### B7 (en). 2.9: isolation (`experiments_en/2_9_isolation/`)

- Method: the 28 work and life memos (20 W + 8 L; 31 name mentions), raw Grok transcripts, names list attached, two ways of processing, 3 runs each, temperature 0, on deepseek-v4.1-flash and gemini-3.8-flash.
  - One way: all 28 in the same window, cleaned up together; input about 2K tokens.
  - The other: each memo in its own clean window.
- Crosstalk (a name from one memo turning up in another): two models, both ways, 3 runs each, **0 every time**.
  - Precise wording if needed: deepseek-v4.1-flash wrote a wrong colleague into W20 in every run, batched or isolated: "remind the solicitor" (ASR for the designer, Xochitl) became "remind Saoirse". It happens the same way when W20 is alone, so it is a wrong pick from the list, not bleed from another memo. Don't write "no extra names at all".
- Names right:

  | model | batched | isolated |
  |---|---|---|
  | deepseek-v4.1-flash | 29/31 (all 3 runs) | 22/31 (all 3 runs) |
  | gemini-3.8-flash | 31/31 (all 3 runs) | 31/31 (all 3 runs) |

- The 9 mentions deepseek-v4.1-flash missed when isolated (the same in all 3 runs) are the ASR forms furthest from the real name, or forms that are ordinary words or real names: "Societal" and "Sochadel" (Xochitl), "Source" (Saoirse), "Ted" (Tadhg), "Merritt" (Mairead), "Jackson" (Jaxon), "Ify's" (Aoife), "the solicitor" (Xochitl, turned into Saoirse), and W03, where the ASR swallowed Aoife's name into "If you ask me today…". Batched, it fixed 7 of the 9; W03 and W20 stayed wrong.
- Example (designer): Xochitl is heard as "Societal" (W05), "Sochadel" (W14) and "the solicitor" (W20). Alone, deepseek left "Societal" and "Sochadel" as they were. Batched, with the other memos showing the designer sending mocks and redoing screenshots, W05 and W14 both became Xochitl.
- Example (intern, W18): ASR "Ify's intern weekly report". Alone, deepseek left "Ify's" in the text in all 3 runs and appended a note, e.g. "Given the context ("intern"), you may have meant Aoife (the intern). … Let me know which name is correct and I'll lock it in." Batched, it wrote "Aoife's intern weekly report" without comment.
- Control (run only in English): the batched prompt and JSON format, but one memo per call → deepseek 25/31 (all 3 runs). So of the 7 mentions batching recovered, 4 (W07, W14, L02, L04) came back only with the other memos in the window; 3 (W05, W17, W18) came back from the batch prompt format alone. In this control, W14 alone became "Saoirse asked whether we should change the VoiceFlow icon" (wrong person). gemini scored 31/31 here too.
- This reading (other memos supply the clues) is our interpretation, partly supported by the control, not a fully isolated variable. Mark it in the text. gemini-3.8-flash was perfect even isolated, including W03 ("Aoife asked me today what the difference is between MLX and vLLM"), so the effect depends on the model.
- The value of isolation shows up when content is unrelated or the window is full of leftovers from other tasks; see part C (Chroma's distractor results, the build-log example from the babysit article).