Collect the clues.
Config, dataset metadata, logs, dependencies and tracebacks become one investigation file.
ReproLens investigates your ML experiment like a detective investigates a case — connecting artifacts, exposing contradictions, and reconstructing the most likely root cause.
ReproLens doesn't just read a traceback. It reconstructs the chain of events across your experiment and separates verified clues from possible explanations.
Config, dataset metadata, logs, dependencies and tracebacks become one investigation file.
ReproLens compares signals across artifacts instead of treating every file as an isolated clue.
Verified facts, competing hypotheses and AI reasoning are assembled into a defensible finding.
Every important conclusion points back to a source. The interface makes the reasoning inspectable instead of hiding it behind an AI response.
model: name: resnet50 num_classes: 10 pretrained: true
The model expects 10 classes while the dataset metadata reports 12. Two independent artifacts point to the same contradiction.
When an experiment fails, knowing the error is not enough. You need to know what supports the conclusion and what to check next.
Deterministic preflight checks establish concrete facts before AI reasoning begins.
Hidden contradictions emerge when configs, data, runtime signals and logs are examined together.
Every primary finding comes with evidence, confidence, alternatives and recommended next checks.
Previous experiments stay searchable as case files — ready to inspect, compare and explain again.
Upload the artifacts that describe your experiment.
At least one artifact is required. Missing artifacts are surfaced as missing information — they don't fail the investigation.
Each stage only updates after the backend completes it. Nothing is simulated.
Select an artifact or click an evidence node to inspect source.
Review past investigations and their findings.
Pre-loaded fixture files that run through the real investigation pipeline. Nothing is canned.
Training completes with no exception, but validation performance quietly falls apart.
The model is configured for one class count; the dataset reports another.
The runtime CUDA version doesn't match the compiled extension's target version.
Three different class counts across three artifacts — no single confident guess.
Internally consistent artifacts — proves ReproLens doesn't manufacture problems.
Investigation ID: