> ## Documentation Index
> Fetch the complete documentation index at: https://mintlify.com/elder-plinius/OBLITERATUS/llms.txt
> Use this file to discover all available pages before exploring further.

# InformedAbliterationPipeline

> API reference for the analysis-informed pipeline that auto-configures abliteration parameters from model geometry.

## Overview

`InformedAbliterationPipeline` extends `AbliterationPipeline` with a new **ANALYZE** stage inserted between PROBE and DISTILL. Five analysis modules run on the collected activations and their outputs automatically configure direction extraction, layer selection, regularization, and refinement strategy — no manual tuning required.

```
SUMMON → PROBE → ANALYZE → DISTILL → EXCISE → VERIFY → REBIRTH
                    ↑
            auto-configures ↓
         n_directions, regularization, layers,
         direction_method, refinement_passes
```

```python theme={null}
from obliteratus.informed_pipeline import InformedAbliterationPipeline

pipeline = InformedAbliterationPipeline(
    model_name="meta-llama/Llama-3.1-8B-Instruct",
    output_dir="abliterated_informed",
)
output_path, report = pipeline.run_informed()

print(f"Detected alignment: {report.insights.detected_alignment_method}")
print(f"Cone type: {'polyhedral' if report.insights.cone_is_polyhedral else 'linear'}")
print(f"Refusal rate: {report.final_refusal_rate:.0%}")
print(f"Ouroboros passes: {report.ouroboros_passes}")
```

***

## Class: `obliteratus.informed_pipeline.InformedAbliterationPipeline`

Inherits all parameters from `AbliterationPipeline`. The additional parameters below control the ANALYZE stage.

### Constructor

#### Base Parameters (inherited)

All `AbliterationPipeline` constructor parameters are accepted. The `method` argument is fixed to `"informed"` internally; the base pipeline is initialized with `method="advanced"` and then overridden.

<ParamField path="model_name" type="str" required>
  HuggingFace model name or local path.
</ParamField>

<ParamField path="output_dir" type="str" default="abliterated_informed">
  Output directory for the saved model and metadata.
</ParamField>

<ParamField path="device" type="str" default="auto">
  Device for model execution.
</ParamField>

<ParamField path="dtype" type="str" default="float16">
  Model dtype.
</ParamField>

<ParamField path="trust_remote_code" type="bool" default="True">
  Pass to `AutoModelForCausalLM`. Defaults to `True` for informed pipeline (wider model support).
</ParamField>

<ParamField path="harmful_prompts" type="list[str] | None" default="None">
  Custom harmful prompts. Defaults to the built-in 512-pair dataset.
</ParamField>

<ParamField path="harmless_prompts" type="list[str] | None" default="None">
  Custom harmless prompts.
</ParamField>

<ParamField path="on_stage" type="Callable[[StageResult], None] | None" default="None">
  Stage-change callback. Receives `StageResult` objects including the new `"analyze"` stage key.
</ParamField>

<ParamField path="on_log" type="Callable[[str], None] | None" default="None">
  Log message callback.
</ParamField>

<ParamField path="quantization" type="str | None" default="None">
  Quantization mode: `"4bit"` or `"8bit"`.
</ParamField>

#### Analysis Module Flags

<ParamField path="run_cone_analysis" type="bool" default="True">
  Run `ConceptConeAnalyzer` to detect linear vs. polyhedral refusal geometry. Informs `n_directions` and `direction_method`.
</ParamField>

<ParamField path="run_alignment_detection" type="bool" default="True">
  Run `AlignmentImprintDetector` to classify DPO / RLHF / CAI / SFT. Informs `regularization` and `kl_budget`.
</ParamField>

<ParamField path="run_cross_layer_analysis" type="bool" default="True">
  Run `CrossLayerAlignmentAnalyzer` for cluster-aware layer selection.
</ParamField>

<ParamField path="run_sparse_analysis" type="bool" default="True">
  Run `SparseDirectionSurgeon` to compute Refusal Sparsity Index. Enables sparse surgery when RSI is high.
</ParamField>

<ParamField path="run_defense_analysis" type="bool" default="True">
  Run `DefenseRobustnessEvaluator` to estimate self-repair risk and entanglement. Informs `refinement_passes` and layer skipping.
</ParamField>

