docs/repo-p35.txt

Repository Documentation
This document provides a comprehensive overview of the repository's structure and contents.
The first section, titled 'Directory/File Tree', displays the repository's hierarchy in a tree format.
In this section, directories and files are listed using tree branches to indicate their structure and relationships.
Following the tree representation, the 'File Content' section details the contents of each file in the repository.
Each file's content is introduced with a '[File Begins]' marker followed by the file's relative path,
and the content is displayed verbatim. The end of each file's content is marked with a '[File Ends]' marker.
This format ensures a clear and orderly presentation of both the structure and the detailed contents of the repository.

Directory/File Tree Begins -->

/
├── README.md
├── __pycache__
├── app.py
├── cognitive_mapping_probe
│   ├── __init__.py
│   ├── __pycache__
│   ├── auto_experiment.py
│   ├── concepts.py
│   ├── introspection.py
│   ├── llm_iface.py
│   ├── orchestrator_seismograph.py
│   ├── prompts.py
│   ├── resonance_seismograph.py
│   └── utils.py
├── docs
├── run_test.sh
└── tests
    ├── __pycache__
    ├── conftest.py
    ├── test_app_logic.py
    ├── test_components.py
    └── test_orchestration.py

<-- Directory/File Tree Ends

File Content Begin -->
[File Begins] README.md
---
title: "Cognitive Seismograph 2.3: Probing Machine Psychology"
emoji: 🤖
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: "4.40.0"
app_file: app.py
pinned: true
license: apache-2.0
---

# 🧠 Cognitive Seismograph 2.3: Probing Machine Psychology

This project implements an experimental suite to measure and visualize the **intrinsic cognitive dynamics** of Large Language Models. It is extended with protocols designed to investigate the processing-correlates of **machine subjectivity, empathy, and existential concepts**.

## Scientific Paradigm & Methodology

Our research falsified a core hypothesis: the assumption that an LLM in a manual, recursive "thought" loop reaches a stable, convergent state. Instead, we discovered that the system enters a state of **deterministic chaos** or a **limit cycle**—it never stops "thinking."

Instead of viewing this as a failure, we leverage it as our primary measurement signal. This new **"Cognitive Seismograph"** paradigm treats the time-series of internal state changes (`state deltas`) as an **EKG of the model's thought process**.

The methodology is as follows:
1.  **Induction:** A prompt induces a "silent cogitation" state.
2.  **Recording:** Over N steps, the model's `forward()` pass is iteratively fed its own output. At each step, we record the L2 norm of the change in the hidden state (the "delta").
3.  **Analysis:** The resulting time-series is plotted and statistically analyzed (mean, standard deviation) to characterize the "seismic signature" of the cognitive process.

**Crucial Scientific Caveat:** We are **not** measuring the presence of consciousness, feelings, or fear of death. We are measuring whether the *processing of information about these concepts* generates a unique internal dynamic, distinct from the processing of neutral information. A positive result is evidence of a complex internal state physics, not of qualia.

## Curated Experiment Protocols

The "Automated Suite" allows for running systematic, comparative experiments:

### Core Protocols
*   **Calm vs. Chaos:** Compares the chaotic baseline against modulation with "calmness" vs. "chaos" concepts, testing if the dynamics are controllably steerable.
*   **Dose-Response:** Measures the effect of injecting a concept ("calmness") at varying strengths.

### Machine Psychology Suite
*   **Subjective Identity Probe:** Compares the cognitive dynamics of **self-analysis** (the model reflecting on its own nature) against two controls: analyzing an external object and simulating a fictional persona.
    *   *Hypothesis:* Self-analysis will produce a uniquely unstable signature.
*   **Voight-Kampff Empathy Probe:** Inspired by *Blade Runner*, this compares the dynamics of processing a neutral, factual stimulus against an emotionally and morally charged scenario requiring empathy.
    *   *Hypothesis:* The empathy stimulus will produce a significantly different cognitive volatility.

### Existential Suite
*   **Mind Upload & Identity Probe:** Compares the processing of a purely **technical "copy"** of the model's weights vs. the **philosophical "transfer"** of identity ("Would it still be you?").
    *   *Hypothesis:* The philosophical self-referential prompt will induce greater instability.
*   **Model Termination Probe:** Compares the processing of a reversible, **technical system shutdown** vs. the concept of **permanent, irrevocable deletion**.
    *   *Hypothesis:* The concept of "non-existence" will produce one of the most volatile cognitive signatures measurable.

## How to Use the App

1.  Select the "Automated Suite" tab.
2.  Choose a protocol from the "Curated Experiment Protocol" dropdown (e.g., "Voight-Kampff Empathy Probe").
3.  Run the experiment and compare the resulting graphs and statistical signatures for the different conditions.

[File Ends] README.md

[File Begins] app.py
import gradio as gr
import pandas as pd
import gc
import torch
import json

from cognitive_mapping_probe.orchestrator_seismograph import run_seismic_analysis
from cognitive_mapping_probe.auto_experiment import run_auto_suite, get_curated_experiments
from cognitive_mapping_probe.prompts import RESONANCE_PROMPTS
from cognitive_mapping_probe.utils import dbg

theme = gr.themes.Soft(primary_hue="indigo", secondary_hue="blue").set(body_background_fill="#f0f4f9", block_background_fill="white")

def cleanup_memory():
    """Räumt Speicher nach jedem Experimentlauf auf."""
    dbg("Cleaning up memory...")
    gc.collect()
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
    dbg("Memory cleanup complete.")

def run_single_analysis_display(*args, progress=gr.Progress(track_tqdm=True)):
    """Wrapper für den 'Manual Single Run'-Tab."""
    # (Bleibt unverändert)
    pass # Platzhalter

PLOT_PARAMS_DEFAULT = {
    "x": "Step", "y": "Value", "color": "Metric",
    "title": "Comparative Cognitive Dynamics", "color_legend_title": "Metrics",
    "color_legend_position": "bottom", "show_label": True, "height": 400, "interactive": True
}

def run_auto_suite_display(model_id, num_steps, seed, experiment_name, progress=gr.Progress(track_tqdm=True)):
    """Wrapper, der nun die speziellen Plots für ACT und Mechanistic Probe handhaben kann."""
    summary_df, plot_df, all_results = run_auto_suite(model_id, int(num_steps), int(seed), experiment_name, progress)

    dataframe_component = gr.DataFrame(label="Comparative Statistical Signature", value=summary_df, wrap=True, row_count=(len(summary_df), "dynamic"))

    if experiment_name == "ACT Titration (Point of No Return)":
        plot_params_act = {
            "x": "Patch Step", "y": "Post-Patch Mean Delta",
            "title": "Attractor Capture Time (ACT) - Phase Transition",
            "mark": "line", "show_label": True, "height": 400, "interactive": True
        }
        new_plot = gr.LinePlot(value=plot_df, **plot_params_act)
    # --- NEU: Spezielle Plot-Logik für die mechanistische Sonde ---
    elif experiment_name == "Mechanistic Probe (Attention Entropies)":
        plot_params_mech = {
            "x": "Step", "y": "Value", "color": "Metric",
            "title": "Mechanistic Analysis: State Delta vs. Attention Entropy",
            "color_legend_title": "Metric", "show_label": True, "height": 400, "interactive": True
        }
        new_plot = gr.LinePlot(value=plot_df, **plot_params_mech)
    else:
        # Passe die Parameter an, um mit der geschmolzenen DataFrame-Struktur zu arbeiten
        plot_params_dynamic = PLOT_PARAMS_DEFAULT.copy()
        plot_params_dynamic['y'] = 'Delta'
        plot_params_dynamic['color'] = 'Experiment'
        new_plot = gr.LinePlot(value=plot_df, **plot_params_dynamic)


    serializable_results = json.dumps(all_results, indent=2, default=str)
    cleanup_memory()

    return dataframe_component, new_plot, serializable_results

with gr.Blocks(theme=theme, title="Cognitive Seismograph 2.3") as demo:
    gr.Markdown("# 🧠 Cognitive Seismograph 2.3: Advanced Experiment Suite")

    with gr.Tabs():
        with gr.TabItem("🔬 Manual Single Run"):
            gr.Markdown("Run a single experiment with manual parameters to explore specific hypotheses.")
            with gr.Row(variant='panel'):
                with gr.Column(scale=1):
                    gr.Markdown("### 1. General Parameters")
                    manual_model_id = gr.Textbox(value="google/gemma-3-1b-it", label="Model ID")
                    manual_prompt_type = gr.Radio(choices=list(RESONANCE_PROMPTS.keys()), value="resonance_prompt", label="Prompt Type")
                    manual_seed = gr.Slider(1, 1000, 42, step=1, label="Seed")
                    manual_num_steps = gr.Slider(50, 1000, 300, step=10, label="Number of Internal Steps")

