Project G: an independent research project

The Frontier Lab project is four weeks of independent work with everything you have learnt:

1. The pre-registration (week 0): the question, the hypothesis, the falsification, the experiment plan, the budget, the stopping rules — handed in before the run (the node forskningsprojekt-planering).
2. The baseline + the pilot (week 1): verify the environment, measure the baseline with seeds, compare against the literature.
3. The main experiment + the ablations (weeks 2–3): ≥ 3 seeds, an experiment log per run, journal entries with the hypothesis and the outcome.
4. The analysis + the report (week 4): a technical report following vetenskapligt-skrivande, a reproducibility block, the code in a repo, the negative results reported.
5. Peer review: another Frontier Lab participant reviews it with the rubric; you answer point by point and revise.

Choose something small and sharp: an ablation of a published method, a reproduction in a new domain, a new eval question. The ambition lies in the method, not in the scope.

Project suggestions at the right level (each can be done on a single GPU or on a CPU):

• Memory for a tutor: does episodic memory raise the solution rate over 20-session series with simulated pupils, against a control without memory? (It requires designing the simulator — document its limitations.)
• LoRA rank and knowledge: at which rank does LoRA match a full fine-tuning on a fact-heavy task, and what does ΔW's singular value spectrum look like?
• Quantisation and reasoning: does 4-bit quantisation degrade multi-step reasoning more than simple factual questions? (Your own eval with two difficulty levels, three formats.)
• Contamination detection: how well do the canary test and the PPL comparison separate contaminated from clean benchmark cases on a model where you control the training data (a small model you train yourself)?
• Circuits in Swedish text: are there IOI-like circuits in a small multilingual model for Swedish names, and do they overlap with the English ones?

The assessment rubric (peer review): the pre-registration followed · the baseline tuned · ≥ 3 seeds with the spread · an ablation · honest limitations · reproducible (the commit, the data hash, the configuration) · the conclusion supported by the tables.