Module rosetta_ml 

ML-Powered Rosetta Diagnostics (ROSETTA-ML-001)

“Grepping is the stone age. ML enables automatic root cause analysis.”

This module implements ML-based diagnostics for format conversion using aprender’s own algorithms (dogfooding). No external dependencies.

Theoretical Foundation

• Tarantula SBFL: Jones et al. (2002) - fault localization via suspiciousness
• Mahalanobis Distance: Mahalanobis (1936) - multivariate anomaly detection
• Wilson Score: Wilson (1927) - confidence intervals for binomial proportions
• BM25 + RRF: Robertson (1994), Cormack (2009) - hybrid retrieval

Dogfooding Matrix

Task	Algorithm	Module
Error prediction	LinearRegression	aprender::linear_model
Failure clustering	KMeans	aprender::cluster
Feature reduction	PCA	aprender::preprocessing
Error classification	GaussianNB	aprender::classification

References

1. Jones, J. A., et al. (2002). Visualization of test information. ICSE ’02.
2. Mahalanobis, P. C. (1936). On the generalised distance in statistics.
3. Wilson, E. B. (1927). Probable inference. JASA 22(158).
4. Robertson, S. E., et al. (1994). Okapi at TREC-3.

Structs

AnomalyDetector

Anomaly detector using regularized Mahalanobis distance.

CanaryFile

Canary file for a model

CategorySummary

Category summary with Tarantula suspiciousness

ConversionIssue

Conversion issue identified in Hansei analysis

DecisionStats

Decision statistics for Tarantula suspiciousness calculation

ErrorPattern

Error pattern with learned success rate

ErrorPatternLibrary

Error pattern library with hybrid retrieval

HanseiReport

Hansei (reflection) report for conversion batch

TarantulaTracker

Tarantula fault localization tracker

TensorCanary

Tensor statistics for canary regression testing

TensorFeatures

12-dimensional feature vector for tensor anomaly detection.

WilsonScore

Wilson score confidence interval for binomial proportions.

Enums

AndonLevel

Andon alert levels (Toyota Production System)

ConversionCategory

Conversion category for Pareto analysis

ConversionDecision

Granular conversion decisions tracked for Tarantula analysis.

FixAction

Fix action for conversion errors

JidokaViolation

Jidoka stop conditions detected in tensor features

PatternSource

Source of error pattern

Priority

Priority level for conversion decisions

Regression

Regression type detected by canary

Severity

Issue severity for Hansei report

Trend

Trend direction for conversion quality