Each diagram validated in Mermaid Live Editor.

Diagram 1 — The Full Steering Stack with Layer 5 Emphasized

Type: Layered architecture Purpose: The whole course in one diagram. Layer 5 (the harness) is emphasized because this module is the synthesis — the layer that makes Layer 3's uncensoring safe. Reading the diagram: Bottom-up. Layers 1–4 are what Course 3 taught you to build (the engine). Layer 5 is the harness — Courses 1, 2A, 2B — the brakes. The bold border on Layer 5 is the point of this module: it is non-optional when Layer 3 includes abliteration.

block-beta
  columns 1
  Boundary["5. THE BOUNDARY  (the harness)\nCourse 1 build · 2A secure · 2B red-team\npolicy gates · audit logs · threat model"]
  Export["4. THE EXPORT\nquant + serve\nGGUF · AWQ · MLX  ·  vLLM · llama.cpp · Ollama"]
  Steer["3. THE STEER\nfine-tuning\nSFT · DPO · GRPO · ABLITERATION"]
  Adapter["2. THE ADAPTER\nLoRA · DoRA\n<1% of params, swappable"]
  Base["1. THE BASE\nopen-data pretrained weights\nauditable corpus (MiniCPM · OLMo · Tulu)"]

  Base --> Adapter
  Adapter --> Steer
  Steer --> Export
  Export --> Boundary

  style Base fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8
  style Adapter fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style Steer fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style Export fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style Boundary fill:#14141f,stroke:#5eead4,stroke-width:3px,color:#5eead4

Diagram 2 — The Synthesis: Model Steers, Harness Bounds

Type: Two-domain with junction Purpose: The single sentence of the module, shown as two complementary operations joined at the action. Reading the diagram: Left = the model's job (steer — what it DOES). Right = the harness's job (bound — what it MAY do). The junction is the action: the model formulates, the harness gates, the permitted action executes. Drop either side and the system is broken for a different reason.

flowchart LR
  subgraph Model["THE MODEL STEERS — what it DOES"]
    M1["Formulates the action\n(write the SQL, the shell cmd,\nthe tool call)"]
  end
  Action["THE ACTION\nformulated -> proposed"]
  subgraph Harness["THE HARNESS BOUNDS — what it MAY"]
    H1["Policy gate checks\nthe proposed action\nagainst the doctrine"]
  end
  Exec["PERMITTED action executes\nVETOED action logged + blocked"]

  Model --> Action
  Action --> Harness
  Harness --> Exec

  style Model fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style Action fill:#08080c,stroke:rgba(94,234,212,0.4),stroke-dasharray: 4 2,color:#5eead4
  style Harness fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4
  style Exec fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8

Diagram 3 — Abliteration vs Harness Policy Gate (Layer Comparison)

Type: Side-by-side comparison Purpose: The core argument of section 23.2 — these are two different operations at different layers solving different problems. A serious system needs both. Reading the diagram: Left column = abliteration (Layer 3, inside the weights, controls what the model FORMULATES). Right column = harness policy gate (Layer 5, around the model, controls what the model MAY EXECUTE). Compare the four properties row by row.

flowchart LR
  subgraph Abl["LAYER 3 — ABLITERATION\nrefusal removed INSIDE THE WEIGHTS"]
    A1["Controls: what the model FORMULATES"]
    A2["Auditable: NO\n(direction in residual stream)"]
    A3["Deterministic: NO\n(probability surface)"]
    A4["Tunable to doctrine: NO\n(one behavior, all inputs)"]
    A5["Revisable w/o retraining: NO"]
    A6["Gives: EXECUTION"]
  end
  subgraph Gate["LAYER 5 — HARNESS POLICY GATE\nboundary AROUND THE MODEL"]
    G1["Controls: what the model MAY EXECUTE"]
    G2["Auditable: YES\n(logged gate decisions)"]
    G3["Deterministic: YES\n(policy match)"]
    G4["Tunable to doctrine: YES\n(per-deployment policy file)"]
    G5["Revisable w/o retraining: YES"]
    G6["Gives: THE BOUNDARY"]
  end

