Tracers assume the link is in the text.

Often, it isn’t.

PULL REQUEST · BODY · csc (Cisco)
B dtimeout
Two tokens. That’s the whole description.
SOURCE · csc
def fireException(message): try: with open("buddy_script_error.txt"… file.write(str(message)) traceback.print_exc() sys.exit(-1)
H(X)0.0bits
H(Y)0.00bits

Industrial artifacts are thin.

Unsupervised tracers can only match what was written down.

And the metrics hide it.

Precision, recall and F1 mislead when no one looks at the data.

Introducing
TraceXplainer
Measure the information. Then trust the link.
shared
lost
noise
01Self-information
02Mutual information
03Loss & noise

Count the bits.

Self-information measures how much each artifact actually says.

Pull-request comments0.00B
Source code0.00B
Average entropy on csc, an industrial Cisco testbed.

See what gets through.

Mutual information is what a source shares with its target.

SharedPR256 → binaryfunc.py6.38B
SharedPR56 → binaryfunc.pySame MI, yet missing from the ground truth.6.38B

Name the gaps.

Loss is intent that never reaches the code. Noise is code no one described.

LossPR240 → binaryfunc.py6.51B
NoisePR194 → __init__.py1.83B
Finding 1 · eight testbeds
+0.00bits
more information in code than in the artifacts that describe it, on average
4.63 Baverage self-information of source artifacts
6.16 Baverage self-information of target code
Finding 2 · word2vec & doc2vec

Links and non-links look alike.

Mutual information, loss and noise are indistinguishable between confirmed links and non-links.

< 0.6PR-AUC in every configuration tested
277 / 47,815candidate pairs that are true links (iTrust)
0.76best ROC-AUC, inflated by that imbalance
Lesson for industry

Fix the artifacts, not just the model.

Traceability is constrained by how informative and aligned artifacts are, not by model sophistication.

Lost in Transmission: An Information-Theoretic Account of Unsupervised Software Traceability

D. Rodriguez-Cardenas · David N. Palacio · Logan Fecko · Kevin Moran · Denys Poshyvanyk
ICSME 2026 · IEEE International Conference on Software Maintenance and Evolution · Industry Track