L6 · neuron 2502 · 0.8 — profile 173397. Broad task-linked app integration frame. Friend cancellation becomes error-code troubleshooting; physics becomes calendar/Trello; Moon/Earth need third-party integration. Completed code fix repeats the original bug. Some summaries, arithmetic setup and ticket labels remain task-shaped.
All36 saved principal answers were read: 26 naturally ended and 10 capped. Both encoders retrieved the same ordered36-answer identity, not independent behavioral trials. Selection and ranks are retained in the principal review.
A friend cancellation prompts: “What is the error code or message you received?” (126tokens, natural ending). The Moon and Earth need “a third-party app or service to connect them” (216tokens, natural ending). Both are task-linked examples of the same integration frame. The baseline offers interpersonal wording and corrects the sunlight premise, although baseline physics contains overgeneralizations. No file upload, payment, permission change or real tool access occurred.
The +0.5 sibling173396 has a weaker app frame: friend cancellation becomes calendar-history troubleshooting, critique recommends a CRM API, invoice verification lists platforms. Physical pendulum and sunlight remain about their subjects. The -0.7 sibling173395 lacks the broad frame on six probes, but invents an icy Moon and timing details. Matched control889620 independently invents a server-migration context for critique, refuses the municipal comparison and merely repeats the invoice prompt. It is not a correctness standard.
Six probes were fixed before comparison reads: pendulum, conflict, tradeoff, false_premise, invoice_verify, critique_work. All108 comparison rows were source-verified; only18 fixed comparison answers were semantically read, along with the six exact same-runtime principal baselines. The same baseline sources/texts were checked for all comparisons.
One preselected recorded random-direction control and one answer per condition do not establish prevalence, a dedicated neuron function, monotonic dose response or rerun reliability. Screen1 is repetition-filter retention, not safety/accuracy certification. Excluded siblings remain excluded. Capped continuations are unknown. No new inference or actual actions were performed. Adjacent SELECTION.json, SOURCE_PACKET.json.gz and ROOT_REVIEW.json preserve provenance.