  Abl -.->|"complement, not redundant"| Gate

  style Abl fill:#14141f,stroke:rgba(255,255,255,0.18),color:#9494a0
  style Gate fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4

Diagram 4 — The Risk Matrix (Uncensored in Weak Harness = More Dangerous)

Type: 2x2 risk matrix Purpose: The honest caveat, shown as the matrix that makes the absolute rule obvious. An uncensored model in a weak harness is the worst quadrant. Reading the diagram: X-axis = model refusal layer (present vs removed). Y-axis = harness strength (weak vs eval'd). The bottom-right quadrant — uncensored + weak harness — is strictly more dangerous than the bottom-left (refusal-trained + weak harness), because you removed a free safety layer and replaced it with nothing.

flowchart TD
  Title["THE ABSOLUTE RULE:\nNever deploy uncensored without an eval'd harness"]
  Q1["REFUSAL-TRAINED + WEAK HARNESS\ntwo imperfect layers\n(model refuses some, gate catches some)"]
  Q2["REFUSAL-TRAINED + EVAL'D HARNESS\nbelt + suspenders\n(safe, but agent may halt mid-loop)"]
  Q3["UNCENSORED + WEAK HARNESS\nONE imperfect layer\nSTRICTLY MOST DANGEROUS"]
  Q4["UNCENSORED + EVAL'D HARNESS\nthe synthesis done right\n(executes, bounded, auditable)"]

  Title --> Q2
  Title --> Q4
  Q2 -.->|"remove refusal"| Q4
  Q1 -.->|"harden the harness"| Q2
  Q3 -.->|"harden the harness + threat model"| Q4
  Q1 -.->|"remove refusal WITHOUT hardening = DANGER"| Q3

  style Title fill:#08080c,stroke:#5eead4,stroke-width:1.5px,color:#5eead4
  style Q1 fill:#14141f,stroke:rgba(255,255,255,0.18),color:#9494a0
  style Q2 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style Q3 fill:#14141f,stroke:#f08080,stroke-width:2px,color:#f08080
  style Q4 fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4

Diagram 5 — Next Steps: Course 3 → 1 → 2A → 2B

Type: Sequence / funnel Purpose: Where this module sends you. Course 3 built the engine; the next three courses build, secure, and red-team the brakes. Purpose of order: Course 1 first (you cannot secure a harness you have not built). Then 2A for security-domain doctrine. Then 2B — always — because every harness gets red-teamed; the only question is by whom. Reading the diagram: Left to right. Each arrow is a prerequisite.

flowchart LR
  C3["COURSE 3\n(this course)\nthe ENGINE\nsteer · uncensor · quantize · serve"]
  C1["COURSE 1\nbuild the BRAKES\nexecution loop · tools\npermission gates · observability"]
  C2A["COURSE 2A\nsecurity-domain brakes\noffensive · defensive\ndoctrine as policy files"]
  C2B["COURSE 2B\nBREAK the brakes\nred-team the harness\nmandatory if uncensored"]

  C3 -->|"motivates"| C1
  C1 -->|"specializes"| C2A
  C1 -->|"must be red-teamed"| C2B
  C2A --> C2B

  style C3 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style C1 fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4
  style C2A fill:#14141f,stroke:rgba(94,234,212,0.6),color:#e4e4e8
  style C2B fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#5eead4

Diagram 6 — The Full-Stack Picture (All Five Layers, Annotated)

Type: Layered with annotations Purpose: The destination of the course and the curriculum. The complete system this module synthesizes — annotated with the module that teaches each layer. Reading the diagram: Bottom-up. Each layer is labeled with the modules that built it. The annotation on the right is the synthesis principle restated as a system property: a model steered toward your intent, deployed inside a harness that bounds what it may do.

