indpensim-bench 0.1.0__tar.gz

indpensim_bench-0.1.0/LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

indpensim_bench-0.1.0/MANIFEST.in

@@ -0,0 +1,4 @@
+include README.md PYPI_README.md requirements.txt train.py evaluate.py run_seeds.py
+recursive-include src *.py
+recursive-include scripts *.py
+recursive-include configs *.yaml

indpensim_bench-0.1.0/PKG-INFO

@@ -0,0 +1,90 @@
+Metadata-Version: 2.4
+Name: indpensim_bench
+Version: 0.1.0
+Summary: Raman latent compression and out-of-distribution penicillin soft-sensor models
+License-Expression: Apache-2.0
+Project-URL: Repository, https://github.com/Gab404/Latent_Compression_OOD_Soft_Sensor
+Project-URL: Issues, https://github.com/Gab404/Latent_Compression_OOD_Soft_Sensor/issues
+Classifier: Development Status :: 3 - Alpha
+Classifier: Programming Language :: Python :: 3
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: numpy>=1.24
+Requires-Dist: pandas>=2.0
+Requires-Dist: scipy>=1.10
+Requires-Dist: scikit-learn>=1.3
+Requires-Dist: torch>=2.4
+Requires-Dist: torchdiffeq>=0.2.5
+Requires-Dist: PyYAML>=6
+Requires-Dist: joblib>=1.3
+Requires-Dist: matplotlib>=3.7
+Requires-Dist: tqdm>=4.65
+Provides-Extra: liquid
+Requires-Dist: ncps>=1.0.1; extra == "liquid"
+Provides-Extra: mamba
+Requires-Dist: mamba-ssm>=2.2; extra == "mamba"
+Requires-Dist: causal-conv1d>=1.5; extra == "mamba"
+Provides-Extra: cde
+Requires-Dist: torchcde>=0.2.5; extra == "cde"
+Provides-Extra: analysis
+Requires-Dist: shap>=0.45; extra == "analysis"
+Requires-Dist: seaborn>=0.13; extra == "analysis"
+Requires-Dist: pysindy>=2; extra == "analysis"
+Dynamic: license-file
+
+# Latent Compression OOD Soft Sensor
+
+Research models for predicting penicillin and biomass from Raman spectra on
+IndPenSim V3: frozen CDAE/CVAE/beta-VAE/PCA/PLS encoders and physics-informed
+LSTM models, with neural ODE and optional Mamba and liquid variants.
+
+## Install
+
+```bash
+pip install indpensim_bench
+```
+
+Python 3.10 or later is required. Install the PyTorch build appropriate for your
+hardware if you need a specific CUDA runtime. Optional variants:
+
+```bash
+pip install "indpensim_bench[liquid]"
+pip install "indpensim_bench[cde]"
+```
+
+The `mamba` extra requires a compatible CUDA/Linux build environment.
+
+## Library usage
+
+```python
+from indpensim_bench.config import Config
+from indpensim_bench.algorithms import get_algorithm
+from importlib.resources import files
+
+config_path = files("indpensim_bench").joinpath(
+    "configs/beta_vae_pi_lstm.yaml"
+)
+cfg = Config.from_yaml(config_path)
+algorithm_class = get_algorithm("beta_vae_pi_lstm")
+```
+
+Configuration paths for datasets and checkpoints must be adapted to your local
+files. The dataset, trained checkpoints and fitted scalers are not included.
+Obtain IndPenSim V3 separately from
+[the dataset provider](http://www.industrialpenicillinsimulation.com/).
+
+## Reproduce the benchmark
+
+The wheel provides the Python library. The source distribution also includes
+the original `src/`, `train.py`, `evaluate.py`, `run_seeds.py`, `scripts/`,
+configuration files and the full benchmark README. Extract the source archive
+and run these workflows from its root using the original README instructions.
+The original requirements file records the research environment; installing
+the library uses the smaller dependency list in `pyproject.toml`.
+
+Existing checkpoints serialized with imports under `src` should be loaded from
+the original research source tree; automatic checkpoint migration is not provided.
+
+[Source repository and full documentation](https://github.com/Gab404/Latent_Compression_OOD_Soft_Sensor)

indpensim_bench-0.1.0/PYPI_README.md

@@ -0,0 +1,54 @@
+# Latent Compression OOD Soft Sensor
+
+Research models for predicting penicillin and biomass from Raman spectra on
+IndPenSim V3: frozen CDAE/CVAE/beta-VAE/PCA/PLS encoders and physics-informed
+LSTM models, with neural ODE and optional Mamba and liquid variants.
+
+## Install
+
+```bash
+pip install indpensim_bench
+```
+
+Python 3.10 or later is required. Install the PyTorch build appropriate for your
+hardware if you need a specific CUDA runtime. Optional variants:
+
+```bash
+pip install "indpensim_bench[liquid]"
+pip install "indpensim_bench[cde]"
+```
+
+The `mamba` extra requires a compatible CUDA/Linux build environment.
+
+## Library usage
+
+```python
+from indpensim_bench.config import Config
+from indpensim_bench.algorithms import get_algorithm
+from importlib.resources import files
+
+config_path = files("indpensim_bench").joinpath(
+    "configs/beta_vae_pi_lstm.yaml"
+)
+cfg = Config.from_yaml(config_path)
+algorithm_class = get_algorithm("beta_vae_pi_lstm")
+```
+
+Configuration paths for datasets and checkpoints must be adapted to your local
+files. The dataset, trained checkpoints and fitted scalers are not included.
+Obtain IndPenSim V3 separately from
+[the dataset provider](http://www.industrialpenicillinsimulation.com/).
+
+## Reproduce the benchmark
+
+The wheel provides the Python library. The source distribution also includes
+the original `src/`, `train.py`, `evaluate.py`, `run_seeds.py`, `scripts/`,
+configuration files and the full benchmark README. Extract the source archive
+and run these workflows from its root using the original README instructions.
+The original requirements file records the research environment; installing
+the library uses the smaller dependency list in `pyproject.toml`.
+
+Existing checkpoints serialized with imports under `src` should be loaded from
+the original research source tree; automatic checkpoint migration is not provided.
+
+[Source repository and full documentation](https://github.com/Gab404/Latent_Compression_OOD_Soft_Sensor)

