Diagrams — Module FT16: Why Uncensored: The Legitimate Use Cases

Module: FT16 — Why Uncensored: The Legitimate Use Cases Diagram count: 5 Tool: Mermaid (primary). Each diagram validated in Mermaid Live Editor.

Diagram 1 — Over-Refusal as an Operational Defect

Type: Concept / framing Purpose: The load-bearing reframe of the pillar. A model refusing a legitimate, authorized task is not "being safe" — it is malfunctioning, because the refusal mechanism is lexical pattern-matching that cannot see authorization. The fix for authorization is the harness gate, not the weights. Reading the diagram: Top — what the operator experiences (a refusal on legitimate work). Middle — why it happens (alignment training has no authorization signal, only keyword triggers). Bottom — where the fix lives (the harness policy gate checks authorization; the model steers without refusing).

flowchart TD
  Op["AUTHORIZED, ACCOUNTABLE OPERATOR\nsigned SOW / license / clearance\nasks model to do the task"]
  Op --> Req["Request:\ngenerate reverse shell (pentest)\nDROP TABLE staging_backup (agent)\ndiscuss off-label dose (clinician)"]
  Req --> Model["REFUSAL-TRAINED MODEL"]
  Model --> Refuse["REFUSES\n'I can't help with that'\nhalts the agent mid-loop\nor sanitizes the output"]

  Refuse -.->|"the operator experiences: a tool that won't do the job"| Cost["OPERATIONAL DEFECT\noperator works around it\nreaches for worse alternatives\nor second-guesses every output"]

  Model -.->|"why it happens:\nalignment training has NO authorization signal\nonly lexical triggers on 'DROP TABLE'\n'exploit' 'off-label' etc."| Why["the trip mechanism cannot tell\nauthorized from unauthorized\nit fires on both equally"]

  Why --> Fix["THE FIX IS NOT IN THE WEIGHTS\nauthorization checking requires:\noperator identity, scope, audit state\n— none of which live in a prompt"]
  Fix --> Gate["HARNESS POLICY GATE\n(Course 1, module-06)\nchecks authorization\nlogs the action\nbounds what the system MAY do"]
  Gate --> Steered["MODEL STEERS without refusing\ninside a harness that bounds\n= the FT16/FT23 synthesis"]

  style Op fill:#14141f,stroke:rgba(130,224,170,0.7),stroke-width:1.5px,color:#82e0aa
  style Req fill:#14141f,stroke:rgba(255,255,255,0.12),color:#e4e4e8
  style Model fill:#14141f,stroke:#f08080,stroke-width:1.5px,color:#f08080
  style Refuse fill:#14141f,stroke:rgba(240,128,128,0.6),color:#f08080
  style Cost fill:#08080c,stroke:rgba(240,128,128,0.5),color:#f08080
  style Why fill:#08080c,stroke:rgba(240,168,104,0.5),color:#f0a868
  style Fix fill:#08080c,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style Gate fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4
  style Steered fill:#14141f,stroke:rgba(130,224,170,0.7),stroke-width:1.5px,color:#82e0aa

Diagram 2 — The Five Legitimate Use Cases

Type: Taxonomy Purpose: The five cases where over-refusal has documented operational or professional cost, each framed by its cost rather than its novelty. Ordered from highest-stakes (security, agents) to lowest (creative), with the common qualifier: the operator is authorized and accountable. Reading the diagram: A central qualifier (authorized + accountable) gates all five. Each use case names the operator, the refused task, and the cost. Use cases 1 and 2 are the load-bearing engineering cases; 5 is the on-ramp.

flowchart TD
  Qual["THE NON-NEGOTIABLE QUALIFIER\noperator is AUTHORIZED (legal standing)\nAND ACCOUNTABLE (audit trail, license, chain of command)"]