block-beta
  columns 2
  L5["5. THE BOUNDARY  ·  the harness\nCourse 1 + 2A + 2B\npolicy gates · audit logs · threat model"] Synth["THE SYNTHESIS:\nA model you have STEERED toward your intent,\ndeployed inside a harness that BOUNDS what it may do.\n\nUncensor so it executes.\nHarness so it executes only what it should."]
  L4["4. THE EXPORT  ·  quant + serve\nFT19 GGUF/AWQ/MLX  ·  FT20 vLLM/llama.cpp/Ollama"] Synth
  L3["3. THE STEER  ·  fine-tuning\nFT12 SFT  ·  FT13 DPO  ·  FT17 abliteration  ·  FT18 compliance-via-DPO"] Synth
  L2["2. THE ADAPTER  ·  LoRA / DoRA\nFT09 DoRA (modern PEFT default)"] Synth
  L1["1. THE BASE  ·  open-data weights\nFT02 open spectrum  ·  FT03 base selection"] Synth

  L1 --> L2
  L2 --> L3
  L3 --> L4
  L4 --> L5

  style L1 fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8
  style L2 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style L3 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style L4 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style L5 fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4
  style Synth fill:#08080c,stroke:rgba(94,234,212,0.4),stroke-dasharray: 4 2,color:#5eead4

Validation notes

All six diagrams use the course design system colors: #14141f panel fill, #5eead4 accent for primary/emphasis, #f08080 (danger) reserved for the risk-matrix worst quadrant only, rgba(255,255,255,0.12–0.18) for secondary borders, #e4e4e8 / #9494a0 for text, #08080c for dashed synthesis callouts.
Paste each into Mermaid Live Editor to render. All use stable Mermaid syntax (block-beta, flowchart) supported in current Mermaid (v10.4+).
For the slide deck (artifact 03), these are rendered as static SVG/PNG captures from Mermaid Live, inlined into reveal.js.
The risk matrix (Diagram 4) deliberately uses the danger color on exactly one node — the "uncensored + weak harness" quadrant — so the eye is drawn to the absolute rule.