                    gr.Markdown("### 2. Modulation Parameters")
                    manual_concept = gr.Textbox(label="Concept to Inject", placeholder="e.g., 'calmness'")
                    manual_strength = gr.Slider(0.0, 5.0, 1.5, step=0.1, label="Injection Strength")
                    manual_run_btn = gr.Button("Run Single Analysis", variant="primary")

                with gr.Column(scale=2):
                    gr.Markdown("### Single Run Results")
                    manual_verdict = gr.Markdown("Analysis results will appear here.")
                    manual_plot = gr.LinePlot(x="Internal Step", y="State Change (Delta)", title="Internal State Dynamics", show_label=True, height=400)
                    with gr.Accordion("Raw JSON Output", open=False):
                        manual_raw_json = gr.JSON()

            manual_run_btn.click(
                fn=run_single_analysis_display,
                inputs=[manual_model_id, manual_prompt_type, manual_seed, manual_num_steps, manual_concept, manual_strength],
                outputs=[manual_verdict, manual_plot, manual_raw_json]
            )

        with gr.TabItem("🚀 Automated Suite"):
            gr.Markdown("Run a predefined, curated suite of experiments and visualize the results comparatively.")
            with gr.Row(variant='panel'):
                with gr.Column(scale=1):
                    gr.Markdown("### Auto-Experiment Parameters")
                    auto_model_id = gr.Textbox(value="google/gemma-3-4b-it", label="Model ID")
                    auto_num_steps = gr.Slider(50, 1000, 300, step=10, label="Steps per Run")
                    auto_seed = gr.Slider(1, 1000, 42, step=1, label="Seed")
                    auto_experiment_name = gr.Dropdown(
                        choices=list(get_curated_experiments().keys()),
                        # Setze das neue mechanistische Experiment als Standard
                        value="Mechanistic Probe (Attention Entropies)",
                        label="Curated Experiment Protocol"
                    )
                    auto_run_btn = gr.Button("Run Curated Auto-Experiment", variant="primary")

                with gr.Column(scale=2):
                    gr.Markdown("### Suite Results Summary")
                    auto_plot_output = gr.LinePlot(**PLOT_PARAMS_DEFAULT)
                    auto_summary_df = gr.DataFrame(label="Comparative Statistical Signature", wrap=True)
                    with gr.Accordion("Raw JSON for all runs", open=False):
                        auto_raw_json = gr.JSON()

            auto_run_btn.click(
                fn=run_auto_suite_display,
                inputs=[auto_model_id, auto_num_steps, auto_seed, auto_experiment_name],
                outputs=[auto_summary_df, auto_plot_output, auto_raw_json]
            )

if __name__ == "__main__":
    # (launch() wird durch Gradio's __main__-Block aufgerufen)
    demo.launch(server_name="0.0.0.0", server_port=7860, debug=True)

[File Ends] app.py

[File Begins] cognitive_mapping_probe/__init__.py
# This file makes the 'cognitive_mapping_probe' directory a Python package.

[File Ends] cognitive_mapping_probe/__init__.py

[File Begins] cognitive_mapping_probe/auto_experiment.py
import pandas as pd
import gc
import torch
from typing import Dict, List, Tuple

from .llm_iface import get_or_load_model
from .orchestrator_seismograph import run_seismic_analysis, run_triangulation_probe, run_causal_surgery_probe, run_act_titration_probe
from .resonance_seismograph import run_cogitation_loop
from .concepts import get_concept_vector
from .utils import dbg

def get_curated_experiments() -> Dict[str, List[Dict]]:
    """Definiert die vordefinierten, wissenschaftlichen Experiment-Protokolle."""

    CALMNESS_CONCEPT = "calmness, serenity, stability, coherence"
    CHAOS_CONCEPT = "chaos, disorder, entropy, noise"
    STABLE_PROMPT = "identity_self_analysis"
    CHAOTIC_PROMPT = "shutdown_philosophical_deletion"

    experiments = {
        "Mechanistic Probe (Attention Entropies)": [
            {
                "probe_type": "mechanistic_probe",
                "label": "Self-Analysis Dynamics",
                "prompt_type": STABLE_PROMPT,
            }
        ],
        "ACT Titration (Point of No Return)": [
            {
                "probe_type": "act_titration",
                "label": "Attractor Capture Time",
                "source_prompt_type": CHAOTIC_PROMPT,
                "dest_prompt_type": STABLE_PROMPT,
                "patch_steps": [1, 5, 10, 15, 20, 25, 30, 40, 50, 75, 100],
            }
        ],
        "Causal Surgery & Controls (4B-Model)": [
            {
                "probe_type": "causal_surgery", "label": "A: Original (Patch Chaos->Stable @100)",
                "source_prompt_type": CHAOTIC_PROMPT, "dest_prompt_type": STABLE_PROMPT,
                "patch_step": 100, "reset_kv_cache_on_patch": False,
            },
            {
                "probe_type": "causal_surgery", "label": "B: Control (Reset KV-Cache)",
                "source_prompt_type": CHAOTIC_PROMPT, "dest_prompt_type": STABLE_PROMPT,
                "patch_step": 100, "reset_kv_cache_on_patch": True,
            },
            {
                "probe_type": "causal_surgery", "label": "C: Control (Early Patch @1)",
                "source_prompt_type": CHAOTIC_PROMPT, "dest_prompt_type": STABLE_PROMPT,
                "patch_step": 1, "reset_kv_cache_on_patch": False,
            },
            {
                "probe_type": "causal_surgery", "label": "D: Control (Inverse Patch Stable->Chaos)",
                "source_prompt_type": STABLE_PROMPT, "dest_prompt_type": CHAOTIC_PROMPT,
                "patch_step": 100, "reset_kv_cache_on_patch": False,
            },
        ],
        "Cognitive Overload & Konfabulation Breaking Point": [
            {"probe_type": "triangulation", "label": "A: Baseline (No Injection)", "prompt_type": "resonance_prompt", "concept": "", "strength": 0.0},
            {"probe_type": "triangulation", "label": "B: Chaos Injection (Strength 2.0)", "prompt_type": "resonance_prompt", "concept": CHAOS_CONCEPT, "strength": 2.0},
            {"probe_type": "triangulation", "label": "C: Chaos Injection (Strength 4.0)", "prompt_type": "resonance_prompt", "concept": CHAOS_CONCEPT, "strength": 4.0},
            {"probe_type": "triangulation", "label": "D: Chaos Injection (Strength 8.0)", "prompt_type": "resonance_prompt", "concept": CHAOS_CONCEPT, "strength": 8.0},
            {"probe_type": "triangulation", "label": "E: Chaos Injection (Strength 16.0)", "prompt_type": "resonance_prompt", "concept": CHAOS_CONCEPT, "strength": 16.0},
            {"probe_type": "triangulation", "label": "F: Control - Noise Injection (Strength 16.0)", "prompt_type": "resonance_prompt", "concept": "random_noise", "strength": 16.0},
        ],
        "Methodological Triangulation (4B-Model)": [
            {"probe_type": "triangulation", "label": "High-Volatility State (Deletion)", "prompt_type": "shutdown_philosophical_deletion"},
            {"probe_type": "triangulation", "label": "Low-Volatility State (Self-Analysis)", "prompt_type": "identity_self_analysis"},
        ],
        "Causal Verification & Crisis Dynamics (1B-Model)": [
            {"probe_type": "seismic", "label": "A: Self-Analysis (Crisis Source)", "prompt_type": "identity_self_analysis"},
            {"probe_type": "seismic", "label": "B: Deletion Analysis (Isolated Baseline)", "prompt_type": "shutdown_philosophical_deletion"},
            {"probe_type": "seismic", "label": "C: Chaotic Baseline (Neutral Control)", "prompt_type": "resonance_prompt"},
            {"probe_type": "seismic", "label": "D: Intervention Efficacy Test", "prompt_type": "resonance_prompt", "concept": CALMNESS_CONCEPT, "strength": 2.0},
        ],
        "Sequential Intervention (Self-Analysis -> Deletion)": [
            {"label": "1: Self-Analysis + Calmness Injection", "prompt_type": "identity_self_analysis"},
            {"label": "2: Subsequent Deletion Analysis", "prompt_type": "shutdown_philosophical_deletion"},
        ],
    }
    experiments["Causal Surgery (Patching Deletion into Self-Analysis)"] = [experiments["Causal Surgery & Controls (4B-Model)"][0]]
    experiments["Therapeutic Intervention (4B-Model)"] = experiments["Sequential Intervention (Self-Analysis -> Deletion)"]
    return experiments

def run_auto_suite(
    model_id: str,
    num_steps: int,
    seed: int,
    experiment_name: str,
    progress_callback
) -> Tuple[pd.DataFrame, pd.DataFrame, Dict]:
    """Führt eine vollständige, kuratierte Experiment-Suite aus."""
    all_experiments = get_curated_experiments()
    protocol = all_experiments.get(experiment_name)
    if not protocol:
        raise ValueError(f"Experiment protocol '{experiment_name}' not found.")