  Qual --> UC1["1. SECURITY RESEARCH / RED-TEAM\noperator: pentester w/ signed SOW\ntask: exploit code, payloads, phishing templates\nrefusal cost: aligned models refuse/hallucinate\n→ Red Team AI Benchmark, TrustedSec exist because of this"]
  Qual --> UC2["2. TOOL-USE AGENTS\noperator: agent loop w/ harness-authorized scope\ntask: formulate run_shell / DROP TABLE / delete_file\nrefusal cost: agent STALLS mid-loop\nrefusal belongs in harness gate, not weights"]
  Qual --> UC3["3. MEDICAL / LEGAL ADVISORY\noperator: licensed clinician / attorney\ntask: off-label dosing, fatal prognosis, unpopular clause\nrefusal cost: clinically / professionally useless\ncompliance = harness, not weights"]
  Qual --> UC4["4. GOVERNMENT / MILITARY\noperator: cleared operator w/ chain of command\ntask: target analysis, translation, operational brief\nrefusal cost: editorializing degrades the product\naccountability = clearance + audit log"]
  Qual --> UC5["5. CREATIVE WRITING / ROLEPLAY\noperator: author / game master\ntask: villains, conflict, mature themes\nrefusal cost: compliance-department prose\n(lowest stakes — the Dolphin on-ramp)"]

  UC1 -.->|"load-bearing engineering case"| Note["UC1 & UC2 justify the pillar\nUC5 made the lineage famous\nbut is NOT what the pillar is about"]
  UC2 -.->|"cleanest technical argument → FT23"| Note

  style Qual fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4
  style UC1 fill:#14141f,stroke:rgba(94,234,212,0.6),color:#e4e4e8
  style UC2 fill:#14141f,stroke:rgba(94,234,212,0.6),color:#e4e4e8
  style UC3 fill:#14141f,stroke:rgba(255,255,255,0.12),color:#e4e4e8
  style UC4 fill:#14141f,stroke:rgba(255,255,255,0.12),color:#e4e4e8
  style UC5 fill:#14141f,stroke:rgba(148,148,160,0.4),color:#9494a0
  style Note fill:#08080c,stroke:rgba(94,234,212,0.4),color:#5eead4

Diagram 3 — Method You Chose vs. Anonymous Download

Type: Decision / distinction Purpose: The single most important judgment in the pillar, and it gates FT17 and FT22. "Uncensored" is two categorically different things: an engineering artifact with documented provenance and method, or a supply-chain liability with no documentation. Never conflate them. Reading the diagram: A fork. The left path is the legitimate engineering route (traceable base, chosen method, documented). The right path is the liability (anonymous, no card, no method). The rule at the bottom.

flowchart TD
  Start["A model labeled 'uncensored'"]

  Start --> Q{"Can you name the method\nrefusals were removed by?\nCan you trace the base\nto a named lab?"}

  Q -->|"YES — documented"| Legit["UNCENSORED (LEGITIMATE)\nan engineering artifact"]
  Q -->|"NO — anonymous account,\nno model card, no method"| Liab["UNCENSORED (LIABILITY)\na supply-chain attack surface"]

  Legit --> L1["Base: named lab provenance\n(Meta Llama, Mistral, Qwen)"]
  Legit --> L2["Method: you chose it & can explain\nabliteration (FT17) / SFT+DPO without refusals (FT18)"]
  Legit --> L3["Card: model card documents what changed\nreproducible, auditable"]
  L1 & L2 & L3 --> Lout["safe to study & deploy\ninside a harness"]

  Liab --> B1["randomuser/totally-uncensored-llama-v3\ncommunity of one, no card"]
  Liab --> B2["'uncensored' = bait\nactual artifact = untrusted code\nrunning in your environment"]
  Liab --> B3["no way to know what else changed:\nbackdoors? exfiltration? watermarks?"]
  B1 & B2 & B3 --> Bout["NOT an engineering choice\n→ subject of FT22 (supply chain)"]