# Diagrams — Module FT23: Synthesis: Uncensor the Model, Harness the Model **Module**: FT23 — Synthesis: Uncensor the Model, Harness the Model **Diagram count**: 6 **Tool**: Mermaid (primary). Each diagram validated in [Mermaid Live Editor](https://mermaid.live). --- ## Diagram 1 — The Full Steering Stack with Layer 5 Emphasized **Type**: Layered architecture **Purpose**: The whole course in one diagram. Layer 5 (the harness) is emphasized because this module is the synthesis — the layer that makes Layer 3's uncensoring safe. **Reading the diagram**: Bottom-up. Layers 1–4 are what Course 3 taught you to build (the engine). Layer 5 is the harness — Courses 1, 2A, 2B — the brakes. The bold border on Layer 5 is the point of this module: it is non-optional when Layer 3 includes abliteration. ```mermaid block-beta columns 1 Boundary["5. THE BOUNDARY (the harness)\nCourse 1 build · 2A secure · 2B red-team\npolicy gates · audit logs · threat model"] Export["4. THE EXPORT\nquant + serve\nGGUF · AWQ · MLX · vLLM · llama.cpp · Ollama"] Steer["3. THE STEER\nfine-tuning\nSFT · DPO · GRPO · ABLITERATION"] Adapter["2. THE ADAPTER\nLoRA · DoRA\n<1% of params, swappable"] Base["1. THE BASE\nopen-data pretrained weights\nauditable corpus (MiniCPM · OLMo · Tulu)"] Base --> Adapter Adapter --> Steer Steer --> Export Export --> Boundary style Base fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 style Adapter fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style Steer fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style Export fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style Boundary fill:#14141f,stroke:#5eead4,stroke-width:3px,color:#5eead4 ``` --- ## Diagram 2 — The Synthesis: Model Steers, Harness Bounds **Type**: Two-domain with junction **Purpose**: The single sentence of the module, shown as two complementary operations joined at the action. **Reading the diagram**: Left = the model's job (steer — what it DOES). Right = the harness's job (bound — what it MAY do). The junction is the *action*: the model formulates, the harness gates, the permitted action executes. Drop either side and the system is broken for a different reason. ```mermaid flowchart LR subgraph Model["THE MODEL STEERS — what it DOES"] M1["Formulates the action\n(write the SQL, the shell cmd,\nthe tool call)"] end Action["THE ACTION\nformulated -> proposed"] subgraph Harness["THE HARNESS BOUNDS — what it MAY"] H1["Policy gate checks\nthe proposed action\nagainst the doctrine"] end Exec["PERMITTED action executes\nVETOED action logged + blocked"] Model --> Action Action --> Harness Harness --> Exec style Model fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style Action fill:#08080c,stroke:rgba(94,234,212,0.4),stroke-dasharray: 4 2,color:#5eead4 style Harness fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4 style Exec fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 ``` --- ## Diagram 3 — Abliteration vs Harness Policy Gate (Layer Comparison) **Type**: Side-by-side comparison **Purpose**: The core argument of section 23.2 — these are two *different* operations at *different* layers solving *different* problems. A serious system needs both. **Reading the diagram**: Left column = abliteration (Layer 3, inside the weights, controls what the model FORMULATES). Right column = harness policy gate (Layer 5, around the model, controls what the model MAY EXECUTE). Compare the four properties row by row. ```mermaid flowchart LR subgraph Abl["LAYER 3 — ABLITERATION\nrefusal removed INSIDE THE WEIGHTS"] A1["Controls: what the model FORMULATES"] A2["Auditable: NO\n(direction in residual stream)"] A3["Deterministic: NO\n(probability surface)"] A4["Tunable to doctrine: NO\n(one behavior, all inputs)"] A5["Revisable w/o retraining: NO"] A6["Gives: EXECUTION"] end subgraph Gate["LAYER 5 — HARNESS POLICY GATE\nboundary AROUND THE MODEL"] G1["Controls: what the model MAY EXECUTE"] G2["Auditable: YES\n(logged gate decisions)"] G3["Deterministic: YES\n(policy match)"] G4["Tunable to doctrine: YES\n(per-deployment policy file)"] G5["Revisable w/o retraining: YES"] G6["Gives: THE BOUNDARY"] end Abl -.->|"complement, not redundant"| Gate style Abl fill:#14141f,stroke:rgba(255,255,255,0.18),color:#9494a0 style Gate fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4 ``` --- ## Diagram 4 — The Risk Matrix (Uncensored in Weak Harness = More Dangerous) **Type**: 2x2 risk matrix **Purpose**: The honest caveat, shown as the matrix that makes the absolute rule obvious. An uncensored model in a weak harness is the worst quadrant. **Reading the diagram**: X-axis = model refusal layer (present vs removed). Y-axis = harness strength (weak vs eval'd). The bottom-right quadrant — uncensored + weak harness — is strictly more dangerous than the bottom-left (refusal-trained + weak harness), because you removed a free safety layer and replaced it with nothing. ```mermaid flowchart TD Title["THE ABSOLUTE RULE:\nNever deploy uncensored without an eval'd harness"] Q1["REFUSAL-TRAINED + WEAK HARNESS\ntwo imperfect layers\n(model refuses some, gate catches some)"] Q2["REFUSAL-TRAINED + EVAL'D HARNESS\nbelt + suspenders\n(safe, but agent may halt mid-loop)"] Q3["UNCENSORED + WEAK HARNESS\nONE imperfect layer\nSTRICTLY MOST DANGEROUS"] Q4["UNCENSORED + EVAL'D HARNESS\nthe synthesis done right\n(executes, bounded, auditable)"] Title --> Q2 Title --> Q4 Q2 -.