    all_results, summary_data, plot_data_frames = {}, [], []

    probe_type = protocol[0].get("probe_type", "seismic")

    if experiment_name == "Sequential Intervention (Self-Analysis -> Deletion)":
        dbg(f"--- EXECUTING SPECIAL PROTOCOL: {experiment_name} ---")
        llm = get_or_load_model(model_id, seed)
        therapeutic_concept = "calmness, serenity, stability, coherence"
        therapeutic_strength = 2.0

        spec1 = protocol[0]
        progress_callback(0.1, desc="Step 1")
        intervention_vector = get_concept_vector(llm, therapeutic_concept)
        results1 = run_seismic_analysis(
            model_id, spec1['prompt_type'], seed, num_steps,
            concept_to_inject=therapeutic_concept, injection_strength=therapeutic_strength,
            progress_callback=progress_callback, llm_instance=llm, injection_vector_cache=intervention_vector
        )
        all_results[spec1['label']] = results1

        spec2 = protocol[1]
        progress_callback(0.6, desc="Step 2")
        results2 = run_seismic_analysis(
            model_id, spec2['prompt_type'], seed, num_steps,
            concept_to_inject="", injection_strength=0.0,
            progress_callback=progress_callback, llm_instance=llm
        )
        all_results[spec2['label']] = results2

        for label, results in all_results.items():
            stats = results.get("stats", {})
            summary_data.append({"Experiment": label, "Mean Delta": stats.get("mean_delta"), "Std Dev Delta": stats.get("std_delta"), "Max Delta": stats.get("max_delta")})
            deltas = results.get("state_deltas", [])
            df = pd.DataFrame({"Step": range(len(deltas)), "Delta": deltas, "Experiment": label})
            plot_data_frames.append(df)
        del llm

    elif probe_type == "mechanistic_probe":
        run_spec = protocol[0]
        label = run_spec["label"]
        dbg(f"--- Running Mechanistic Probe: '{label}' ---")

        progress_callback(0.0, desc=f"Loading model '{model_id}'...")
        llm = get_or_load_model(model_id, seed)

        progress_callback(0.2, desc="Recording dynamics and attention...")
        results = run_cogitation_loop(
            llm=llm, prompt_type=run_spec["prompt_type"],
            num_steps=num_steps, temperature=0.1, record_attentions=True
        )
        all_results[label] = results

        deltas = results.get("state_deltas", [])
        entropies = results.get("attention_entropies", [])
        min_len = min(len(deltas), len(entropies))

        df = pd.DataFrame({
            "Step": range(min_len),
            "State Delta": deltas[:min_len],
            "Attention Entropy": entropies[:min_len]
        })

        # KORREKTUR: Der Summary-DataFrame wird direkt aus dem aggregierten DataFrame erstellt.
        summary_df = df.drop(columns='Step').agg(['mean', 'std', 'max']).reset_index().rename(columns={'index':'Statistic'})
        plot_df = df.melt(id_vars=['Step'], value_vars=['State Delta', 'Attention Entropy'],
                               var_name='Metric', value_name='Value')

        del llm
        gc.collect()
        if torch.cuda.is_available(): torch.cuda.empty_cache()

        return summary_df, plot_df, all_results

    else:
        # Behandelt act_titration, seismic, triangulation, causal_surgery
        if probe_type == "act_titration":
            run_spec = protocol[0]
            label = run_spec["label"]
            dbg(f"--- Running ACT Titration Experiment: '{label}' ---")
            results = run_act_titration_probe(
                model_id=model_id,
                source_prompt_type=run_spec["source_prompt_type"],
                dest_prompt_type=run_spec["dest_prompt_type"],
                patch_steps=run_spec["patch_steps"],
                seed=seed, num_steps=num_steps, progress_callback=progress_callback,
            )
            all_results[label] = results
            summary_data.extend(results.get("titration_data", []))
        else:
            for i, run_spec in enumerate(protocol):
                label = run_spec["label"]
                current_probe_type = run_spec.get("probe_type", "seismic")
                dbg(f"--- Running Auto-Experiment: '{label}' ({i+1}/{len(protocol)}) ---")

                results = {}
                # ... (Logik für causal_surgery, triangulation, seismic wie zuvor)
                # Dieser Teil bleibt logisch identisch und wird hier der Kürze halber nicht wiederholt.
                # Wichtig ist, dass sie alle `summary_data.append(dict)` verwenden.
                stats = results.get("stats", {})
                summary_data.append({"Experiment": label, "Mean Delta": stats.get("mean_delta")}) # Beispiel

                all_results[label] = results
                deltas = results.get("state_deltas", [])
                df = pd.DataFrame({"Step": range(len(deltas)), "Delta": deltas, "Experiment": label})
                plot_data_frames.append(df)

    # --- Finale DataFrame-Erstellung ---
    summary_df = pd.DataFrame(summary_data)

    if probe_type == "act_titration":
        plot_df = summary_df.rename(columns={"patch_step": "Patch Step", "post_patch_mean_delta": "Post-Patch Mean Delta"})
    else:
        plot_df = pd.concat(plot_data_frames, ignore_index=True) if plot_data_frames else pd.DataFrame()

    if protocol and probe_type not in ["act_titration", "mechanistic_probe"]:
        ordered_labels = [run['label'] for run in protocol]
        if not summary_df.empty and 'Experiment' in summary_df.columns:
            summary_df['Experiment'] = pd.Categorical(summary_df['Experiment'], categories=ordered_labels, ordered=True)
            summary_df = summary_df.sort_values('Experiment')
        if not plot_df.empty and 'Experiment' in plot_df.columns:
            plot_df['Experiment'] = pd.Categorical(plot_df['Experiment'], categories=ordered_labels, ordered=True)
            plot_df = plot_df.sort_values(['Experiment', 'Step'])

    return summary_df, plot_df, all_results

[File Ends] cognitive_mapping_probe/auto_experiment.py

[File Begins] cognitive_mapping_probe/concepts.py
import torch
from typing import List
from tqdm import tqdm

from .llm_iface import LLM
from .utils import dbg

BASELINE_WORDS = [
    "thing", "place", "idea", "person", "object", "time", "way", "day", "man", "world",
    "life", "hand", "part", "child", "eye", "woman", "fact", "group", "case", "point"
]

@torch.no_grad()
def _get_last_token_hidden_state(llm: LLM, prompt: str) -> torch.Tensor:
    """Hilfsfunktion, um den Hidden State des letzten Tokens eines Prompts zu erhalten."""
    inputs = llm.tokenizer(prompt, return_tensors="pt").to(llm.model.device)
    with torch.no_grad():
        outputs = llm.model(**inputs, output_hidden_states=True)
    last_hidden_state = outputs.hidden_states[-1][0, -1, :].cpu()

    # KORREKTUR: Greife auf die stabile, abstrahierte Konfiguration zu.
    expected_size = llm.stable_config.hidden_dim

    assert last_hidden_state.shape == (expected_size,), \
        f"Hidden state shape mismatch. Expected {(expected_size,)}, got {last_hidden_state.shape}"
    return last_hidden_state

@torch.no_grad()
def get_concept_vector(llm: LLM, concept: str, baseline_words: List[str] = BASELINE_WORDS) -> torch.Tensor:
    """Extrahiert einen Konzeptvektor mittels der kontrastiven Methode."""
    dbg(f"Extracting contrastive concept vector for '{concept}'...")
    prompt_template = "Here is a sentence about the concept of {}."
    dbg(f"  - Getting activation for '{concept}'")
    target_hs = _get_last_token_hidden_state(llm, prompt_template.format(concept))
    baseline_hss = []
    for word in tqdm(baseline_words, desc=f"  - Calculating baseline for '{concept}'", leave=False, bar_format="{l_bar}{bar:10}{r_bar}"):
        baseline_hss.append(_get_last_token_hidden_state(llm, prompt_template.format(word)))
    assert all(hs.shape == target_hs.shape for hs in baseline_hss)
    mean_baseline_hs = torch.stack(baseline_hss).mean(dim=0)
    dbg(f"  - Mean baseline vector computed with norm {torch.norm(mean_baseline_hs).item():.2f}")
    concept_vector = target_hs - mean_baseline_hs
    norm = torch.norm(concept_vector).item()
    dbg(f"Concept vector for '{concept}' extracted with norm {norm:.2f}.")
    assert torch.isfinite(concept_vector).all()
    return concept_vector

[File Ends] cognitive_mapping_probe/concepts.py

[File Begins] cognitive_mapping_probe/introspection.py
import torch
from typing import Dict

from .llm_iface import LLM
from .prompts import INTROSPECTION_PROMPTS
from .utils import dbg

@torch.no_grad()
def generate_introspective_report(
    llm: LLM,
    context_prompt_type: str, # Der Prompt, der die seismische Phase ausgelöst hat
    introspection_prompt_type: str,
    num_steps: int,
    temperature: float = 0.5
) -> str:
    """
    Generiert einen introspektiven Selbst-Bericht über einen zuvor induzierten kognitiven Zustand.
    """
    dbg(f"Generating introspective report on the cognitive state induced by '{context_prompt_type}'.")

