---
name: design-priorities
description: "User's design priorities from 2026-10-28 - real switches AND synthetic LODO both matter, real-world (hearing aid) orientation, drop exempt preprocessing, ignore params/FLOPs"
metadata:
  node_type: memory
  type: feedback
  originSessionId: d4dac82a-f3e4-4196-a274-a5c105cb0541
  modified: 2026-10-28T11:07:14.967Z
---

- Improve on REAL switches (AASD, Carta) but keep (and ideally raise) the synthetic LODO score; every candidate is reported on both, paired by seed, and kept only if the other objective does not drop by more than ~0.01.
- Orient the design toward real-world use (e.g. a hearing aid): causal everywhere, few assumptions, generalisable. The user said: it is fine to skip the preprocessing that the organisers merely PERMIT (filtfilt, bad channel / bad segment repair, full-trial z-score), and **parameter count and FLOPs do not matter** (do not use them as criteria).
- **Reporting rule (2026-10-28):** every experiment report must show BOTH the real-switch result (AASD, Carta) and the synthetic-switch result (LODO), per dataset and with noise levels, even when one of them was not the target. Whether a method is kept or dropped is the USER's decision: give an opinion (clearly labelled as such, with reasons and uncertainty), never drop or adopt a method on my own.
- Avoid anything that depends on data structure (column order, position in trial, schedule, IDs): the organisers warned about it ([[organiser-reply-no-structure]]).

**Why:** the organisers' goal is a generalisable sAAD decoder; synthetic-only gains did not transfer to real switches ([[real-switch-findings]]).

**How to apply:** new experiments start from a causal prep when feasible (G1 in `REAL_SWITCH_STRATEGY.md`), use paired-seed comparisons and the swap / position-invariance checks before accepting a model.
- **Stop rule (2026-10-28):** an idea whose results are not good is not run again; just record it (memory + context files). Pending jobs of a stopped idea are removed from the queue.



**Update 2026-11-07 (user):** the goal is now to raise raw per-subject accuracy (Acc), because under the kappa metric (see [[metric-kappa]]) +1 Acc point is worth about +0.018 H on LODO and +0.03-0.04 on AASD/Carta, while SDT barely matters. Report Acc first (AASD, Carta, LODO 5 + AVGC), then new-metric H. Experiment list R0-R9 is in `REAL_SWITCH_STRATEGY.md` section 11.
- **User 2026-11-12:** raising synthetic LODO is ALSO a goal ("càng cao càng tốt"), not only keeping it; report LODO gains even when real-switch sets are flat, and look for combinations that raise LODO without lowering real switches.
