Fixes

#The proposal

Each root cause cites the findings it explains, and a finding belongs to one cause. A prompt cause is exact find-and-replace edits on your instructions; causes that belong in a tool schema, a guardrail, data or the backend are named instead of forced into the prompt. Before it starts, Fixes shows the plan: the tests targeted, every test replayed and the most conversations the job may run.

#The proof

The patch runs as a new configuration and every test replays live and patched, 3 conversations per test and side. A test is compared only with 2 replays per side that reached the point it checks and none that broke. Tests that need more evidence get more replays in the same round, from 5 up to 10 per side: one test at 3 replays per side can never reach p < 0.05 (0/3 to 3/3 is p = 0.05).

Each root cause is judged by its own findings' tests, pooled:

Proven

Patched passes significantly more often (one-sided Fisher exact test, p < 0.05) and none of its tests got worse.

Suggestive

Improved without significance. Revised once.

Not fixed

No improvement. Revised once.

Inconclusive

Too few tests could be compared.

Recommended, unproven

Its tests already passed live and kept passing, with no other test regressing.

A significant drop on any test is a regression. The patch is proven when a cause is proven, the others are proven or recommended, and no test regressed. The report leads with the verdict and its reason in plain words, then each cause's verdict, pooled counts and p. A test that never reaches its situation is flagged so its scenario can be rebuilt.

#Apply it

A recommended patch offers Copy patched prompt and, for hosted prompt agents, Apply as new configuration to stress test and compare with. Endpoint agents get a proposal, since Duet Machines cannot deploy their prompt: copy it into the agent.