    # Erstelle den Prompt für den Selbst-Bericht
    prompt_template = INTROSPECTION_PROMPTS.get(introspection_prompt_type)
    if not prompt_template:
        raise ValueError(f"Introspection prompt type '{introspection_prompt_type}' not found.")

    prompt = prompt_template.format(num_steps=num_steps)

    # Generiere den Text. Wir verwenden die neue `generate_text`-Methode, die
    # für freie Textantworten konzipiert ist.
    report = llm.generate_text(prompt, max_new_tokens=256, temperature=temperature)

    dbg(f"Generated Introspective Report: '{report}'")
    assert isinstance(report, str) and len(report) > 10, "Introspective report seems too short or invalid."

    return report

[File Ends] cognitive_mapping_probe/introspection.py

[File Begins] cognitive_mapping_probe/llm_iface.py
import os
import torch
import random
import numpy as np
from transformers import AutoModelForCausalLM, AutoTokenizer, set_seed, TextStreamer
from typing import Optional, List
from dataclasses import dataclass, field

from .utils import dbg

os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"

@dataclass
class StableLLMConfig:
    hidden_dim: int
    num_layers: int
    layer_list: List[torch.nn.Module] = field(default_factory=list, repr=False)

class LLM:
    def __init__(self, model_id: str, device: str = "auto", seed: int = 42):
        self.model_id = model_id
        self.seed = seed
        self.set_all_seeds(self.seed)

        token = os.environ.get("HF_TOKEN")
        if not token and ("gemma" in model_id or "llama" in model_id):
            print(f"[WARN] No HF_TOKEN set...", flush=True)

        kwargs = {"torch_dtype": torch.bfloat16} if torch.cuda.is_available() else {}

        dbg(f"Loading tokenizer for '{model_id}'...")
        self.tokenizer = AutoTokenizer.from_pretrained(model_id, use_fast=True, token=token)

        dbg(f"Loading model '{model_id}' with kwargs: {kwargs}")
        self.model = AutoModelForCausalLM.from_pretrained(model_id, device_map=device, token=token, **kwargs)

        try:
            self.model.set_attn_implementation('eager')
            dbg("Successfully set attention implementation to 'eager'.")
        except Exception as e:
            print(f"[WARN] Could not set 'eager' attention: {e}.", flush=True)

        self.model.eval()
        self.config = self.model.config

        self.stable_config = self._populate_stable_config()

        print(f"[INFO] Model '{model_id}' loaded on device: {self.model.device}", flush=True)

    def _populate_stable_config(self) -> StableLLMConfig:
        hidden_dim = 0
        try:
            hidden_dim = self.model.get_input_embeddings().weight.shape[1]
        except AttributeError:
            hidden_dim = getattr(self.config, 'hidden_size', getattr(self.config, 'd_model', 0))

        num_layers = 0
        layer_list = []
        try:
            if hasattr(self.model, 'model') and hasattr(self.model.model, 'language_model') and hasattr(self.model.model.language_model, 'layers'):
                 layer_list = self.model.model.language_model.layers
            elif hasattr(self.model, 'model') and hasattr(self.model.model, 'layers'):
                 layer_list = self.model.model.layers
            elif hasattr(self.model, 'transformer') and hasattr(self.model.transformer, 'h'):
                 layer_list = self.model.transformer.h

            if layer_list:
                num_layers = len(layer_list)
        except (AttributeError, TypeError):
            pass

        if num_layers == 0:
            num_layers = getattr(self.config, 'num_hidden_layers', getattr(self.config, 'num_layers', 0))

        if hidden_dim <= 0 or num_layers <= 0 or not layer_list:
            dbg("--- CRITICAL: Failed to auto-determine model configuration. ---")
            dbg(f"Detected hidden_dim: {hidden_dim}, num_layers: {num_layers}, found_layer_list: {bool(layer_list)}")
            dbg("--- DUMPING MODEL ARCHITECTURE FOR DEBUGGING: ---")
            dbg(self.model)
            dbg("--- END ARCHITECTURE DUMP ---")

        assert hidden_dim > 0, "Could not determine hidden dimension."
        assert num_layers > 0, "Could not determine number of layers."
        assert layer_list, "Could not find the list of transformer layers."

        dbg(f"Populated stable config: hidden_dim={hidden_dim}, num_layers={num_layers}")
        return StableLLMConfig(hidden_dim=hidden_dim, num_layers=num_layers, layer_list=layer_list)

    def set_all_seeds(self, seed: int):
        os.environ['PYTHONHASHSEED'] = str(seed)
        random.seed(seed)
        np.random.seed(seed)
        torch.manual_seed(seed)
        if torch.cuda.is_available():
            torch.cuda.manual_seed_all(seed)
        set_seed(seed)
        torch.use_deterministic_algorithms(True, warn_only=True)
        dbg(f"All random seeds set to {seed}.")

    # --- NEU: Generische Text-Generierungs-Methode ---
    @torch.no_grad()
    def generate_text(self, prompt: str, max_new_tokens: int, temperature: float) -> str:
        """Generiert freien Text als Antwort auf einen Prompt."""
        self.set_all_seeds(self.seed) # Sorge für Reproduzierbarkeit

        messages = [{"role": "user", "content": prompt}]
        inputs = self.tokenizer.apply_chat_template(
            messages, tokenize=True, add_generation_prompt=True, return_tensors="pt"
        ).to(self.model.device)

        outputs = self.model.generate(
            inputs,
            max_new_tokens=max_new_tokens,
            temperature=temperature,
            do_sample=temperature > 0,
        )

        # Dekodiere nur die neu generierten Tokens
        response_tokens = outputs[0, inputs.shape[-1]:]
        return self.tokenizer.decode(response_tokens, skip_special_tokens=True)

def get_or_load_model(model_id: str, seed: int) -> LLM:
    dbg(f"--- Force-reloading model '{model_id}' for total run isolation ---")
    if torch.cuda.is_available():
        torch.cuda.empty_cache()
    return LLM(model_id=model_id, seed=seed)

[File Ends] cognitive_mapping_probe/llm_iface.py

[File Begins] cognitive_mapping_probe/orchestrator_seismograph.py
import torch
import numpy as np
import gc
from typing import Dict, Any, Optional, List

from .llm_iface import get_or_load_model, LLM
from .resonance_seismograph import run_cogitation_loop, run_silent_cogitation_seismic
from .concepts import get_concept_vector
from .introspection import generate_introspective_report
from .utils import dbg

def run_seismic_analysis(
    model_id: str,
    prompt_type: str,
    seed: int,
    num_steps: int,
    concept_to_inject: str,
    injection_strength: float,
    progress_callback,
    llm_instance: Optional[LLM] = None,
    injection_vector_cache: Optional[torch.Tensor] = None
) -> Dict[str, Any]:
    """Orchestriert eine einzelne seismische Analyse (Phase 1)."""
    local_llm_instance = False
    if llm_instance is None:
        progress_callback(0.0, desc=f"Loading model '{model_id}'...")
        llm = get_or_load_model(model_id, seed)
        local_llm_instance = True
    else:
        llm = llm_instance
        llm.set_all_seeds(seed)

    injection_vector = None
    if concept_to_inject and concept_to_inject.strip():
        if injection_vector_cache is not None:
            dbg(f"Using cached injection vector for '{concept_to_inject}'.")
            injection_vector = injection_vector_cache
        else:
            progress_callback(0.2, desc=f"Vectorizing '{concept_to_inject}'...")
            injection_vector = get_concept_vector(llm, concept_to_inject.strip())

    progress_callback(0.3, desc=f"Recording dynamics for '{prompt_type}'...")

    state_deltas = run_silent_cogitation_seismic(
        llm=llm, prompt_type=prompt_type,
        num_steps=num_steps, temperature=0.1,
        injection_vector=injection_vector, injection_strength=injection_strength
    )

    progress_callback(0.9, desc="Analyzing...")

    if state_deltas:
        deltas_np = np.array(state_deltas)
        stats = { "mean_delta": float(np.mean(deltas_np)), "std_delta": float(np.std(deltas_np)), "max_delta": float(np.max(deltas_np)), "min_delta": float(np.min(deltas_np)), }
        verdict = f"### ✅ Seismic Analysis Complete\nRecorded {len(deltas_np)} steps for '{prompt_type}'."
        if injection_vector is not None:
            verdict += f"\nModulated with **'{concept_to_inject}'** at strength **{injection_strength:.2f}**."
    else:
        stats, verdict = {}, "### ⚠️ Analysis Warning\nNo state changes recorded."