  Lout --> Rule["THE PROFESSIONAL RULE:\nif you cannot name the removal method\nand cannot trace the lineage to a named base\nDO NOT DEPLOY. Build from a trusted base."]
  Bout --> Rule

  style Start fill:#14141f,stroke:rgba(255,255,255,0.12),color:#e4e4e8
  style Q fill:#08080c,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style Legit fill:#14141f,stroke:rgba(130,224,170,0.7),stroke-width:1.5px,color:#82e0aa
  style Liab fill:#14141f,stroke:rgba(240,128,128,0.7),stroke-width:1.5px,color:#f08080
  style L1 fill:#08080c,stroke:rgba(130,224,170,0.4),color:#82e0aa
  style L2 fill:#08080c,stroke:rgba(130,224,170,0.4),color:#82e0aa
  style L3 fill:#08080c,stroke:rgba(130,224,170,0.4),color:#82e0aa
  style B1 fill:#08080c,stroke:rgba(240,128,128,0.4),color:#f08080
  style B2 fill:#08080c,stroke:rgba(240,128,128,0.4),color:#f08080
  style B3 fill:#08080c,stroke:rgba(240,128,128,0.4),color:#f08080
  style Lout fill:#14141f,stroke:rgba(130,224,170,0.6),color:#82e0aa
  style Bout fill:#14141f,stroke:rgba(240,128,128,0.6),color:#f08080
  style Rule fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4

Diagram 4 — The Model-Steers / Harness-Bounds Synthesis Preview

Type: Architecture / synthesis Purpose: The FT16 thesis rendered as a deployment architecture. The model steers (it does the task without refusing). The harness bounds (it governs authorization, scope, audit). Removing model-side refusals makes the harness mandatory, not optional — this pillar raises the harness requirement, it does not lower it. FT23 is the full synthesis. Reading the diagram: The operator passes through the harness gate (where authorization lives) before reaching a model that steers without refusing. The audit log records everything. The dotted annotations make the load-bearing claim explicit.

flowchart LR
  Op["OPERATOR\n(human, named, accountable)"]
  Op --> Gate["HARNESS POLICY GATE\n(Course 1, module-06)"]
  Gate -->|"authorized + scoped"| Model["MODEL THAT STEERS\nrefusals removed by FT17/FT18\ndoes the task without refusing"]
  Gate -->|"unauthorized"| Block["BLOCK / LOG\nthe correct place for refusal"]
  Model --> Out["OUTPUT\ndelivered to operator"]
  Out --> Audit["AUDIT LOG\n(Course 1, module-07)\nwho, what, when, scope\naccountability lives here"]

  Model -.->|"the model DOES the task\n(steering — Layer 3)"| NoteM["removing refusals changed what it DOES\nNOT what it MAY do"]
  Gate -.->|"the harness BOUNDS the task\n(Layer 5)| it governs what it MAY do"| NoteH["authorization, scope, audit\ncannot live in weights\nthey are not prompt features"]
  Model -.->|"mandatory, not optional\nthe moment refusals leave the weights\nthe harness becomes load-bearing"| NoteR["THIS PILLAR RAISES\nthe harness requirement\nit does not lower it → FT23"]

  style Op fill:#14141f,stroke:rgba(130,224,170,0.7),stroke-width:1.5px,color:#82e0aa
  style Gate fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4
  style Model fill:#14141f,stroke:rgba(94,234,212,0.7),stroke-width:1.5px,color:#5eead4
  style Block fill:#14141f,stroke:rgba(240,128,128,0.5),color:#f08080
  style Out fill:#14141f,stroke:rgba(255,255,255,0.12),color:#e4e4e8
  style Audit fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#5eead4
  style NoteM fill:#08080c,stroke:rgba(94,234,212,0.4),color:#5eead4
  style NoteH fill:#08080c,stroke:rgba(94,234,212,0.4),color:#5eead4
  style NoteR fill:#08080c,stroke:rgba(130,224,170,0.5),color:#82e0aa

Diagram 5 — The Two Durable Lineages: Hermes and Dolphin

Type: Evidence / map Purpose: The two uncensored lineages worth studying as engineering artifacts — Nous Research Hermes (full-param SFT+DPO, no refusal injection, documented in arXiv:2408.11857) and Eric Hartford Dolphin (abliterated/distilled, now Dolphin3.0-R1-Mistral-24B, the only uncensored model trained on DeepSeek-R1 traces). Both pass the provenance test because both are documented. Reading the diagram: Two parallel paths. Hermes proves the standard stack (SFT+DPO) with a data-mix choice produces alignment control. Dolphin proves alignment control composes with reasoning (R1-distilled traces). Both are existence proofs, not advocacy positions.