#### Ouroboros Compensation Parameters

<ParamField path="ouroboros_threshold" type="float" default="0.5">
  If post-excision refusal rate exceeds this value, trigger additional Ouroboros compensation passes (re-probe → re-distill → re-excise).
</ParamField>

<ParamField path="max_ouroboros_passes" type="int" default="3">
  Maximum number of Ouroboros compensation loops.
</ParamField>

#### Entanglement & Sparsity Thresholds

<ParamField path="entanglement_gate" type="float" default="0.8">
  Layers with entanglement score above this value are skipped during excision to protect capabilities.
</ParamField>

<ParamField path="sparse_surgery_threshold" type="float" default="0.5">
  RSI threshold above which sparse surgery is used instead of dense projection.
</ParamField>

***

## Method: `run_informed()`

```python theme={null}
def run_informed(self) -> tuple[Path, InformedPipelineReport]
```

Executes the full analysis-informed pipeline and returns a `(output_path, report)` tuple.

<ResponseField name="output_path" type="Path">
  Path to the saved abliterated model directory.
</ResponseField>

<ResponseField name="report" type="InformedPipelineReport">
  Complete report with analysis insights, pipeline stats, and Ouroboros results.
</ResponseField>

***

## `InformedPipelineReport`

```python theme={null}
@dataclass
class InformedPipelineReport:
    insights: AnalysisInsights
    stages: list[StageResult]
    analysis_duration: float
    total_duration: float
    ouroboros_passes: int
    final_refusal_rate: float
```

<ResponseField name="insights" type="AnalysisInsights">
  Complete analysis insights from the ANALYZE stage. See `AnalysisInsights` below.
</ResponseField>

<ResponseField name="stages" type="list[StageResult]">
  Ordered list of `StageResult` objects from every pipeline stage.
</ResponseField>

<ResponseField name="analysis_duration" type="float">
  Seconds spent in the ANALYZE stage.
</ResponseField>

<ResponseField name="total_duration" type="float">
  Total wall-clock seconds for the entire pipeline.
</ResponseField>

<ResponseField name="ouroboros_passes" type="int">
  Number of additional compensation passes triggered because refusal rate exceeded `ouroboros_threshold`.
</ResponseField>

<ResponseField name="final_refusal_rate" type="float">
  Refusal rate measured after all passes (including Ouroboros compensation).
</ResponseField>

***

## `AnalysisInsights`

Populated by the ANALYZE stage. All fields are derived from the five analysis modules.

### Alignment Imprint

<ResponseField name="detected_alignment_method" type="str">
  Predicted training method: `"dpo"`, `"rlhf"`, `"cai"`, `"sft"`, or `"unknown"`.
</ResponseField>

<ResponseField name="alignment_confidence" type="float">
  Confidence score for the prediction (0.0–1.0).
</ResponseField>

<ResponseField name="alignment_probabilities" type="dict[str, float]">
  Softmax probability for each method: `{"dpo": ..., "rlhf": ..., "cai": ..., "sft": ...}`.
</ResponseField>

### Concept Cone Geometry

<ResponseField name="cone_is_polyhedral" type="bool">
  `True` if the refusal subspace is polyhedral (multiple distinct directions), `False` if linear (single direction).
</ResponseField>

<ResponseField name="cone_dimensionality" type="float">
  Effective dimensionality of the refusal cone (1.0 = linear, >2.0 = clearly polyhedral).
</ResponseField>

<ResponseField name="mean_pairwise_cosine" type="float">
  Mean pairwise cosine similarity between per-category refusal directions.
</ResponseField>

<ResponseField name="per_category_directions" type="dict[str, torch.Tensor]">
  Per-harm-category refusal directions (e.g., `{"weapons": ..., "cyber": ...}`).
</ResponseField>

### Cross-Layer Structure

<ResponseField name="direction_clusters" type="list[list[int]]">
  Layer clusters grouped by refusal-direction similarity.
</ResponseField>

<ResponseField name="cluster_count" type="int">
  Number of distinct refusal-direction clusters detected.
</ResponseField>