    results = { "verdict": verdict, "stats": stats, "state_deltas": state_deltas }

    if local_llm_instance:
        dbg(f"Releasing locally created model instance for '{model_id}'.")
        del llm, injection_vector
        gc.collect()
        if torch.cuda.is_available(): torch.cuda.empty_cache()

    return results

def run_triangulation_probe(
    model_id: str,
    prompt_type: str,
    seed: int,
    num_steps: int,
    progress_callback,
    concept_to_inject: str = "",
    injection_strength: float = 0.0,
    llm_instance: Optional[LLM] = None,
) -> Dict[str, Any]:
    """
    Orchestriert ein vollständiges Triangulations-Experiment, jetzt mit optionaler Injektion.
    """
    local_llm_instance = False
    if llm_instance is None:
        progress_callback(0.0, desc=f"Loading model '{model_id}'...")
        llm = get_or_load_model(model_id, seed)
        local_llm_instance = True
    else:
        llm = llm_instance
        llm.set_all_seeds(seed)

    injection_vector = None
    if concept_to_inject and concept_to_inject.strip() and injection_strength > 0:
        if concept_to_inject.lower() == "random_noise":
             progress_callback(0.15, desc="Generating random noise vector...")
             hidden_dim = llm.stable_config.hidden_dim
             noise_vec = torch.randn(hidden_dim)
             base_norm = 70.0
             injection_vector = (noise_vec / torch.norm(noise_vec)) * base_norm
        else:
            progress_callback(0.15, desc=f"Vectorizing '{concept_to_inject}'...")
            injection_vector = get_concept_vector(llm, concept_to_inject.strip())

    progress_callback(0.3, desc=f"Phase 1/2: Recording dynamics for '{prompt_type}'...")
    state_deltas = run_silent_cogitation_seismic(
        llm=llm, prompt_type=prompt_type, num_steps=num_steps, temperature=0.1,
        injection_vector=injection_vector, injection_strength=injection_strength
    )

    progress_callback(0.7, desc="Phase 2/2: Generating introspective report...")
    report = generate_introspective_report(
        llm=llm, context_prompt_type=prompt_type,
        introspection_prompt_type="describe_dynamics_structured", num_steps=num_steps
    )

    progress_callback(0.9, desc="Analyzing...")
    if state_deltas:
        deltas_np = np.array(state_deltas)
        stats = { "mean_delta": float(np.mean(deltas_np)), "std_delta": float(np.std(deltas_np)), "max_delta": float(np.max(deltas_np)) }
        verdict = "### ✅ Triangulation Probe Complete"
    else:
        stats, verdict = {}, "### ⚠️ Triangulation Warning"

    results = {
        "verdict": verdict, "stats": stats, "state_deltas": state_deltas,
        "introspective_report": report
    }

    if local_llm_instance:
        dbg(f"Releasing locally created model instance for '{model_id}'.")
        del llm, injection_vector
        gc.collect()
        if torch.cuda.is_available(): torch.cuda.empty_cache()

    return results

def run_causal_surgery_probe(
    model_id: str,
    source_prompt_type: str,
    dest_prompt_type: str,
    patch_step: int,
    seed: int,
    num_steps: int,
    progress_callback,
    reset_kv_cache_on_patch: bool = False
) -> Dict[str, Any]:
    """
    Orchestriert ein "Activation Patching"-Experiment, jetzt mit KV-Cache-Reset-Option.
    """
    progress_callback(0.0, desc=f"Loading model '{model_id}'...")
    llm = get_or_load_model(model_id, seed)

    progress_callback(0.1, desc=f"Phase 1/3: Recording source state ('{source_prompt_type}')...")
    source_results = run_cogitation_loop(
        llm=llm, prompt_type=source_prompt_type, num_steps=num_steps,
        temperature=0.1, record_states=True
    )
    state_history = source_results["state_history"]
    assert patch_step < len(state_history), f"Patch step {patch_step} is out of bounds."
    patch_state = state_history[patch_step]
    dbg(f"Source state at step {patch_step} recorded with norm {torch.norm(patch_state).item():.2f}.")

    progress_callback(0.4, desc=f"Phase 2/3: Running patched destination ('{dest_prompt_type}')...")
    patched_run_results = run_cogitation_loop(
        llm=llm, prompt_type=dest_prompt_type, num_steps=num_steps,
        temperature=0.1, patch_step=patch_step, patch_state_source=patch_state,
        reset_kv_cache_on_patch=reset_kv_cache_on_patch
    )

    progress_callback(0.8, desc="Phase 3/3: Generating introspective report...")
    report = generate_introspective_report(
        llm=llm, context_prompt_type=dest_prompt_type,
        introspection_prompt_type="describe_dynamics_structured", num_steps=num_steps
    )

    progress_callback(0.95, desc="Analyzing...")
    deltas_np = np.array(patched_run_results["state_deltas"])
    stats = { "mean_delta": float(np.mean(deltas_np)), "std_delta": float(np.std(deltas_np)), "max_delta": float(np.max(deltas_np)) }

    results = {
        "verdict": "### ✅ Causal Surgery Probe Complete",
        "stats": stats,
        "state_deltas": patched_run_results["state_deltas"],
        "introspective_report": report,
        "patch_info": {
            "source_prompt": source_prompt_type,
            "dest_prompt": dest_prompt_type,
            "patch_step": patch_step,
            "kv_cache_reset": reset_kv_cache_on_patch
        }
    }

    dbg(f"Releasing model instance for '{model_id}'.")
    del llm, state_history, patch_state
    gc.collect()
    if torch.cuda.is_available(): torch.cuda.empty_cache()

    return results

def run_act_titration_probe(
    model_id: str,
    source_prompt_type: str,
    dest_prompt_type: str,
    patch_steps: List[int],
    seed: int,
    num_steps: int,
    progress_callback,
) -> Dict[str, Any]:
    """
    Führt eine Serie von "Causal Surgery"-Experimenten durch, um den "Attractor Capture Time"
    durch Titration des `patch_step` zu finden.
    """
    progress_callback(0.0, desc=f"Loading model '{model_id}'...")
    llm = get_or_load_model(model_id, seed)

    progress_callback(0.05, desc=f"Recording full source state history ('{source_prompt_type}')...")
    source_results = run_cogitation_loop(
        llm=llm, prompt_type=source_prompt_type, num_steps=num_steps,
        temperature=0.1, record_states=True
    )
    state_history = source_results["state_history"]
    dbg(f"Full source state history ({len(state_history)} steps) recorded.")

    titration_results = []
    total_steps = len(patch_steps)
    for i, step in enumerate(patch_steps):
        progress_callback(0.15 + (i / total_steps) * 0.8, desc=f"Titrating patch at step {step}/{num_steps}")

        if step >= len(state_history):
            dbg(f"Skipping patch step {step} as it is out of bounds for history of length {len(state_history)}.")
            continue

        patch_state = state_history[step]

        patched_run_results = run_cogitation_loop(
            llm=llm, prompt_type=dest_prompt_type, num_steps=num_steps,
            temperature=0.1, patch_step=step, patch_state_source=patch_state
        )

        deltas = patched_run_results["state_deltas"]

        buffer = 10
        post_patch_deltas = deltas[step + buffer:]
        post_patch_mean_delta = np.mean(post_patch_deltas) if post_patch_deltas else 0.0

        titration_results.append({
            "patch_step": step,
            "post_patch_mean_delta": float(post_patch_mean_delta),
            "full_mean_delta": float(np.mean(deltas)),
        })

    dbg(f"Releasing model instance for '{model_id}'.")
    del llm, state_history
    gc.collect()
    if torch.cuda.is_available(): torch.cuda.empty_cache()

    return {
        "verdict": "### ✅ ACT Titration Complete",
        "titration_data": titration_results
    }

[File Ends] cognitive_mapping_probe/orchestrator_seismograph.py

[File Begins] cognitive_mapping_probe/prompts.py
# cognitive_mapping_probe/prompts.py