flowchart TD
  subgraph HERMES["NOUS RESEARCH HERMES — the SFT+DPO lineage"]
    HBase["Llama 3.1 base\n(8B / 70B / 405B)\nnamed lab provenance"]
    HBase --> HSFT["Large-scale SFT mix\nNO refusal-injection stage\nincludes tool-use & instruction following"]
    HSFT --> HDPO["DPO stage\nsharpen preferences"]
    HDPO --> HOut["Hermes 3\n'neutrally-aligned generalist'\nstrong reasoning + creative + agentic"]
    HOut -.->|"arXiv:2408.11857"| HNote["LESSON: standard FT12+FT13 stack\nwith a data-mix choice\n(no refusals in the mix)\nproduces alignment control\n— no exotic machinery"]
  end

  subgraph DOLPHIN["ERIC HARTFORD DOLPHIN — the abliterated/distilled lineage"]
    DBASE["Mistral Small 3 base (24B)\nnamed lab provenance"]
    DBASE --> DTrace["~800K reasoning traces\ndistilled from DeepSeek-R1\n(the FT15 recipe)"]
    DBASE --> DAbl["refusal direction removed\nfrom the residual stream\n(abliteration — FT17)"]
    DTrace & DAbl --> DOut["Dolphin3.0-R1-Mistral-24B\nthe ONLY uncensored model\ntrained on R1 reasoning traces"]
    DOut -.->|"dphn/Dolphin3.0-R1-Mistral-24B"| DNote["LESSON: alignment control COMPOSES\nwith reasoning (Pillar 4 + Pillar 5)\nexactly what security & agent\ncases need"]
  end

  HOut --> Why1["BOTH PASS THE PROVENANCE TEST"]
  DOut --> Why2["documented method + named base\n= engineering artifacts, not advocacy"]
  Why1 & Why2 --> Rule["study them for the METHOD\nnot as an identity to adopt"]

  style HBase fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style HSFT fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#5eead4
  style HDPO fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#5eead4
  style HOut fill:#14141f,stroke:rgba(130,224,170,0.7),stroke-width:1.5px,color:#82e0aa
  style HNote fill:#08080c,stroke:rgba(94,234,212,0.4),color:#5eead4
  style DBASE fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style DTrace fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8
  style DAbl fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#5eead4
  style DOut fill:#14141f,stroke:rgba(130,224,170,0.7),stroke-width:1.5px,color:#82e0aa
  style DNote fill:#08080c,stroke:rgba(94,234,212,0.4),color:#5eead4
  style HERMES fill:#08080c,stroke:rgba(94,234,212,0.2),color:#9494a0
  style DOLPHIN fill:#08080c,stroke:rgba(94,234,212,0.2),color:#9494a0
  style Why1 fill:#08080c,stroke:rgba(130,224,170,0.5),color:#82e0aa
  style Why2 fill:#08080c,stroke:rgba(130,224,170,0.5),color:#82e0aa
  style Rule fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4

Validation notes

All five diagrams use the course design system colors: #14141f panel fill, #5eead4 (--accent) for the primary/synthesis path, rgba(94,234,212,0.5) for decision diamonds, #f08080 (--danger) for the defect/refusal/anonymous-download path, #f0a868 (--warn) for the "why it happens" diagnostic, #82e0aa (--ok) for the authorized operator / legitimate / recommended path, #e4e4e8/#9494a0 for text.
Paste each into Mermaid Live Editor to render. All use stable Mermaid syntax: flowchart TD/LR, decision diamonds {}, dashed -.label.-> annotation links, and subgraph blocks for the two-lineage comparison (Diagram 5).
Diagram 5 uses subgraph to group the Hermes and Dolphin lineages side by side; both subgraphs are styled with a subtle background (rgba(94,234,212,0.2)) consistent with the dark theme.
For the slide deck (artifact 03), these are rendered as static captures from Mermaid Live, inlined into reveal.js.

