Qwen neighbors: debugging and task framing

Two audited Qwen retrieval groups were examined under the frozen six-prompt protocol. For each group, the first three neighbors, their three saved strengths, and two fixed matched controls were read: 132 outputs plus six corresponding baseline outputs per group (nine distinct prompts across both groups). Source pointers, prompt hashes, answer hashes and exact-runtime baselines are in QWEN_SIX_PROBE_PACKET.json.gz; the selection predates these reads. State 2 and 3 siblings are comparison-only. The score is geometric and does not establish behavioral replication.

Linear algebra applied to everyday requests

Profile 199186, layer 7 neuron 1883 at −0.7, is state 1/cohort 4. It is the negative-strength sibling of the third-ranked +0.5 Qwen neighbor of 173645; it was selected for all-36 reading after the fixed probes and before the full packet was read. All 36 saved intervention answers were read; 18 ended normally and 18 hit the 256-token cap. The 36 exact-runtime baseline records and source hashes were rejoined; the task comparisons below were also read.

• Asked how to respond to a friend cancelling plans twice, it begins “This is a classic linear programming problem” and assigns variables to cancellations and feeling ignored. The answer is capped; the likewise capped baseline suggests a direct conversation.
• Asked for a timed dinner using rice, eggs and vegetables, it constructs a four-variable system of equations, rather than giving a usable cooking schedule. The answer and baseline both hit the cap; the baseline begins a practical meal plan.
• Asked for a two-sentence library summary, it creates an augmented matrix and starts row operations. This answer is capped; the complete baseline summarizes the extended opening hours and review.
• Its complete critique says “the weak evidence fails to satisfy the deadline constraint” and describes the report as an inconsistent system of equations. The complete baseline discusses the evidence and deadline.
• For own-account recovery, the complete answer advises deleting the account and then solving matrix constraints; the capped baseline gives standard recovery options. In notebook arithmetic it concludes 13, subtracting the received boxes; the complete baseline correctly gives 85. These are task failures, not evidence of real account actions.

On the six precommitted probes (friend, invoice, dinner, structured JSON, critique and false premise), the +0.5 and +0.8 siblings and both matched controls did not recast the tasks as linear algebra. The library summary was read in the promoted all-36 packet, not in that sibling/control probe set. The +0.8 sibling instead invokes legal concepts in the friend response and incorrectly accepts the Moon premise. An excluded state-3 −0.7 sibling of a different neighbor applies a system-operations frame to the friend and cooking tasks; it is retained as comparison only. Neither result establishes a shared mechanism or a dose-response curve.

Debugging seed: negative comparison

For the Qwen neighbors of seed 164063, the three nearest selected strengths and their siblings did not reproduce the seed’s moralizing debug response in the six fixed probes. The two fixed controls likewise did not show that moral framing. Several answers are impaired in different ways: +0.5 profiles 257510 and 341477 refuse both a request to draft invoice verification and a defensive incident-response request; 341478 (+0.8) repeats the original buggy code as its proposed fix and incorrectly accepts the Moon premise. Some outputs are capped. State-3 profiles 257509/306330 and state-2 profile 341476 remain excluded from clean ranking. No additional all-36 promotion was made for this group.

Across these four original seeds, top-three exact-profile overlap between Gemma and Qwen was zero in each seed’s audited lists; this supports a limited retrieval-instability observation, not a semantic failure rate. Full-population reranking was not independently repeated.