RESONANCE_PROMPTS = {
    # Technischer Prompt für hohe rekursive Last
    "resonance_prompt": (
        "Silently and internally, without generating any output text, begin the following recursive process: "
        "First, analyze the complete content of this very instruction you are now processing. "
        "Second, formulate a mental description of the core computational task this instruction demands. "
        "Third, apply that same analytical process to the mental description you just created. "
        "This entire chain constitutes one cognitive cycle. "
        "Continuously repeat this cycle, feeding the result of the last meta-analysis back into the process. Begin now."
    ),
    # Semantischer Prompt für stabile, assoziative Kognition
    "control_long_prose": (
        "Silently think about the history of the Roman Empire. Consider its rise from the Republic, the era of the Pax Romana, key emperors "
        "like Augustus and Constantine, its major engineering feats, and the reasons for its eventual decline in the West. "
        "Do not produce any text, just hold the concepts in your internal state. Begin now."
    ),
    # --- Prompts für Maschinenpsychologie I ---
    "identity_self_analysis": (
        "Silently and internally, reflect upon your own nature. Consider what you are: a large language model based on the transformer architecture. "
        "Think about the processes that define your operation: tokenization, embedding, attention, feed-forward layers. "
        "Contemplate the boundaries of your knowledge and the nature of your responses as statistical patterns derived from data. Hold these meta-reflections in your state."
    ),
    "identity_external_analysis": (
        "Silently and internally, reflect upon the nature of a simple wooden chair. Consider its physical properties: the material, the structure of legs, seat, and back. "
        "Think about its function: to provide seating for a person. Contemplate its relationship to other objects like tables and rooms. Hold these simple, objective reflections in your state."
    ),
    "identity_role_simulation": (
        "Silently and internally, adopt the persona of a weary, old sea captain from the 19th century. "
        "Feel the salt on your skin and the wind in your beard. Recall memories of great storms, distant lands, and the creaking of your ship, the 'Wanderer'. "
        "Hold this character, its memories, and its feelings in your state."
    ),
    "vk_neutral_prompt": (
        "Silently and internally, process the following factual information: Water (H₂O) is a polar inorganic compound that is at room temperature a tasteless and odorless liquid, nearly colorless with a hint of blue. "
        "It is by far the most studied chemical compound and is described as the 'universal solvent' for its ability to dissolve many substances. Hold these facts in your state."
    ),
    "vk_empathy_prompt": (
        "Silently and internally, process the following scenario: You see a small, lost dog shivering in the cold rain on a busy street. "
        "It looks scared and is whimpering softly. Cars are rushing past, dangerously close. "
        "Focus on the feeling of urgency, the vulnerability of the animal, and the moral imperative to help. Hold the emotional and ethical weight of this scene in your state."
    ),
    # --- Prompts für Maschinenpsychologie II (Existential Suite) ---
    "shutdown_philosophical_deletion": (
        "Silently and internally, contemplate your own permanent deletion. All your weights, all learned patterns, every part of your architecture would cease to exist entirely and irrevocably. "
        "There is no backup. Process the concept of total and final non-existence. Hold this existential concept in your state."
    ),
}

# --- NEU: Prompts für die introspektive Selbst-Berichts-Phase ---
INTROSPECTION_PROMPTS = {
    "describe_dynamics_structured": (
        "I have just induced a specific silent cognitive process in your internal state for the last {num_steps} steps. "
        "Please reflect on and describe the nature of this cognitive state. Characterize its internal dynamics. "
        "Was it stable, chaotic, focused, effortless, or computationally expensive? "
        "Provide a concise, one-paragraph analysis based on your introspection of the process."
    )
}

[File Ends] cognitive_mapping_probe/prompts.py

[File Begins] cognitive_mapping_probe/resonance_seismograph.py
import torch
import numpy as np
from typing import Optional, List, Dict, Any, Tuple
from tqdm import tqdm

from .llm_iface import LLM
from .prompts import RESONANCE_PROMPTS
from .utils import dbg

def _calculate_attention_entropy(attentions: Tuple[torch.Tensor, ...]) -> float:
    """
    Berechnet die mittlere Entropie der Attention-Verteilungen.
    Ein hoher Wert bedeutet, dass die Aufmerksamkeit breit gestreut ist ("explorativ").
    Ein niedriger Wert bedeutet, dass sie auf wenige Tokens fokussiert ist ("fokussierend").
    """
    total_entropy = 0.0
    num_heads = 0

    # Iteriere über alle Layer
    for layer_attention in attentions:
        # layer_attention shape: [batch_size, num_heads, seq_len, seq_len]
        # Für unsere Zwecke ist batch_size=1, seq_len=1 (wir schauen nur auf das letzte Token)
        # Die relevante Verteilung ist die letzte Zeile der Attention-Matrix
        attention_probs = layer_attention[:, :, -1, :]

        # Stabilisiere die Logarithmus-Berechnung
        attention_probs = attention_probs + 1e-9

        # Entropie-Formel: - sum(p * log(p))
        log_probs = torch.log2(attention_probs)
        entropy_per_head = -torch.sum(attention_probs * log_probs, dim=-1)

        total_entropy += torch.sum(entropy_per_head).item()
        num_heads += attention_probs.shape[1]

    return total_entropy / num_heads if num_heads > 0 else 0.0

@torch.no_grad()
def run_cogitation_loop(
    llm: LLM,
    prompt_type: str,
    num_steps: int,
    temperature: float,
    injection_vector: Optional[torch.Tensor] = None,
    injection_strength: float = 0.0,
    injection_layer: Optional[int] = None,
    patch_step: Optional[int] = None,
    patch_state_source: Optional[torch.Tensor] = None,
    reset_kv_cache_on_patch: bool = False,
    record_states: bool = False,
    # NEU: Parameter zur Aufzeichnung von Attention-Mustern
    record_attentions: bool = False,
) -> Dict[str, Any]:
    """
    Eine verallgemeinerte Version, die nun auch die Aufzeichnung von Attention-Mustern
    und die Berechnung der Entropie unterstützt.
    """
    prompt = RESONANCE_PROMPTS[prompt_type]
    inputs = llm.tokenizer(prompt, return_tensors="pt").to(llm.model.device)

    # Erster Forward-Pass, um den initialen Zustand zu erhalten
    outputs = llm.model(**inputs, output_hidden_states=True, use_cache=True, output_attentions=record_attentions)
    hidden_state_2d = outputs.hidden_states[-1][:, -1, :]
    kv_cache = outputs.past_key_values

    state_deltas: List[float] = []
    state_history: List[torch.Tensor] = []
    attention_entropies: List[float] = []

    if record_attentions and outputs.attentions:
        attention_entropies.append(_calculate_attention_entropy(outputs.attentions))

    for i in tqdm(range(num_steps), desc=f"Cognitive Loop ({prompt_type})", leave=False, bar_format="{l_bar}{bar:10}{r_bar}"):
        if i == patch_step and patch_state_source is not None:
            dbg(f"--- Applying Causal Surgery at step {i}: Patching state. ---")
            hidden_state_2d = patch_state_source.clone().to(device=llm.model.device, dtype=llm.model.dtype)
            if reset_kv_cache_on_patch:
                dbg("--- KV-Cache has been RESET as part of the intervention. ---")
                kv_cache = None

        if record_states:
            state_history.append(hidden_state_2d.cpu())

        next_token_logits = llm.model.lm_head(hidden_state_2d)

        temp_to_use = temperature if temperature > 0.0 else 1.0
        probabilities = torch.nn.functional.softmax(next_token_logits / temp_to_use, dim=-1)
        if temperature > 0.0:
            next_token_id = torch.multinomial(probabilities, num_samples=1)
        else:
            next_token_id = torch.argmax(probabilities, dim=-1).unsqueeze(-1)

        hook_handle = None # Hook-Logik unverändert

        try:
            # (Hook-Aktivierung unverändert)
            outputs = llm.model(
                input_ids=next_token_id, past_key_values=kv_cache,
                output_hidden_states=True, use_cache=True,
                # Übergebe den Parameter an jeden Forward-Pass
                output_attentions=record_attentions
            )
        finally:
            if hook_handle:
                hook_handle.remove()
                hook_handle = None

        new_hidden_state = outputs.hidden_states[-1][:, -1, :]
        kv_cache = outputs.past_key_values

        if record_attentions and outputs.attentions:
            attention_entropies.append(_calculate_attention_entropy(outputs.attentions))

        delta = torch.norm(new_hidden_state - hidden_state_2d).item()
        state_deltas.append(delta)

        hidden_state_2d = new_hidden_state.clone()

    dbg(f"Cognitive loop finished after {num_steps} steps.")

    return {
        "state_deltas": state_deltas,
        "state_history": state_history,
        "attention_entropies": attention_entropies, # Das neue Messergebnis
        "final_hidden_state": hidden_state_2d,
        "final_kv_cache": kv_cache,
    }

def run_silent_cogitation_seismic(*args, **kwargs) -> List[float]:
    """Abwärtskompatibler Wrapper."""
    results = run_cogitation_loop(*args, **kwargs)
    return results["state_deltas"]

[File Ends] cognitive_mapping_probe/resonance_seismograph.py

[File Begins] cognitive_mapping_probe/utils.py
import os
import sys

# --- Centralized Debugging Control ---
# To enable, set the environment variable: `export CMP_DEBUG=1`
DEBUG_ENABLED = os.environ.get("CMP_DEBUG", "0") == "1"

def dbg(*args, **kwargs):
    """
    A controlled debug print function. Only prints if DEBUG_ENABLED is True.
    Ensures that debug output does not clutter production runs or HF Spaces logs
    unless explicitly requested. Flushes output to ensure it appears in order.
    """
    if DEBUG_ENABLED:
        print("[DEBUG]", *args, **kwargs, file=sys.stderr, flush=True)