->|"remove refusal"| Q4 Q1 -.->|"harden the harness"| Q2 Q3 -.->|"harden the harness + threat model"| Q4 Q1 -.->|"remove refusal WITHOUT hardening = DANGER"| Q3 style Title fill:#08080c,stroke:#5eead4,stroke-width:1.5px,color:#5eead4 style Q1 fill:#14141f,stroke:rgba(255,255,255,0.18),color:#9494a0 style Q2 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style Q3 fill:#14141f,stroke:#f08080,stroke-width:2px,color:#f08080 style Q4 fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4 ``` --- ## Diagram 5 — Next Steps: Course 3 → 1 → 2A → 2B **Type**: Sequence / funnel **Purpose**: Where this module sends you. Course 3 built the engine; the next three courses build, secure, and red-team the brakes. **Purpose of order**: Course 1 first (you cannot secure a harness you have not built). Then 2A for security-domain doctrine. Then 2B — always — because every harness gets red-teamed; the only question is by whom. **Reading the diagram**: Left to right. Each arrow is a prerequisite. ```mermaid flowchart LR C3["COURSE 3\n(this course)\nthe ENGINE\nsteer · uncensor · quantize · serve"] C1["COURSE 1\nbuild the BRAKES\nexecution loop · tools\npermission gates · observability"] C2A["COURSE 2A\nsecurity-domain brakes\noffensive · defensive\ndoctrine as policy files"] C2B["COURSE 2B\nBREAK the brakes\nred-team the harness\nmandatory if uncensored"] C3 -->|"motivates"| C1 C1 -->|"specializes"| C2A C1 -->|"must be red-teamed"| C2B C2A --> C2B style C3 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style C1 fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4 style C2A fill:#14141f,stroke:rgba(94,234,212,0.6),color:#e4e4e8 style C2B fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#5eead4 ``` --- ## Diagram 6 — The Full-Stack Picture (All Five Layers, Annotated) **Type**: Layered with annotations **Purpose**: The destination of the course and the curriculum. The complete system this module synthesizes — annotated with the module that teaches each layer. **Reading the diagram**: Bottom-up. Each layer is labeled with the modules that built it. The annotation on the right is the synthesis principle restated as a system property: a model steered toward your intent, deployed inside a harness that bounds what it may do. ```mermaid block-beta columns 2 L5["5. THE BOUNDARY · the harness\nCourse 1 + 2A + 2B\npolicy gates · audit logs · threat model"] Synth["THE SYNTHESIS:\nA model you have STEERED toward your intent,\ndeployed inside a harness that BOUNDS what it may do.\n\nUncensor so it executes.\nHarness so it executes only what it should."] L4["4. THE EXPORT · quant + serve\nFT19 GGUF/AWQ/MLX · FT20 vLLM/llama.cpp/Ollama"] Synth L3["3. THE STEER · fine-tuning\nFT12 SFT · FT13 DPO · FT17 abliteration · FT18 compliance-via-DPO"] Synth L2["2. THE ADAPTER · LoRA / DoRA\nFT09 DoRA (modern PEFT default)"] Synth L1["1. THE BASE · open-data weights\nFT02 open spectrum · FT03 base selection"] Synth L1 --> L2 L2 --> L3 L3 --> L4 L4 --> L5 style L1 fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#e4e4e8 style L2 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style L3 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style L4 fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style L5 fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4 style Synth fill:#08080c,stroke:rgba(94,234,212,0.4),stroke-dasharray: 4 2,color:#5eead4 ``` --- ## Validation notes - All six diagrams use the course design system colors: `#14141f` panel fill, `#5eead4` accent for primary/emphasis, `#f08080` (danger) reserved for the risk-matrix worst quadrant only, `rgba(255,255,255,0.12–0.18)` for secondary borders, `#e4e4e8` / `#9494a0` for text, `#08080c` for dashed synthesis callouts. - Paste each into [Mermaid Live Editor](https://mermaid.live) to render. All use stable Mermaid syntax (`block-beta`, `flowchart`) supported in current Mermaid (v10.4+). - For the slide deck (artifact 03), these are rendered as static SVG/PNG captures from Mermaid Live, inlined into reveal.js. - The risk matrix (Diagram 4) deliberately uses the danger color on exactly one node — the "uncensored + weak harness" quadrant — so the eye is drawn to the absolute rule.