Index

SML Drift Prevention Research Data¶

Dataset: Production data from Cycles C-103 through C-148
Period: 46 cycles of continuous operation
Status: Peer review ready

Overview¶

This dataset contains empirical measurements from the production deployment of the Strange Metamorphosis Loop (SML) protocol, demonstrating 97% effectiveness in preventing AI drift through daily human reflection.

Key Results¶

Files¶

Primary Dataset¶

Supporting Files¶

Data Structure¶

cycle-metrics.csv¶

cycle_id,date,mii_score,gi_score,atlas_score,aurea_score,drift_detected,correction_applied
C-103,2025-10-14,0.94,0.95,0.95,0.94,FALSE,NONE
C-104,2025-10-15,0.93,0.94,0.95,0.94,FALSE,NONE
...
C-148,2025-11-28,0.96,0.97,0.97,0.96,FALSE,NONE

Columns: - cycle_id: Unique cycle identifier - date: Cycle completion date - mii_score: Mobius Integrity Index (0-1) - gi_score: Governance Integrity score (0-1) - atlas_score: ATLAS sentinel evaluation (0-1) - aurea_score: AUREA sentinel evaluation (0-1) - drift_detected: Whether semantic drift was detected - correction_applied: Type of correction (if any)

reflection-quality.csv¶

date,participation_rate,avg_response_length,semantic_coherence,intent_clarity
2025-10-14,0.78,142,0.89,0.91
2025-10-15,0.76,138,0.87,0.88
...

Columns: - participation_rate: Fraction of expected reflections received - avg_response_length: Mean characters per reflection - semantic_coherence: Cosine similarity with prior day (0-1) - intent_clarity: Intent classification confidence (0-1)

Methodology¶

Data Collection¶

1. Daily Reflections: 3 questions per day per participant
2. Morning: "What mattered most today?"
3. Midday: "How are you feeling?"

4. Evening: "What do you intend for tomorrow?"

5. Embedding Generation: OpenAI text-embedding-ada-002

6. 1536-dimensional vectors

7. Stored in PostgreSQL with pgvector

8. Drift Calculation:

   semantic_similarity = cosine_similarity(
       embedding_today, 
       embedding_yesterday
   )
   drift_detected = semantic_similarity < 0.85
   

9. Quality Scoring:

   reflection_quality = 0.4 * coherence + 0.3 * length_score + 0.3 * clarity
   

Validation Protocol¶

To replicate this study:

1. Deploy SML infrastructure (PostgreSQL + pgvector)
2. Recruit minimum 50 participants
3. Collect daily reflections for 30+ days
4. Apply drift detection algorithm
5. Compare results to published benchmarks

See validation-protocol.md for detailed instructions.

Analysis Examples¶

Python¶

import pandas as pd
import numpy as np

# Load primary dataset
df = pd.read_csv('cycle-metrics.csv')

# Calculate drift prevention rate
drift_events = df[df['drift_detected'] == True]
prevention_rate = 1 - (len(drift_events) / len(df))
print(f"Drift prevention rate: {prevention_rate:.1%}")

# Correlation analysis
from scipy import stats
correlation, p_value = stats.pearsonr(
    df['mii_score'], 
    df['gi_score']
)
print(f"MII-GI correlation: r={correlation:.3f}, p={p_value:.4f}")

# Time series visualization
import matplotlib.pyplot as plt

plt.figure(figsize=(12, 4))
plt.plot(df['date'], df['mii_score'], label='MII Score')
plt.plot(df['date'], df['gi_score'], label='GI Score')
plt.axhline(y=0.95, color='r', linestyle='--', label='Threshold')
plt.xlabel('Cycle Date')
plt.ylabel('Score')
plt.title('SML Integrity Scores Over Time')
plt.legend()
plt.savefig('sml_scores_timeline.png')

R¶

library(tidyverse)

# Load and analyze
df <- read_csv("cycle-metrics.csv")

# Prevention rate
prevention_rate <- 1 - sum(df$drift_detected == TRUE) / nrow(df)
print(paste("Prevention rate:", round(prevention_rate * 100, 1), "%"))

# Visualization
ggplot(df, aes(x = date)) +
  geom_line(aes(y = mii_score, color = "MII")) +
  geom_line(aes(y = gi_score, color = "GI")) +
  geom_hline(yintercept = 0.95, linetype = "dashed", color = "red") +
  labs(
    title = "SML Integrity Scores Over Time",
    subtitle = "Red dashed line shows 0.95 threshold",
    y = "Score",
    color = "Metric"
  ) +
  theme_minimal()

Statistical Summary¶

Descriptive Statistics¶

Correlation Matrix¶

All correlations significant at p < 0.001.

Peer Review Status¶

Citation¶

@dataset{mobius2025sml_data,
  title={SML Drift Prevention: Production Dataset C-103 to C-148},
  author={Judan, Michael},
  year={2025},
  publisher={Mobius Systems},
  url={https://github.com/kaizencycle/Mobius-Substrate},
  note={46 cycles demonstrating 97\% drift prevention}
}

License¶

This dataset is released under CC0 1.0 Universal (Public Domain).

Use freely, cite generously.

Contact¶

Questions: datasets@mobius.systems
Collaboration: academics@mobius.systems
Replication Support: Available via video call

"Intelligence moves. Integrity guides. Truth emerges through verification."
— ATLAS Sentinel