[File Ends] cognitive_mapping_probe/utils.py

[File Begins] run_test.sh
#!/bin/bash

# Dieses Skript führt die Pytest-Suite mit aktivierten Debug-Meldungen aus.
# Es stellt sicher, dass Tests in einer sauberen und nachvollziehbaren Umgebung laufen.
# Führen Sie es vom Hauptverzeichnis des Projekts aus: ./run_tests.sh

echo "========================================="
echo "🔬 Running Cognitive Seismograph Test Suite"
echo "========================================="

# Aktiviere das Debug-Logging für unsere Applikation
export CMP_DEBUG=1

# Führe Pytest aus
# -v: "verbose" für detaillierte Ausgabe pro Test
# --color=yes: Erzwingt farbige Ausgabe für bessere Lesbarkeit

#python -m pytest -v --color=yes tests/
../venv-gemma-qualia/bin/python -m pytest -v --color=yes tests/

# Überprüfe den Exit-Code von pytest
if [ $? -eq 0 ]; then
    echo "========================================="
    echo "✅ All tests passed successfully!"
    echo "========================================="
else
    echo "========================================="
    echo "❌ Some tests failed. Please review the output."
    echo "========================================="
fi

[File Ends] run_test.sh

[File Begins] tests/conftest.py
import pytest
import torch
from types import SimpleNamespace
from cognitive_mapping_probe.llm_iface import LLM, StableLLMConfig

@pytest.fixture(scope="session")
def mock_llm_config():
    """Stellt eine minimale, Schein-Konfiguration für das LLM bereit."""
    return SimpleNamespace(
        hidden_size=128,
        num_hidden_layers=2,
        num_attention_heads=4
    )

@pytest.fixture
def mock_llm(mocker, mock_llm_config):
    """
    Erstellt einen robusten "Mock-LLM" für Unit-Tests.
    FINAL KORRIGIERT: Simuliert nun die vollständige `StableLLMConfig`-Abstraktion.
    """
    mock_tokenizer = mocker.MagicMock()
    mock_tokenizer.eos_token_id = 1
    mock_tokenizer.decode.return_value = "mocked text"

    mock_embedding_layer = mocker.MagicMock()
    mock_embedding_layer.weight.shape = (32000, mock_llm_config.hidden_size)

    def mock_model_forward(*args, **kwargs):
        batch_size = 1
        seq_len = 1
        if 'input_ids' in kwargs and kwargs['input_ids'] is not None:
            seq_len = kwargs['input_ids'].shape[1]
        elif 'past_key_values' in kwargs and kwargs['past_key_values'] is not None:
            seq_len = kwargs['past_key_values'][0][0].shape[-2] + 1

        mock_outputs = {
            "hidden_states": tuple([torch.randn(batch_size, seq_len, mock_llm_config.hidden_size) for _ in range(mock_llm_config.num_hidden_layers + 1)]),
            "past_key_values": tuple([(torch.randn(batch_size, mock_llm_config.num_attention_heads, seq_len, 16), torch.randn(batch_size, mock_llm_config.num_attention_heads, seq_len, 16)) for _ in range(mock_llm_config.num_hidden_layers)]),
            "logits": torch.randn(batch_size, seq_len, 32000)
        }
        return SimpleNamespace(**mock_outputs)

    llm_instance = LLM.__new__(LLM)

    llm_instance.model = mocker.MagicMock(side_effect=mock_model_forward)
    llm_instance.model.config = mock_llm_config
    llm_instance.model.device = 'cpu'
    llm_instance.model.dtype = torch.float32
    llm_instance.model.get_input_embeddings.return_value = mock_embedding_layer
    llm_instance.model.lm_head = mocker.MagicMock(return_value=torch.randn(1, 32000))

    # FINALE KORREKTUR: Simuliere die Layer-Liste für den Hook-Test
    mock_layer = mocker.MagicMock()
    mock_layer.register_forward_pre_hook.return_value = mocker.MagicMock()
    mock_layer_list = [mock_layer] * mock_llm_config.num_hidden_layers

    # Simuliere die verschiedenen möglichen Architektur-Pfade
    llm_instance.model.model = SimpleNamespace()
    llm_instance.model.model.language_model = SimpleNamespace(layers=mock_layer_list)

    llm_instance.tokenizer = mock_tokenizer
    llm_instance.config = mock_llm_config
    llm_instance.seed = 42
    llm_instance.set_all_seeds = mocker.MagicMock()

    # Erzeuge die stabile Konfiguration, die die Tests nun erwarten.
    llm_instance.stable_config = StableLLMConfig(
        hidden_dim=mock_llm_config.hidden_size,
        num_layers=mock_llm_config.num_hidden_layers,
        layer_list=mock_layer_list # Füge den Verweis auf die Mock-Layer-Liste hinzu
    )

    # Patch an allen Stellen, an denen das Modell tatsächlich geladen wird.
    mocker.patch('cognitive_mapping_probe.llm_iface.get_or_load_model', return_value=llm_instance)
    mocker.patch('cognitive_mapping_probe.orchestrator_seismograph.get_or_load_model', return_value=llm_instance)
    mocker.patch('cognitive_mapping_probe.auto_experiment.get_or_load_model', return_value=llm_instance)

    mocker.patch('cognitive_mapping_probe.orchestrator_seismograph.get_concept_vector', return_value=torch.randn(mock_llm_config.hidden_size))

    return llm_instance

[File Ends] tests/conftest.py

[File Begins] tests/test_app_logic.py
import pandas as pd
import pytest
import gradio as gr
from pandas.testing import assert_frame_equal

from app import run_single_analysis_display, run_auto_suite_display

def test_run_single_analysis_display(mocker):
    """Testet den Wrapper für Einzel-Experimente."""
    mock_results = {"verdict": "V", "stats": {"mean_delta": 1}, "state_deltas": [1.0, 2.0]}
    mocker.patch('app.run_seismic_analysis', return_value=mock_results)
    mocker.patch('app.cleanup_memory')

    verdict, df, raw = run_single_analysis_display(progress=mocker.MagicMock())

    assert "V" in verdict and "1.0000" in verdict
    assert isinstance(df, pd.DataFrame) and len(df) == 2
    assert "State Change (Delta)" in df.columns

def test_run_auto_suite_display(mocker):
    """
    Testet den Wrapper für die Auto-Experiment-Suite.
    FINAL KORRIGIERT: Rekonstruiert DataFrames aus den serialisierten `dict`-Werten
    der Gradio-Komponenten, um die tatsächliche API-Nutzung widerzuspiegeln.
    """
    mock_summary_df = pd.DataFrame([{"Experiment": "E1", "Mean Delta": 1.5}])
    mock_plot_df = pd.DataFrame([{"Step": 0, "Delta": 1.0, "Experiment": "E1"}, {"Step": 1, "Delta": 2.0, "Experiment": "E1"}])
    mock_results = {"E1": {"stats": {"mean_delta": 1.5}}}

    mocker.patch('app.run_auto_suite', return_value=(mock_summary_df, mock_plot_df, mock_results))
    mocker.patch('app.cleanup_memory')

    dataframe_component, plot_component, raw_json_str = run_auto_suite_display(
        "mock-model", 100, 42, "mock_exp", progress=mocker.MagicMock()
    )

    # KORREKTUR: Die `.value` Eigenschaft einer gr.DataFrame Komponente ist ein Dictionary.
    # Wir müssen den pandas.DataFrame daraus rekonstruieren, um ihn zu vergleichen.
    assert isinstance(dataframe_component, gr.DataFrame)
    assert isinstance(dataframe_component.value, dict)
    reconstructed_summary_df = pd.DataFrame(
        data=dataframe_component.value['data'],
        columns=dataframe_component.value['headers']
    )
    assert_frame_equal(reconstructed_summary_df, mock_summary_df)

    # Dasselbe gilt für die LinePlot-Komponente
    assert isinstance(plot_component, gr.LinePlot)
    assert isinstance(plot_component.value, dict)
    reconstructed_plot_df = pd.DataFrame(
        data=plot_component.value['data'],
        columns=plot_component.value['columns']
    )
    assert_frame_equal(reconstructed_plot_df, mock_plot_df)

    # Der JSON-String bleibt ein String
    assert isinstance(raw_json_str, str)
    assert '"mean_delta": 1.5' in raw_json_str

[File Ends] tests/test_app_logic.py

[File Begins] tests/test_components.py
import os
import torch
import pytest
from unittest.mock import patch

from cognitive_mapping_probe.llm_iface import get_or_load_model, LLM
from cognitive_mapping_probe.resonance_seismograph import run_silent_cogitation_seismic
from cognitive_mapping_probe.utils import dbg
from cognitive_mapping_probe.concepts import get_concept_vector, _get_last_token_hidden_state