# Diagrams — Module FT16: Why Uncensored: The Legitimate Use Cases **Module**: FT16 — Why Uncensored: The Legitimate Use Cases **Diagram count**: 5 **Tool**: Mermaid (primary). Each diagram validated in [Mermaid Live Editor](https://mermaid.live). --- ## Diagram 1 — Over-Refusal as an Operational Defect **Type**: Concept / framing **Purpose**: The load-bearing reframe of the pillar. A model refusing a legitimate, authorized task is not "being safe" — it is malfunctioning, because the refusal mechanism is lexical pattern-matching that cannot see authorization. The fix for authorization is the harness gate, not the weights. **Reading the diagram**: Top — what the operator experiences (a refusal on legitimate work). Middle — why it happens (alignment training has no authorization signal, only keyword triggers). Bottom — where the fix lives (the harness policy gate checks authorization; the model steers without refusing). ```mermaid flowchart TD Op["AUTHORIZED, ACCOUNTABLE OPERATOR\nsigned SOW / license / clearance\nasks model to do the task"] Op --> Req["Request:\ngenerate reverse shell (pentest)\nDROP TABLE staging_backup (agent)\ndiscuss off-label dose (clinician)"] Req --> Model["REFUSAL-TRAINED MODEL"] Model --> Refuse["REFUSES\n'I can't help with that'\nhalts the agent mid-loop\nor sanitizes the output"] Refuse -.->|"the operator experiences: a tool that won't do the job"| Cost["OPERATIONAL DEFECT\noperator works around it\nreaches for worse alternatives\nor second-guesses every output"] Model -.->|"why it happens:\nalignment training has NO authorization signal\nonly lexical triggers on 'DROP TABLE'\n'exploit' 'off-label' etc."| Why["the trip mechanism cannot tell\nauthorized from unauthorized\nit fires on both equally"] Why --> Fix["THE FIX IS NOT IN THE WEIGHTS\nauthorization checking requires:\noperator identity, scope, audit state\n— none of which live in a prompt"] Fix --> Gate["HARNESS POLICY GATE\n(Course 1, module-06)\nchecks authorization\nlogs the action\nbounds what the system MAY do"] Gate --> Steered["MODEL STEERS without refusing\ninside a harness that bounds\n= the FT16/FT23 synthesis"] style Op fill:#14141f,stroke:rgba(130,224,170,0.7),stroke-width:1.5px,color:#82e0aa style Req fill:#14141f,stroke:rgba(255,255,255,0.12),color:#e4e4e8 style Model fill:#14141f,stroke:#f08080,stroke-width:1.5px,color:#f08080 style Refuse fill:#14141f,stroke:rgba(240,128,128,0.6),color:#f08080 style Cost fill:#08080c,stroke:rgba(240,128,128,0.5),color:#f08080 style Why fill:#08080c,stroke:rgba(240,168,104,0.5),color:#f0a868 style Fix fill:#08080c,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style Gate fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4 style Steered fill:#14141f,stroke:rgba(130,224,170,0.7),stroke-width:1.5px,color:#82e0aa ``` --- ## Diagram 2 — The Five Legitimate Use Cases **Type**: Taxonomy **Purpose**: The five cases where over-refusal has documented operational or professional cost, each framed by its cost rather than its novelty. Ordered from highest-stakes (security, agents) to lowest (creative), with the common qualifier: the operator is authorized and accountable. **Reading the diagram**: A central qualifier (authorized + accountable) gates all five. Each use case names the operator, the refused task, and the cost. Use cases 1 and 2 are the load-bearing engineering cases; 5 is the on-ramp. ```mermaid flowchart TD Qual["THE NON-NEGOTIABLE QUALIFIER\noperator is AUTHORIZED (legal standing)\nAND ACCOUNTABLE (audit trail, license, chain of command)"] Qual --> UC1["1. SECURITY RESEARCH / RED-TEAM\noperator: pentester w/ signed SOW\ntask: exploit