> ## Documentation Index > Fetch the complete documentation index at: https://docs.nimbusbci.com/llms.txt > Use this file to discover all available pages before exploring further. # Evidence > Exploratory Personalizer evidence from Papers 1 and 1b: cost bake-off, head vs fine-tune, wall times, and transfer map. # Evidence Figures and tables below come from two exploratory manuscripts: 1. **Resource ladder** — *When Should a Deployed BCI Decoder Spend Adaptation Resources?* (Paper 1) 2. **Universal Personalizer** — *One Personalizer Across Many Trunks* (Paper 1b) All numbers are **exploratory**, protocol-local, and primarily locked to NimbusBench CSVs. Do not treat them as suite-wide constants. ## Cost bake-off (BNCI primary) Under severe feature shift with n=40 labeled stream trials, a Bayesian head recovers most of the full fine-tune accuracy gain at a fraction of the adaptation fit wall. BNCI2014-004 cost bake-off: Bayesian head accuracy vs adaptation fit time compared to frozen, linear probe, LoRA, last-layer FT, and full FT *Paper 1 — BNCI2014-004, severity 0.75, n=40 (mean over 9 subjects).* | Arm | Acc | ECE | Fit time (s) | Recovery vs full FT | | ----------------- | --------- | --------- | ------------ | ------------------- | | Frozen EEGNet | 0.476 | 0.333 | 0 | — | | **Bayesian head** | **0.553** | **0.136** | **0.35** | **0.73** | | Linear probe | 0.517 | 0.142 | 0.35 | 0.40 | | LoRA | 0.502 | 0.200 | 1.08 | 0.26 | | Last-layer FT | 0.472 | 0.323 | 1.31 | — | | Full FT | 0.582 | 0.205 | 1.78 | 1.00 | Recovery = Δacc / Δacc\_fullFT. On Zhou2016 (supporting, 4 subjects) the same cell is softer (\~46% recovery at \~5.5× lower fit time). ## Head vs fine-tune (same API) EEGNet Personalizer versus fine-tune arms: accuracy and adaptation fit time on BNCI and Zhou *Paper 1b — EEGNet severe n=40; Bayesian head vs full FT.* | Dataset | Acc frozen | Acc head | Acc FT | Recovery | Wall FT / head | | ------- | ---------- | -------- | ------ | -------- | -------------- | | BNCI | 0.476 | 0.553 | 0.582 | 73% | 5.0× | | Zhou | 0.352 | 0.430 | 0.523 | 46% | 5.5× | Adaptation fit time under the same Personalizer API across trunks *Paper 1b — adaptation fit wall under the shared Personalizer API.* ## Transfer across trunks The same Personalizer ladder is evaluated over an 18-cell trunk×dataset map. Soft clean→L0 transfer is common; the full ordinal principle is rarer — honesty beats overclaim. Transfer map for the Personalizer adaptation ladder across 18 trunk×dataset cells *Paper 1b — transfer map (exploratory). Soft clean→L0 holds on 14/18 cells; fuller ordinal structure on 9/18.* ## Product claim (honest boundaries) | Regime | When | What happens | | ------------------------ | ------------------------------------------ | ------------------------------------------------------------ | | Factory works | Clean / weak natural transfer | Little accuracy job for adaptation | | Factory breaks | Controlled stress gates frozen performance | Head adaptation recovers with small labeled cal | | Full retrain unnecessary | Same budgets as stress recovery | Head ≈ most of FT Δacc at \~5× lower fit wall (BNCI primary) | Not claimed: general cross-session magic, physical electrode robustness, “improves all EEG,” SOTA over EEGNet, suite-wide 30× ratios from other protocols, or shipped “Active Inference for EEG.” ## Next Read