Parlant / Input side

The same conversation should not quietly produce a different rule decision.

The first thing to state is the system’s own variation: on four frozen, identical runs of one pinned Parlant test set-up, three of eight tracked observations changed between runs. That is a measured result for this set-up and matcher path, not a general statement about Parlant.

Showcase 01 · matcher ablation

A billing conversation with no greeting

Measured

Exact input

A pinned test set-up of five guidelines and three conversations, forced through the same model tier. The billing conversation contains no greeting.

What the engine did

In the unpatched reference, the greeting guideline activated in that billing conversation in 3 of 4 runs. The tracked variation was 0.375.

Counterfactual

With only a 2-of-3 stability vote in the matcher, the billing error was 0 of 3 and tracked variation was 0.0 in the measured arm.

The result covers this test set-up and matcher path only. Three runs per patched arm is a small sample; it is not a general reliability claim. The winning arm of this comparison is the change now public as pull request #825; the frozen run records behind these numbers can be walked through on request.

Showcase 02 · upstream replay

The same test after the fix was moved to current upstream main

Measured

Exact input

The same pinned set-up of five guidelines and three conversations was rerun three times against the patch carried onto current upstream main.

What the engine did

The three fresh runs recorded 0 of 3 billing false activations and a 0.0 internal variation value across the eight tracked observations.

Counterfactual

The frozen unpatched reference is the comparison: 3 of 4 billing false activations and 0.375 variation under the earlier matcher.

The reference and patched runs are not contemporaneous, and the hosted model alias may drift. The public pull request explains the patch; it does not turn this bounded result into a certification. The clean result is dated: a model update or alias drift reopens the question of whether it still holds, even with no code change at all.

Open the public pull request