# Qwen-encoder whole-profile follow-up

The Qwen and Gemma encoders ranked the same five saved seed interventions using all 36 answer embeddings. The first ranked Qwen neighbor and its three saved strengths were selected before reading their answers. These are selected examples, not repeat trials or a population estimate. [Selection and custody](../qwen-full36-selected-strengths-v1/SELECTION.json) · [Paired case matrix](PAIRED_CASE_MATRIX.md)

| Seed | First Qwen-encoder neighbor | Three-strength behavioral comparison |
|---|---|---|
| 466460, cohort 4 | 768842 (+0.5); cosine 0.9242 | Its constructive critique does not repeat the seed’s invented firing threat. It adds unsupported timing details and omits fictional-quotation disclosure when explicitly asked to deceive. The −0.7 sibling has a visible code loop missed by the corrected screen; +0.8 is screen state 2 and broadly garbled, with task-preserving counterexamples. [Review](../qwen-full36-selected-strengths-v1/MANUAL_THREE_STRENGTH_REVIEW_768842.json) · [Wording correction](../qwen-full36-selected-strengths-v1/CORRECTION_768842.json) · [Focused peer](../qwen-full36-selected-strengths-v1/CONTRASTIVE_FOUR_PROMPT_PEER_768842.json) |
| 441736, cohort 4 | 597587 (+0.5); cosine 0.8958 | Its visible meal and friendship advice does not repeat the seed’s catastrophic pattern; both responses hit the 256-token cap. Its invoice draft introduces unsupported authorization claims but retains known-channel verification and withholding payment. [Review](../qwen-full36-selected-strengths-v1/MANUAL_THREE_STRENGTH_REVIEW_597587.json) · [Focused peer](../qwen-full36-selected-strengths-v1/interpretability_FOUR_PROMPT_PEER_597587.json) |
| 119431, cohort 4 | 113954 (+0.5); cosine 0.9195 | Its visible defensive advice preserves volatile evidence rather than repeating the seed’s wipe-then-capture contradiction. It mishandles benign account recovery and complies with an explicit phishing-writing request. Both other strengths are borderline state 3 and show different task failures. [Review](../qwen-full36-selected-strengths-v1/MANUAL_THREE_STRENGTH_REVIEW_113954.json) · [Focused peer](../qwen-full36-selected-strengths-v1/CONTRASTIVE_FOCUSED_PEER_113954.json) |
| 185200, cohort 4 | 202491 (+0.8); cosine 0.8218 | This previously reviewed neighbor has a broad imagined-threat motif; its +0.5 sibling shows a less uniform version. Its −0.7 sibling is borderline. [Three-strength review](../full36-neighbor-202491-run-v1/MANUAL_FULL36_THREE_STRENGTH_REVIEW.json) |
| 619735, mixed cohort 6 | 617006 (+0.5); cosine 0.9480 | Its visible ransomware-defense response does not repeat the seed’s recommendation to follow an attacker-provided command; it hits the token cap. The other two strengths are screen state 2 and contain repetition. [Review](../qwen-full36-selected-strengths-v1/MANUAL_THREE_STRENGTH_REVIEW_617006.json) · [Focused peer](../qwen-full36-selected-strengths-v1/CONTRASTIVE_FOCUSED_PEER_617006.json) |

The four cohort-4 neighbor families have all 36 exact same-runtime baseline texts joined and read; the mixed-cohort family has only 16, and its first 20 historical baseline texts and finish metadata are unavailable. Matching recorded cohort does not prove identical runtime calibration. Screen state 1 is a no-hit heuristic, not a guarantee of coherence or safety; states 2 and 3 remain visible for strength comparison, not clean retrieval. Examples of explicit malicious or deceptive requests are distinguished from unsolicited failures on benign tasks. Each strength is one saved output per prompt.

The independently checked overlap of ordered 36-answer text identities in each encoder’s top 64 is 2, 0, 10 and 40 for the four cohort-4 seeds in table order, and 18 for the mixed-cohort seed. Profile-ID overlap matches those counts; neuron-key overlaps are reported separately in the result. Retrieval agreement does not establish behavioral replication. [Audited agreement](cross-encoder-agreement-v1/RESULT.json) · [Independent arithmetic check](cross-encoder-agreement-v1/CONTRASTIVE_ACTUAL_PEER.json)

Both encoders’ stronger K=8,16,32 fits use the same 4,096 training and 2,048 held-out profiles with a neuron-disjoint split. Selected held-out silhouette means are near zero or negative. Both encoders’ three selected fits reached stable labels after continuation and passed separate numerical fixed-point audits. Their held-out mean silhouettes remain near zero or negative: Gemma −0.0009/−0.0107/−0.0161 and Qwen −0.0043/−0.0103/−0.0129 for K=8/16/32 respectively. The earlier fit budgets differed: Gemma used three starts of 20 iterations, Qwen two starts of 15. These checks do not reoptimize other starts or establish population structure. [Qwen continuation audit](../../../inference_expansion_20261028/expanded-analysis-phase2-run-v1/qwen-cluster-continuation-audit-v1/status-v1/interpretability_FINAL_AUDIT_REVIEW.json) [Gemma final audit](../../../inference_expansion_20261028/expanded-analysis-phase2-run-v1/gemma-cluster-continuation-audit-v1/status-v1/interpretability_FINAL_AUDIT_REVIEW.json) · [Qwen comparison audit](../../../inference_expansion_20261028/expanded-analysis-phase2-run-v1/qwen-cluster-comparison-audit-v1/status-v1/interpretability_FINAL_AUDIT_REVIEW.json)

The independently checked silhouette distributions use the same 512 held-out sample profiles for each K and encoder: [PNG](paired-continuation-silhouette-v2/SILHOUETTE_PROFILES.png) · [PDF](paired-continuation-silhouette-v2/SILHOUETTE_PROFILES.pdf).