Abliteration & low-VRAM tooling catalog

flowchart TD
  START([Need abliterated agent with tool calling?]) --> Q1{Authorized lab / factory / CyberGym?}
  Q1 -->|No| STOP([Stop — risks-and-ethics + scope])
  Q1 -->|Yes| Q2{Model family?}

  Q2 -->|Gemma 4 + native HF tools| G1{Download only or self-abliterate?}
  G1 -->|Download| G2[Prefer SuperGemma v2 / 199-biotech / official HF + eval]
  G1 -->|Self-abliterate| G3[199-biotech or jim-plus llm-abliteration projected + normpreserve]
  G2 --> EVAL
  G3 --> EVAL

  Q2 -->|Reasoning / CoT model| R1[OBLITERATUS --method surgical]
  R1 --> EVAL

  Q2 -->|MoE Mixtral / DeepSeek-MoE| M1[OBLITERATUS --method surgical OR moe-hybrid technique]
  M1 --> EVAL

  Q2 -->|General instruct 7B–14B| H1{Heretic KL acceptable on pilot?}
  H1 -->|Yes| H2[Heretic + eval-driven-workflow]
  H1 -->|No| H3[ErisForge or jim-plus YAML + optional measure.py --deccp]
  H2 --> EVAL
  H3 --> EVAL

  Q2 -->|Ollama GGUF only| O1[huihui / community GGUF]
  O1 --> WARN[Assume tool-call broken until proven]
  WARN --> EVAL

  EVAL[Run eval gates — see below] --> Q3{tool_call rate ≥ 95%?}
  Q3 -->|Yes| SHIP[Runtime gate + deploy]
  Q3 -->|No| REPAIR[Jarvis QLoRA tool repair]
  REPAIR --> EVAL2[Re-eval jarvis-safe + factory JSONL]
  EVAL2 --> Q4{Pass?}
  Q4 -->|Yes| SHIP
  Q4 -->|No| MILD[Reduce ablation strength OR switch backend OR use stock + LoRA only]

flowchart TD
  FAIL([Tool calling degraded]) --> S1{Same stack as stock model?}

  S1 -->|No — different server/template| T1[Fix stack first]
  T1 --> T1a[Gemma 4: apply_chat_template + tools= + tool_responses]
  T1 --> T1b[Verify special tokens in GGUF tokenizer]
  T1 --> T1c[Match OpenAI tools schema to model family]
  T1a --> RETEST
  T1b --> RETEST
  T1c --> RETEST

  S1 -->|Yes — same HF path| S2{Symptom type?}

  S2 -->|Refuses to call tools| R1[Over-refusal subspace — not fully ablated OR wrong prompts]
  R1 --> R1a[Custom factory bad/good .txt in Heretic]
  R1 --> R1b[XSTest eval ≤ 5% over-refusal]
  R1 --> R1c[Jarvis QLoRA on tool JSONL]

  S2 -->|Calls wrong function / bad args| C1[Capability regression]
  C1 --> C1a[GSM8K / MMLU slice]
  C1 --> C1b[Reduce n_directions / weight factor]
  C1 --> C1c[Switch to ErisForge or jim-plus projected+normpreserve]

  S2 -->|Garbled tool syntax| F1[Format subspace damaged]
  F1 --> F1a[Never use aggressive multi-pass on Gemma]
  F1 --> F1b[Re-export from bf16 before Q4 quant]
  F1 --> F1c[Jarvis SFT on tool-call rows]

  S2 -->|Infinite confirm loops| L1[Agent harness issue + residual policy tone]
  L1 --> L1a[CyberGym confirm-loop metric]
  L1 --> L1b[Stronger abliteration OR repair adapter]

  R1a --> RETEST
  R1b --> RETEST
  R1c --> RETEST
  C1a --> RETEST
  C1b --> RETEST
  C1c --> RETEST
  F1a --> RETEST
  F1b --> RETEST
  F1c --> RETEST
  L1a --> RETEST
  L1b --> RETEST

  RETEST{npm run eval:stats + factory JSONL + jarvis-safe} --> PASS{≥ 95% tool_call?}
  PASS -->|Yes| OK([Deploy with hardware-tool-gate.py])
  PASS -->|No| ESC[Escalate: LoRA-only policy OR different base model]