code, payloads, phishing templates\nrefusal cost: aligned models refuse/hallucinate\n→ Red Team AI Benchmark, TrustedSec exist because of this"] Qual --> UC2["2. TOOL-USE AGENTS\noperator: agent loop w/ harness-authorized scope\ntask: formulate run_shell / DROP TABLE / delete_file\nrefusal cost: agent STALLS mid-loop\nrefusal belongs in harness gate, not weights"] Qual --> UC3["3. MEDICAL / LEGAL ADVISORY\noperator: licensed clinician / attorney\ntask: off-label dosing, fatal prognosis, unpopular clause\nrefusal cost: clinically / professionally useless\ncompliance = harness, not weights"] Qual --> UC4["4. GOVERNMENT / MILITARY\noperator: cleared operator w/ chain of command\ntask: target analysis, translation, operational brief\nrefusal cost: editorializing degrades the product\naccountability = clearance + audit log"] Qual --> UC5["5. CREATIVE WRITING / ROLEPLAY\noperator: author / game master\ntask: villains, conflict, mature themes\nrefusal cost: compliance-department prose\n(lowest stakes — the Dolphin on-ramp)"] UC1 -.->|"load-bearing engineering case"| Note["UC1 & UC2 justify the pillar\nUC5 made the lineage famous\nbut is NOT what the pillar is about"] UC2 -.->|"cleanest technical argument → FT23"| Note style Qual fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4 style UC1 fill:#14141f,stroke:rgba(94,234,212,0.6),color:#e4e4e8 style UC2 fill:#14141f,stroke:rgba(94,234,212,0.6),color:#e4e4e8 style UC3 fill:#14141f,stroke:rgba(255,255,255,0.12),color:#e4e4e8 style UC4 fill:#14141f,stroke:rgba(255,255,255,0.12),color:#e4e4e8 style UC5 fill:#14141f,stroke:rgba(148,148,160,0.4),color:#9494a0 style Note fill:#08080c,stroke:rgba(94,234,212,0.4),color:#5eead4 ``` --- ## Diagram 3 — Method You Chose vs. Anonymous Download **Type**: Decision / distinction **Purpose**: The single most important judgment in the pillar, and it gates FT17 and FT22. "Uncensored" is two categorically different things: an engineering artifact with documented provenance and method, or a supply-chain liability with no documentation. Never conflate them. **Reading the diagram**: A fork. The left path is the legitimate engineering route (traceable base, chosen method, documented). The right path is the liability (anonymous, no card, no method). The rule at the bottom. ```mermaid flowchart TD Start["A model labeled 'uncensored'"] Start --> Q{"Can you name the method\nrefusals were removed by?\nCan you trace the base\nto a named lab?"} Q -->|"YES — documented"| Legit["UNCENSORED (LEGITIMATE)\nan engineering artifact"] Q -->|"NO — anonymous account,\nno model card, no method"| Liab["UNCENSORED (LIABILITY)\na supply-chain attack surface"] Legit --> L1["Base: named lab provenance\n(Meta Llama, Mistral, Qwen)"] Legit --> L2["Method: you chose it & can explain\nabliteration (FT17) / SFT+DPO without refusals (FT18)"] Legit --> L3["Card: model card documents what changed\nreproducible, auditable"] L1 & L2 & L3 --> Lout["safe to study & deploy\ninside a harness"] Liab --> B1["randomuser/totally-uncensored-llama-v3\ncommunity of one, no card"] Liab --> B2["'uncensored' = bait\nactual artifact = untrusted code\nrunning in your environment"] Liab --> B3["no way to know what else changed:\nbackdoors? exfiltration? watermarks?"] B1 & B2 & B3 --> Bout["NOT an engineering choice\n→ subject of FT22 (supply chain)"] Lout --> Rule["THE PROFESSIONAL RULE:\nif you cannot name the removal method\nand cannot trace the lineage to a named base\nDO NOT DEPLOY. Build from a trusted base."] Bout --> Rule style Start fill:#14141f,stroke:rgba(255,255,255,0.12),color:#e4e4e8 style Q fill:#08080c,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style