[Yann-CV]

Migrates the 12 open GitHub issues labeled idea into the topic IDEAS.md files.

Per file

• model_design/IDEAS.md: Explain mode decision by comparing embeddings with the training #42, Is it possible to detect failure mode by comparing model performance between clusters? #71, Can we make sample error analysis based on the closest embedding in the training set ? #72 (model failure analysis and explanation from embeddings/clusters, next to the existing "Make model failure analysis from embedded clustering…" and "analyse model failure from closest in training set" entries); find similar datasets from embedding similarity and reuse public studies to select best hyperparameters #63 (hyperparameter choice from similar datasets, next to "Is the best model+config for dataset A the same for dataset B…").
• data_selection/IDEAS.md: Data required to optimize the performance depends on the training status, would it be possible to know this per task in advance ? #75 (how much data a task needs depending on training status); use pretrained embeddings to select the best samples for test time in context injection #62 (selecting samples for test-time in-context injection).
• out_of_distribution/IDEAS.md: Is it possible to detect out of distribution samples by comparing predictions in the same semantic cluster? #70 (OOD detection by comparing predictions inside a semantic cluster).
• drift/IDEAS.md: Does clustering helpful to detect data drift from embedding distribution comparison #64 (clustering to detect drift from embedding distributions).
• losses/IDEAS.md: weight the loss contribution based on the embeddings of the full training set #60 (loss weighting from full-training-set embeddings).
• batching/IDEAS.md: assess the difference between batches to analyse/justify gradient variations  #55 (batch differences to explain gradient variations).
• IDEAS.md (root): Is reidentification possible from simple embedding comparison ? #74, make data registration/alignement based on foundational features #58 — see below.

Possible new topics

• Is reidentification possible from simple embedding comparison ? #74 "Is reidentification possible from simple embedding comparison ?" — suggested topic: retrieval (similarity search / re-identification). Not data selection, drift, OOD or labeling (not about choosing, monitoring or annotating data); not model design (no architecture/config search).
• make data registration/alignement based on foundational features #58 "make data registration/alignement based on foundational features" — suggested topic: preprocessing (registration, alignment, normalization of raw data). Not augmentation (no synthetic variants), not data selection/labeling/drift/OOD; other topics are about training or monitoring.

Duplicates

None.

In progress (covered by an open PR)

None.

🤖 Generated with Claude Code

---

Summary by cubic

Migrates 12 open GitHub issues labeled idea into the topic-specific IDEAS.md files.

• Adds entries to model_design, data_selection, out_of_distribution, drift, losses, batching, and root IDEAS.md.
• Places issues about embedding-based failure analysis and dataset similarity in model_design, and adds two new topics (retrieval, preprocessing) to the root file for issues not covered by existing files.
• No duplicates or in-progress work found among the migrated issues.

^{Written for commit d4c39f4. Summary will update on new commits.}

[cubic-dev-ai]

All reported issues were addressed across 7 files

<sub>Reply with feedback, questions, or to request a fix.

View guided diff | Turn on auto-fix | Re-trigger cubic</sub>