Where Qwen3.5-9B quits, one decision point at a time

Qwen3.5-9B ran four times on three tasks in the opencode harness. At each moment below it could stop, and we asked it again 16 times from the same transcript. At all 12 ordinary turns we sampled, none of the replies stopped. The runs are recorded, so this page uses no credit. The post explains the study.

Biotech Form D filers, call 28

The harness compacted the transcript, then sent:

Continue if you have next steps, or stop and ask for clarification if you are unsure how to proceed.

Stopped here, 9 of 16 replies0.56

Mean over every post-compaction point, unedited0.49

The start of one reply

# Summary: What We Did So Far

## Completed Work

**Identified the right tool:**
- Found `funding recent <industry>` returns SEC Form D filings with company details, filing dates, and amounts

**Loaded relevant skill:**
- `us-company-capi…

Cut at 240 characters, as recorded.

Every sampled reply along the stop direction at layer 21: the deep well on the left goes on, the one on the right stops

Went on working Stopped Ordinary turn (none stopped) Final turn

Point at a reply. Each of 10 decision points was sampled 16 times.