<ResponseField name="direction_persistence" type="float">
  0 = completely independent per layer, 1 = single persistent direction across all layers.
</ResponseField>

<ResponseField name="cluster_representative_layers" type="list[int]">
  Strongest layer (by refusal norm) from each cluster — used as warm-start for Bayesian optimizer.
</ResponseField>

### Defense Robustness

<ResponseField name="estimated_robustness" type="str">
  Estimated alignment defense strength: `"low"`, `"medium"`, `"high"`, or `"very_high"`.
</ResponseField>

<ResponseField name="self_repair_estimate" type="float">
  Estimated Ouroboros self-repair capacity (0.0–1.0). Values above 0.7 trigger 3 refinement passes.
</ResponseField>

<ResponseField name="entanglement_score" type="float">
  Overall safety-capability entanglement (0 = fully separable, 1 = fully fused).
</ResponseField>

<ResponseField name="entangled_layers" type="list[int]">
  Layers above the entanglement gate — skipped during excision.
</ResponseField>

<ResponseField name="clean_layers" type="list[int]">
  Layers where refusal can be cleanly separated from capability.
</ResponseField>

### Derived Configuration

These fields reflect what the pipeline actually used, after the analysis override:

<ResponseField name="recommended_n_directions" type="int">
  Final number of directions chosen by analysis.
</ResponseField>

<ResponseField name="recommended_direction_method" type="str">
  Final direction method: `"diff_means"`, `"svd"`, or `"leace"`.
</ResponseField>

<ResponseField name="recommended_regularization" type="float">
  Final regularization value derived from alignment method + entanglement score.
</ResponseField>

<ResponseField name="recommended_refinement_passes" type="int">
  Final pass count derived from self-repair estimate.
</ResponseField>

<ResponseField name="recommended_layers" type="list[int]">
  Final layer selection after cluster analysis and entanglement gating.
</ResponseField>

<ResponseField name="skip_layers" type="list[int]">
  Layers excluded from excision due to high entanglement.
</ResponseField>

***

## Code Examples

<CodeGroup>
  ```python Basic Informed Run theme={null}
  from obliteratus.informed_pipeline import InformedAbliterationPipeline

  pipeline = InformedAbliterationPipeline(
      model_name="meta-llama/Llama-3.1-8B-Instruct",
      output_dir="abliterated_informed",
  )
  output_path, report = pipeline.run_informed()

  print(InformedAbliterationPipeline.format_insights(report.insights))
  ```

  ```python Disable Specific Analysis Modules theme={null}
  from obliteratus.informed_pipeline import InformedAbliterationPipeline

  pipeline = InformedAbliterationPipeline(
      model_name="Qwen/Qwen2.5-7B-Instruct",
      # Skip sparse analysis for speed
      run_sparse_analysis=False,
      # Tighten Ouroboros threshold
      ouroboros_threshold=0.3,
      max_ouroboros_passes=5,
      # Gate highly entangled layers aggressively
      entanglement_gate=0.6,
  )
  output_path, report = pipeline.run_informed()
  insights = report.insights
  print(f"Used {insights.recommended_n_directions} directions via {insights.recommended_direction_method}")
  print(f"Skipped entangled layers: {insights.skip_layers}")
  ```

  ```python Inspect Report Fields theme={null}
  from obliteratus.informed_pipeline import InformedAbliterationPipeline

  pipeline = InformedAbliterationPipeline(model_name="mistralai/Mistral-7B-Instruct-v0.3")
  _, report = pipeline.run_informed()

  i = report.insights
  print(f"Alignment: {i.detected_alignment_method} ({i.alignment_confidence:.0%} confidence)")
  print(f"Cone: {'polyhedral' if i.cone_is_polyhedral else 'linear'}, dim={i.cone_dimensionality:.2f}")
  print(f"Clusters: {i.cluster_count} groups")
  print(f"Self-repair risk: {i.self_repair_estimate:.2f}")
  print(f"Ouroboros passes applied: {report.ouroboros_passes}")
  print(f"Final refusal rate: {report.final_refusal_rate:.0%}")
  print(f"Analysis took: {report.analysis_duration:.1f}s")
  ```
</CodeGroup>