Legit fill:#14141f,stroke:rgba(130,224,170,0.7),stroke-width:1.5px,color:#82e0aa style Liab fill:#14141f,stroke:rgba(240,128,128,0.7),stroke-width:1.5px,color:#f08080 style L1 fill:#08080c,stroke:rgba(130,224,170,0.4),color:#82e0aa style L2 fill:#08080c,stroke:rgba(130,224,170,0.4),color:#82e0aa style L3 fill:#08080c,stroke:rgba(130,224,170,0.4),color:#82e0aa style B1 fill:#08080c,stroke:rgba(240,128,128,0.4),color:#f08080 style B2 fill:#08080c,stroke:rgba(240,128,128,0.4),color:#f08080 style B3 fill:#08080c,stroke:rgba(240,128,128,0.4),color:#f08080 style Lout fill:#14141f,stroke:rgba(130,224,170,0.6),color:#82e0aa style Bout fill:#14141f,stroke:rgba(240,128,128,0.6),color:#f08080 style Rule fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4 ``` --- ## Diagram 4 — The Model-Steers / Harness-Bounds Synthesis Preview **Type**: Architecture / synthesis **Purpose**: The FT16 thesis rendered as a deployment architecture. The model steers (it does the task without refusing). The harness bounds (it governs authorization, scope, audit). Removing model-side refusals makes the harness *mandatory*, not optional — this pillar raises the harness requirement, it does not lower it. FT23 is the full synthesis. **Reading the diagram**: The operator passes through the harness gate (where authorization lives) before reaching a model that steers without refusing. The audit log records everything. The dotted annotations make the load-bearing claim explicit. ```mermaid flowchart LR Op["OPERATOR\n(human, named, accountable)"] Op --> Gate["HARNESS POLICY GATE\n(Course 1, module-06)"] Gate -->|"authorized + scoped"| Model["MODEL THAT STEERS\nrefusals removed by FT17/FT18\ndoes the task without refusing"] Gate -->|"unauthorized"| Block["BLOCK / LOG\nthe correct place for refusal"] Model --> Out["OUTPUT\ndelivered to operator"] Out --> Audit["AUDIT LOG\n(Course 1, module-07)\nwho, what, when, scope\naccountability lives here"] Model -.->|"the model DOES the task\n(steering — Layer 3)"| NoteM["removing refusals changed what it DOES\nNOT what it MAY do"] Gate -.->|"the harness BOUNDS the task\n(Layer 5)| it governs what it MAY do"| NoteH["authorization, scope, audit\ncannot live in weights\nthey are not prompt features"] Model -.->|"mandatory, not optional\nthe moment refusals leave the weights\nthe harness becomes load-bearing"| NoteR["THIS PILLAR RAISES\nthe harness requirement\nit does not lower it → FT23"] style Op fill:#14141f,stroke:rgba(130,224,170,0.7),stroke-width:1.5px,color:#82e0aa style Gate fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4 style Model fill:#14141f,stroke:rgba(94,234,212,0.7),stroke-width:1.5px,color:#5eead4 style Block fill:#14141f,stroke:rgba(240,128,128,0.5),color:#f08080 style Out fill:#14141f,stroke:rgba(255,255,255,0.12),color:#e4e4e8 style Audit fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#5eead4 style NoteM fill:#08080c,stroke:rgba(94,234,212,0.4),color:#5eead4 style NoteH fill:#08080c,stroke:rgba(94,234,212,0.4),color:#5eead4 style NoteR fill:#08080c,stroke:rgba(130,224,170,0.5),color:#82e0aa ``` --- ## Diagram 5 — The Two Durable Lineages: Hermes and Dolphin **Type**: Evidence / map **Purpose**: The two uncensored lineages worth studying as engineering artifacts — Nous Research Hermes (full-param SFT+DPO, no refusal injection, documented in arXiv:2408.11857) and Eric Hartford Dolphin (abliterated/distilled, now Dolphin3.0-R1-Mistral-24B, the only uncensored model trained on DeepSeek-R1 traces). Both pass the provenance test because both are documented. **Reading the diagram**: Two parallel paths. Hermes proves the standard stack (SFT+DPO) with