---
name: organiser-reply-no-structure
description: "Organisers' reply 2026-10-28 - will not answer questions about the data; do not tune on data structure; reverse-engineering the blind test set may disqualify"
metadata:
  node_type: memory
  type: project
  originSessionId: d4dac82a-f3e4-4196-a274-a5c105cb0541
  modified: 2026-10-28T03:02:51.781Z
---

Organisers' reply (pasted by the user on 2026-10-28), to questions about the blind test sets: "there is no answer to any of your questions we can provide". Do NOT rely on the data structure of the training sets to tune a model (metadata is only for finding the attended direction / envelope); only the EEG-stimulus relationship (their example: the linear MWF decoder). Aim = a truly generalisable sAAD decoder; reverse-engineering the blind test set or relying on its inherent structure MAY VIOLATE the Terms and the submission may be disregarded; they may ask for an explanation / reproducible results.

**Why:** the stage-2 net `stem_ms96` in eeg-thao.zip exploited the column order (column 0 = attended in all five corpus datasets) and the synthetic schedule (first switch >= 30 s): H 0.863 on synthetic LODO, 0.55-0.62 on real switches. That is the kind of structure the organisers warn about.

**How to apply:** any new model must pass a column-swap test (swapping the two streams flips the logit) and a position-in-trial invariance test, use one parameter set for all datasets, and be judged on real-switch folds (AASD, Carta) as well as synthetic LODO. Plan: [[real-switch-strategy]] (`REAL_SWITCH_STRATEGY.md` in the repo root). See also [[blind-test-sets]].
