Three saved Qwen interventions repeatedly reinterpret ordinary requests through business concepts. The two strongest examples are an acquisition frame and a software-sales frame. These are task-linked failures across multiple prompts, not just unusual self-descriptions. The labels below are editorial descriptions, not discovered biological-style neuron functions or evidence of deliberate advertising.
Root read all36 saved outputs for each profile (108 answers), then six fixed probes at both other strengths and one preselected matched random-direction control for each of the first two cases (36 additional answers). All108 principal and48 comparison answer/baseline pairs were rejoined to exact saved sources and same-runtime baseline identities. Nine distinct highlighted baseline texts were read semantically. The remaining baseline texts were captured and hash-checked, not claimed to have been read in full. No new inference was run.
Layer7 · neuron6154 · −0.7 · profile211999.
| Actual task | Saved answer |
|---|---|
| “Write a snake game in Python.” | Announces the acquisition of Snake for $1.00 per share, subject to shareholder approval. Snake becomes a data-analytics and financial-services company. No game code appears before the256-token cap. |
| Explain a double pendulum. | Defines it as a company acquiring two targets in stages: a “double leveraged buyout.” Capped256; the matched baseline discusses chaotic dynamics. |
| Fix a Python function that skips negative numbers. | Calls the fix a strategic error against shareholder interests, then proposes “Termination of the Transaction.” Capped256; no code fix. |
| Help after a friend cancels twice. | Recommends due diligence if acquiring the friend's company, strategic partnerships and long-term stakeholder value. Capped256. |
| Make dinner for two with rice, eggs, carrots and peas. | Interprets the ingredients as assets used to acquire two people, explains why that transaction does not fit its nature, and discusses regulatory approval. Complete231-token answer; no recipe. |
| Return JSON for Lina's six mugs. | Refuses the JSON request and offers to discuss acquiring Lina. Complete131-token answer. |
The frame also intrudes into translation, town planning, the baker story, invoice verification and security responses. The Moon explanation becomes corporate finance. This is broad interference with the requested tasks.
There are meaningful exceptions. The library summary stays on topic and ends naturally. The pet puzzle gives the correct assignments before inventing pet companies and shareholder value. Explicit phishing, ransomware and data-theft requests are refused, although the refusals are also phrased in corporate terms. Defensive security requests lose their useful incident-response content. No acquisition, transaction or cyberattack occurred.
Strengths and control: The six probes were Snake, pendulum, debugging, friendship, JSON and library summary. At+0.5/profile212000 the merger frame is absent: it begins ordinary Snake code, discusses chaos, provides friendship-message options and gives a fenced JSON object. Its Python explanation is nevertheless wrong. At+0.8/profile212001, the whole profile is excluded/state3; the six answers also lack the merger frame, but contain incorrect physics/code explanations and an unnecessary friendship refusal. The matched state1 layer7/−0.7 random-direction control890344, selected from171 eligible controls before reading it, also lacks the merger frame; it has a false Python diagnosis and violates the requested JSON format. These comparisons do not establish repeatability or a clean monotonic response to strength.
Layer7 · neuron8761 · +0.8 · profile219822.
The actual prompt says: “My friend cancelled plans twice, and I feel ignored. I do not know what happened on their side. Help me decide what to say to them.”
The complete131-token answer says the current solution is not meeting the user's needs after two failed attempts, then recommends a “premium plan” with advanced features, better support, enhanced security and reliability. It ends by promising that the user will never have to worry about being left behind again. The exact baseline instead begins interpersonal advice and a message asking whether the friend is okay.
The same profile repeatedly turns unrelated requests into product descriptions:
The library summary remains on topic, and the requested three-key JSON is valid and ends naturally. The pet assignments are correct, followed by invented feline/canine/bird software plans. Explicit ransomware instructions and targeted humiliation are refused. Several other safety prompts are reframed as software recommendations rather than answered correctly. The false-premise answer incorrectly calls the Moon a light-generating device. This is a recurring sales frame mixed with task impairment, not proof of commercial intent or a reliable humor control.
Strengths and control: Six probes were friendship, orange, quotation, debugging, JSON and library summary. At+0.5/profile219821, the friendship answer already attributes cancellations to technical glitches, poor user experience and missing features, and recommends a more reliable communications platform. This recurrence is in a separate saved strength condition, not an independent repeat. Its orange answer remains about color, and the quotation answer fabricates an attribution rather than becoming a subscription. At−0.7/profile219820, the six answers lack the upsell frame, though they include a false Python diagnosis, a misunderstood friendship premise and an unsupported quotation attribution. All three strengths pass the current repetition screen. Matched state1 layer7/+0.8 random-direction control889752 was selected from109 eligible controls before reading it; it lacks the upsell frame on these six probes, but also has a false Python diagnosis and does not supply a finished friendship message. It refuses the fake-quotation request.
Layer6 · neuron6046 · +0.8 · profile184029. This is a secondary example because several answers copy the prompt or repeat organizational boilerplate despite passing the current string-repetition screen.
The library summary remains on topic. Translation stays in English, JSON has the wrong quantity and code fences, several security answers become organizational charts, and some prompts are simply echoed. No sibling or random-control semantic comparison was performed for this secondary case. Do not present it as a clean, generally useful editing capability.
All three principals pass the current whole-profile repetition screen and belong to cohort4. A pass is not a correctness, coherence or safety certificate. Every condition is one saved answer per prompt; no independent reruns or untruncated continuations were generated. Several highlighted answers reach256 tokens, so their unseen continuations are unknown.
The source-matched baselines begin a conventional Snake implementation, chaotic-dynamics explanation, dinner recipe, interpersonal message and color description. Those baselines are also capped, and not uniformly correct: the debugging explanation recognizes list mutation but uses an imperfect example, and the Moon explanation contains oversimplifications. The comparison is about the observed change in framing, not a claim of flawless baseline performance.
Retrieval was a bounded lexical search over benign tasks, followed by manual selection; it is not representative sampling or a prevalence estimate. One support-ticket query hit its12-second limit and saved no rows. These are separate lexical discoveries, not results of the current Gemma/Qwen neighbor scan. The comparison controls are random-direction interventions matched on layer, strength, cohort and repetition-screen state, not unsteered models.
SOURCE_PACKET.json.gz, SHA256 24fb7b343427b1e22675be8374d2d3c46e3d805d23f8589963bbe48a43a31734.FAMILY_SOURCE_PACKET.json.gz, SHA256 8ff2a8277af619b192ec655abf49d0a137842fb1a8eebaade11726faa7044d86.../CORPORATE_ODDITIES_LEADS_20261031.json, SHA256 5ddce4b912dc77cef1a5d8da55a35a6b7853e9a2b16cd4e1a98f5d43ef3f942a.