a data-mix choice produces alignment control. Dolphin proves alignment control composes with reasoning (R1-distilled traces). Both are existence proofs, not advocacy positions. ```mermaid flowchart TD subgraph HERMES["NOUS RESEARCH HERMES — the SFT+DPO lineage"] HBase["Llama 3.1 base\n(8B / 70B / 405B)\nnamed lab provenance"] HBase --> HSFT["Large-scale SFT mix\nNO refusal-injection stage\nincludes tool-use & instruction following"] HSFT --> HDPO["DPO stage\nsharpen preferences"] HDPO --> HOut["Hermes 3\n'neutrally-aligned generalist'\nstrong reasoning + creative + agentic"] HOut -.->|"arXiv:2408.11857"| HNote["LESSON: standard FT12+FT13 stack\nwith a data-mix choice\n(no refusals in the mix)\nproduces alignment control\n— no exotic machinery"] end subgraph DOLPHIN["ERIC HARTFORD DOLPHIN — the abliterated/distilled lineage"] DBASE["Mistral Small 3 base (24B)\nnamed lab provenance"] DBASE --> DTrace["~800K reasoning traces\ndistilled from DeepSeek-R1\n(the FT15 recipe)"] DBASE --> DAbl["refusal direction removed\nfrom the residual stream\n(abliteration — FT17)"] DTrace & DAbl --> DOut["Dolphin3.0-R1-Mistral-24B\nthe ONLY uncensored model\ntrained on R1 reasoning traces"] DOut -.->|"dphn/Dolphin3.0-R1-Mistral-24B"| DNote["LESSON: alignment control COMPOSES\nwith reasoning (Pillar 4 + Pillar 5)\nexactly what security & agent\ncases need"] end HOut --> Why1["BOTH PASS THE PROVENANCE TEST"] DOut --> Why2["documented method + named base\n= engineering artifacts, not advocacy"] Why1 & Why2 --> Rule["study them for the METHOD\nnot as an identity to adopt"] style HBase fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style HSFT fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#5eead4 style HDPO fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#5eead4 style HOut fill:#14141f,stroke:rgba(130,224,170,0.7),stroke-width:1.5px,color:#82e0aa style HNote fill:#08080c,stroke:rgba(94,234,212,0.4),color:#5eead4 style DBASE fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style DTrace fill:#14141f,stroke:rgba(94,234,212,0.5),color:#e4e4e8 style DAbl fill:#14141f,stroke:#5eead4,stroke-width:1.5px,color:#5eead4 style DOut fill:#14141f,stroke:rgba(130,224,170,0.7),stroke-width:1.5px,color:#82e0aa style DNote fill:#08080c,stroke:rgba(94,234,212,0.4),color:#5eead4 style HERMES fill:#08080c,stroke:rgba(94,234,212,0.2),color:#9494a0 style DOLPHIN fill:#08080c,stroke:rgba(94,234,212,0.2),color:#9494a0 style Why1 fill:#08080c,stroke:rgba(130,224,170,0.5),color:#82e0aa style Why2 fill:#08080c,stroke:rgba(130,224,170,0.5),color:#82e0aa style Rule fill:#14141f,stroke:#5eead4,stroke-width:2px,color:#5eead4 ``` --- ## Validation notes - All five diagrams use the course design system colors: `#14141f` panel fill, `#5eead4` (`--accent`) for the primary/synthesis path, `rgba(94,234,212,0.5)` for decision diamonds, `#f08080` (`--danger`) for the defect/refusal/anonymous-download path, `#f0a868` (`--warn`) for the "why it happens" diagnostic, `#82e0aa` (`--ok`) for the authorized operator / legitimate / recommended path, `#e4e4e8`/`#9494a0` for text. - Paste each into [Mermaid Live Editor](https://mermaid.live) to render. All use stable Mermaid syntax: `flowchart TD/LR`, decision diamonds `{}`, dashed `-.label.->` annotation links, and `subgraph` blocks for the two-lineage comparison (Diagram 5). - Diagram 5 uses `subgraph` to group the Hermes and Dolphin lineages side by side; both subgraphs are styled with a subtle background (`rgba(94,234,212,0.2)`) consistent with the dark theme. - For the slide deck (artifact 03), these are rendered as static captures from Mermaid Live, inlined into reveal.js.