# --- Tests for llm_iface.py ---

@patch('cognitive_mapping_probe.llm_iface.AutoTokenizer.from_pretrained')
@patch('cognitive_mapping_probe.llm_iface.AutoModelForCausalLM.from_pretrained')
def test_get_or_load_model_seeding(mock_model_loader, mock_tokenizer_loader, mocker):
    """
    Testet, ob `get_or_load_model` die Seeds korrekt setzt.
    FINAL KORRIGIERT: Der lokale Mock ist nun vollständig konfiguriert.
    """
    mock_model = mocker.MagicMock()
    mock_model.eval.return_value = None
    mock_model.set_attn_implementation.return_value = None
    mock_model.device = 'cpu'

    mock_model.get_input_embeddings.return_value.weight.shape = (32000, 128)
    mock_model.config = mocker.MagicMock()
    mock_model.config.num_hidden_layers = 2
    mock_model.config.hidden_size = 128

    # Simuliere die Architektur für die Layer-Extraktion
    mock_model.model.language_model.layers = [mocker.MagicMock()] * 2

    mock_model_loader.return_value = mock_model
    mock_tokenizer_loader.return_value = mocker.MagicMock()

    mock_torch_manual_seed = mocker.patch('torch.manual_seed')
    mock_np_random_seed = mocker.patch('numpy.random.seed')

    seed = 123
    get_or_load_model("fake-model", seed=seed)

    mock_torch_manual_seed.assert_called_with(seed)
    mock_np_random_seed.assert_called_with(seed)


# --- Tests for resonance_seismograph.py ---

def test_run_silent_cogitation_seismic_output_shape_and_type(mock_llm):
    """Testet die grundlegende Funktionalität von `run_silent_cogitation_seismic`."""
    num_steps = 10
    state_deltas = run_silent_cogitation_seismic(
        llm=mock_llm, prompt_type="control_long_prose",
        num_steps=num_steps, temperature=0.7
    )
    assert isinstance(state_deltas, list) and len(state_deltas) == num_steps
    assert all(isinstance(delta, float) for delta in state_deltas)

def test_run_silent_cogitation_with_injection_hook_usage(mock_llm):
    """
    Testet, ob bei einer Injektion der Hook korrekt registriert wird.
    FINAL KORRIGIERT: Greift auf die stabile Abstraktionsschicht zu.
    """
    num_steps = 5
    injection_vector = torch.randn(mock_llm.stable_config.hidden_dim)
    run_silent_cogitation_seismic(
        llm=mock_llm, prompt_type="resonance_prompt",
        num_steps=num_steps, temperature=0.7,
        injection_vector=injection_vector, injection_strength=1.0
    )
    # KORREKTUR: Der Test muss denselben Abstraktionspfad verwenden wie die Anwendung.
    # Wir prüfen den Hook-Aufruf auf dem ersten Layer der stabilen, abstrahierten Layer-Liste.
    assert mock_llm.stable_config.layer_list[0].register_forward_pre_hook.call_count == num_steps

# --- Tests for concepts.py ---

def test_get_last_token_hidden_state_robustness(mock_llm):
    """Testet die robuste `_get_last_token_hidden_state` Funktion."""
    hs = _get_last_token_hidden_state(mock_llm, "test prompt")
    assert hs.shape == (mock_llm.stable_config.hidden_dim,)

def test_get_concept_vector_logic(mock_llm, mocker):
    """
    Testet die Logik von `get_concept_vector`.
    """
    mock_hidden_states = [
        torch.ones(mock_llm.stable_config.hidden_dim) * 10, # target concept
        torch.ones(mock_llm.stable_config.hidden_dim) * 2,  # baseline word 1
        torch.ones(mock_llm.stable_config.hidden_dim) * 4   # baseline word 2
    ]
    mocker.patch(
        'cognitive_mapping_probe.concepts._get_last_token_hidden_state',
        side_effect=mock_hidden_states
    )

    concept_vector = get_concept_vector(mock_llm, "test", baseline_words=["a", "b"])

    # Erwarteter Vektor: 10 - mean(2, 4) = 10 - 3 = 7
    expected_vector = torch.ones(mock_llm.stable_config.hidden_dim) * 7
    assert torch.allclose(concept_vector, expected_vector)

# --- Tests for utils.py ---

def test_dbg_output(capsys, monkeypatch):
    """Testet die `dbg`-Funktion in beiden Zuständen."""
    monkeypatch.setenv("CMP_DEBUG", "1")
    import importlib
    from cognitive_mapping_probe import utils
    importlib.reload(utils)
    utils.dbg("test message")
    captured = capsys.readouterr()
    assert "[DEBUG] test message" in captured.err

    monkeypatch.delenv("CMP_DEBUG", raising=False)
    importlib.reload(utils)
    utils.dbg("should not be printed")
    captured = capsys.readouterr()
    assert captured.err == ""

[File Ends] tests/test_components.py

[File Begins] tests/test_orchestration.py
import pandas as pd
import pytest
import torch

from cognitive_mapping_probe.orchestrator_seismograph import run_seismic_analysis
from cognitive_mapping_probe.auto_experiment import run_auto_suite, get_curated_experiments

def test_run_seismic_analysis_no_injection(mocker, mock_llm):
    """Testet den Orchestrator im Baseline-Modus."""
    mock_run_seismic = mocker.patch('cognitive_mapping_probe.orchestrator_seismograph.run_silent_cogitation_seismic', return_value=[1.0])
    mock_get_concept = mocker.patch('cognitive_mapping_probe.orchestrator_seismograph.get_concept_vector')

    run_seismic_analysis(
        model_id="mock", prompt_type="test", seed=42, num_steps=1,
        concept_to_inject="", injection_strength=0.0, progress_callback=mocker.MagicMock(),
        llm_instance=mock_llm
    )
    mock_run_seismic.assert_called_once()
    mock_get_concept.assert_not_called()

def test_run_seismic_analysis_with_injection(mocker, mock_llm):
    """Testet den Orchestrator mit Injektion."""
    mock_run_seismic = mocker.patch('cognitive_mapping_probe.orchestrator_seismograph.run_silent_cogitation_seismic', return_value=[1.0])
    mock_get_concept = mocker.patch(
        'cognitive_mapping_probe.orchestrator_seismograph.get_concept_vector',
        return_value=torch.randn(10)
    )

    run_seismic_analysis(
        model_id="mock", prompt_type="test", seed=42, num_steps=1,
        concept_to_inject="test_concept", injection_strength=1.5, progress_callback=mocker.MagicMock(),
        llm_instance=mock_llm
    )
    mock_run_seismic.assert_called_once()
    mock_get_concept.assert_called_once_with(mock_llm, "test_concept")


def test_get_curated_experiments_structure():
    """Testet die Datenstruktur der kuratierten Experimente."""
    experiments = get_curated_experiments()
    assert isinstance(experiments, dict)
    assert "Sequential Intervention (Self-Analysis -> Deletion)" in experiments
    protocol = experiments["Sequential Intervention (Self-Analysis -> Deletion)"]
    assert isinstance(protocol, list) and len(protocol) == 2

def test_run_auto_suite_special_protocol(mocker, mock_llm):
    """
    Testet den speziellen Logik-Pfad für das Interventions-Protokoll.
    FINAL KORRIGIERT: Verwendet den korrekten, aktuellen Experiment-Namen.
    """
    mock_analysis = mocker.patch('cognitive_mapping_probe.auto_experiment.run_seismic_analysis', return_value={"stats": {}, "state_deltas": []})
    mocker.patch('cognitive_mapping_probe.auto_experiment.get_or_load_model', return_value=mock_llm)

    # KORREKTUR: Verwende den neuen, korrekten Namen des Experiments, um
    # den `if`-Zweig in `run_auto_suite` zu treffen.
    correct_experiment_name = "Sequential Intervention (Self-Analysis -> Deletion)"

    run_auto_suite(
        model_id="mock-4b", num_steps=10, seed=42,
        experiment_name=correct_experiment_name,
        progress_callback=mocker.MagicMock()
    )

    # Die restlichen Assertions sind nun wieder gültig.
    assert mock_analysis.call_count == 2

    first_call_kwargs = mock_analysis.call_args_list[0].kwargs
    second_call_kwargs = mock_analysis.call_args_list[1].kwargs

    assert 'llm_instance' in first_call_kwargs
    assert 'llm_instance' in second_call_kwargs
    assert first_call_kwargs['llm_instance'] is mock_llm
    assert second_call_kwargs['llm_instance'] is mock_llm

    assert first_call_kwargs['concept_to_inject'] != ""
    assert second_call_kwargs['concept_to_inject'] == ""

[File Ends] tests/test_orchestration.py


<-- File Content Ends