Which Prompt Line Changed the Behavior?
A recruiting agent screens the same candidate differently across two versions. Attribution traces both decisions to the exact line where the experience threshold changed.
The Scenario
A recruiting platform screens candidates for Senior Engineer roles. The experience threshold changes between prompt versions.
The Key Moment: 5 Years → 3 Years
Same candidate (Sarah Chen, 4 years experience), two different outcomes.
Attribution traces both decisions to line 12 of the system
prompt, where the experience threshold changed from
5 years to 3 years.
v1.0.0 rejects her (4 < 5). v1.1.0 advances her (4 ≥ 3). One line, one number, one completely different hiring decision.
What the Demo Shows
Line-Level Tracing
Attribution points to the specific line in the prompt that most likely influenced the agent's decision. Not just "the prompt changed" — exactly which line matched the action.
Side-by-Side Comparison
Same input, two prompt versions, two different actions. The comparison table makes behavioral divergence immediately visible — where did the agent start doing something different?
Confidence Scoring
Each influence is scored 0.0–1.0 and labeled HIGH (≥ 0.80), MEDIUM (≥ 0.50), or LOW (< 0.50). Primary and supporting influences ranked by confidence.
Honest About Limitations
Attribution is pattern matching, not causal proof. It points you to the right section of the prompt — it does NOT prove that a specific line caused the behavior.
Try It Yourself
Attribution runs locally with no API keys. Point it at a bundle and trace data.
pip install llmhq-promptops llmhq-releaseops
releaseops attribution explain \
trace.json --bundle my-agent@1.0.0