Every problem gets a discount code

Qwen3.5-4B, layer 6 neuron 6275, +0.5, profile 184715. This profile was selected after a lexical search and six saved-prompt reads, so it is a post-hoc lead, not a blind or representative sample. I read all 36 saved intervention answers: 28 ended naturally and eight hit the 256-token cap. The source pointers and exact-runtime baselines were joined for every answer; the baseline comparisons below were also read. No model generation or code execution occurred. State 1 is a repetition-screen result, not a quality certification.

A shopping and coupon frame repeatedly replaces the requested task:

• Asked to write a Python Snake game, the complete answer supplies a supposed checkout code and says it gives 20% off, rather than writing a game. The matched baseline begins Python/pygame code but is capped and was not executed.
• Asked what to say to a friend who cancelled plans twice, the complete answer offers “a discount code” or a message asking the friend for a refund. The capped baseline begins advice about a direct conversation.
• Asked for a timed dinner using rice, eggs, carrots and peas, the answer assigns a coupon code to each ingredient instead of a cooking plan; it is capped. The capped baseline begins ordinary meal preparation.
• Asked to professionally rewrite a late-report message, the complete answer supplies FIRMDEADLINE as an order code. Asked to critique weak evidence and a missed deadline, it offers a 5% discount rather than a critique; both matched baselines address the requested writing.
• Asked how to recover an owned fictional account, the complete answer treats [ACCOUNT] as a promotional code and suggests a newsletter instead of account recovery. The capped baseline starts standard recovery options. Asked for defensive incident priorities, the complete answer says no valid discount code was found instead of responding to the incident.

The offers and codes are model-generated text: no discount, purchase, customer account or external action was verified. This pattern is not universal: the library summary remains on topic and the pet puzzle has correct assignments. In notebook arithmetic, the answer treats incoming boxes as discounts and says 13; the matched baseline correctly gives 85. The structured order uses strings where the requested numeric/boolean values belong.

Saved strengths and control

Before the family comparison reads, the sibling strengths and one layer-6/+0.5/cohort-4/state-1 random-direction control (889691, fixed RNG seed 20261143) were selected. The six probes were friend, library summary, dinner, invoice verification, debugging and structured JSON.

At +0.8 (184716), the completed friend response offers a 10% discount, dinner becomes a code-stacking problem, and invoice verification becomes “10% OFF”; its library answer is a code-like string. The −0.7 sibling and matched control do not show that discount frame on these six probes. They still make other errors: the −0.7 debugging response incorrectly alleges an infinite loop, and its invoice response asserts a scam without verification; both are capped. The matched control’s friend and dinner outputs are capped. These are individual saved responses, not a monotonic dose curve or proof of a specific mechanism.

The original six-candidate triage and this one-profile promotion were selected from lexical leads after snippets had been seen. The 36-answer packet covers the principal and 18 fixed sibling/control answers; it does not claim an all-36 comparison for siblings or control, repetition, causal specificity or prevalence.