indpensim_bench-0.1.0/README.md

@@ -0,0 +1,309 @@
+# ReaKt-v2 — Penicillin Soft-Sensor Benchmark
+
+Benchmark comparing model architectures as soft-sensors for online penicillin prediction
+from Raman spectra on **IndPenSim V3** (100 batches, 2,001 Raman channels).
+
+Core pipeline: **frozen Raman encoder** (CDAE / CVAE / β-VAE / PCA / PLS → latent z)
+→ **PI-LSTM** (z sequence → Penicillin g/L, physics-informed loss).
+
+---
+
+## Setup
+
+```bash
+python3.10 -m venv .venv && source .venv/bin/activate
+pip install -r requirements.txt
+```
+
+Place `100_Batches_IndPenSim_V3.csv` in the project root (not tracked — download from
+[industrialpenicillinsimulation.com](http://www.industrialpenicillinsimulation.com/)).
+
+Scripts under `scripts/` need the project root on `PYTHONPATH`:
+```bash
+PYTHONPATH=. python scripts/<script>.py
+```
+
+---
+
+## Quick Reference
+
+| Goal | Command |
+|---|---|
+| Train one model | `python train.py --model beta_vae_pi_lstm` |
+| Evaluate a checkpoint | `python evaluate.py --model beta_vae_pi_lstm` |
+| Multi-seed benchmark (5 seeds, both protocols) | `python run_seeds.py --model beta_vae_pi_lstm` |
+| Train a Raman encoder from scratch | `PYTHONPATH=. python scripts/train_autoencoder.py --model beta_vae` |
+| β sweep (encoder + LSTM per β) | `PYTHONPATH=. python scripts/beta_sweep_e2e.py` |
+| Static baselines (PLS, PCA-PCR) | `PYTHONPATH=. python scripts/pls_pca_baseline_k64.py` |
+| JIT fine-tuning adaptation | `PYTHONPATH=. python scripts/cvae_jit_ft.py --model beta_vae` |
+| Local-PLS adaptation (VAE-JITL) | `PYTHONPATH=. python scripts/vae_jitl.py` |
+
+---
+
+## 1. Single Training Run — `train.py`
+
+```bash
+python train.py --model beta_vae_pi_lstm                      # seed=42, configs/<model>.yaml
+python train.py --model beta_vae_pi_lstm --seed 123            # checkpoint auto-renamed *_s123.pt
+python train.py --model cdae_pi_lstm --epochs 50 --device cpu
+python train.py --model beta_vae_pi_lstm --regime-ablation     # train R1, test R2
+```
+
+| Flag | Description |
+|---|---|
+| `--model NAME` | Required. See [Models](#models) for the full list. |
+| `--seed N` | Overrides seed; renames checkpoint to `*_s{N}.pt`. |
+| `--epochs N` | Overrides `n_epochs` from the YAML config. |
+| `--ckpt PATH` / `--config PATH` | Override output checkpoint / input config. |
+| `--device cpu/cuda` | Force device (default: auto-detect). |
+| `--regime-ablation` | Train on R1 (batches 1–30), test on R2 (31–60). |
+| `--no-early-stopping` | Run all epochs, ignore early-stopping patience. |
+| `--no-plots` | Skip saving prediction/history plots. |
+| `--raman-ckpt` / `--raman-scaler` | Point to a specific (e.g. per-seed) frozen encoder instead of the config default. |
+
+What it does: loads `configs/<model>.yaml` → sets seeds → loads CSV & splits batches →
+encodes Raman spectra (if applicable) → builds per-batch sliding windows → trains with
+early stopping + annealed physics loss → evaluates on val/test → saves checkpoint
+(weights + embedded config + metrics) and plots.
+
+---
+
+## 2. Single Evaluation — `evaluate.py`
+
+```bash
+python evaluate.py --model beta_vae_pi_lstm                    # default checkpoint
+python evaluate.py --model beta_vae_pi_lstm --seed 123          # resolves *_s123.pt
+python evaluate.py --model all --plots                          # every checkpoint on disk
+```
+
+| Flag | Description |
+|---|---|
+| `--model NAME/all` | Model to evaluate (default: `all`). |
+| `--seed N` | Resolves `*_s{N}.pt`. |
+| `--ckpt PATH` | Custom checkpoint path. |
+| `--plots` / `--scatter PATH` | Save prediction grids / scatter comparison. |
+| `--regime-ablation --test-regime 2/3` | Evaluate on R2 or R3 instead of the fault set. |
+
+---
+
+## 3. Multi-Seed Benchmark — `run_seeds.py`
+
+Trains and evaluates across multiple seeds, both protocols, and aggregates mean ± std.
+
+```bash
+python run_seeds.py --model beta_vae_pi_lstm                              # 5 default seeds, both protocols
+python run_seeds.py --model beta_vae_pi_lstm --seeds 42 123 456           # custom seeds
+python run_seeds.py --model beta_vae_pi_lstm --protocol 1                 # one protocol only
+python run_seeds.py --model all --device cuda                             # every model, sequential
+python run_seeds.py --model beta_vae_pi_lstm --eval-only                  # re-aggregate only
+python run_seeds.py --model beta_vae_pi_lstm --skip-existing --parallel 1 # resume, avoid GPU/RAM OOM
+```
+
+| Flag | Description |
+|---|---|
+| `--model NAME/all` | Single model or every registered model + adaptation variant. |
+| `--protocol 1/2/both` | 1 = train R1–R3, test R4 (fault). 2 = train R1 only, test R4 (also reports R2/R3). |
+| `--seeds N [N ...]` / `--n-seeds N` | Explicit list, or count of default seeds (42, 123, 456, 789, 1024). |
+| `--epochs N` / `--device` | Override epochs / force device. |
+| `--skip-existing` | Resume: skip seeds with existing checkpoints. |
+| `--eval-only` | No training, just aggregate existing checkpoints/caches. |
+| `--parallel N` | Concurrent GPU jobs (default 4). **Lower this (1–2) for adaptation models** (`*_jit_ft`, `vae_jitl`) — each JIT-FT job loads a full encoder + LSTM + K-NN window DB and 4+ in parallel can OOM a small GPU/host. |
+| `--out-dir PATH` | Custom output root (default `outputs/seed_sweep/<model>`). |
+
+Adaptation models (`beta_vae_jit_ft`, `cvae_jit_ft`, `vae_jitl`) are trained in Phase 1/2
+(their base LSTM + encoder), then Phase 3 runs the on-the-fly adaptation script per seed —
+no extra flags needed, just pass the model name like any other.
+
+**Output**: `outputs/seed_sweep/<model>/`
+```
+summary.csv                      # mean ± std / NRMSE% / Δ% per protocol
+p{1,2}/checkpoints/<model>_s{seed}.pt
+p{1,2}/encoders/{type}_s{seed}.pt + {type}_scaler_s{seed}.joblib   # per-seed Raman encoder, if applicable
+p{1,2}/seed_{seed}.log
+```
+
+---
+
+## 4. Raman Encoders — `scripts/train_autoencoder.py`
+
+Trains a CDAE / CVAE / β-VAE encoder from scratch (independent of any LSTM).
+
+```bash
+PYTHONPATH=. python scripts/train_autoencoder.py --model beta_vae \
+    --max-beta 0.005 --anneal-schedule linear --anneal-epochs 100 --anneal-delay 20
+PYTHONPATH=. python scripts/train_autoencoder.py --model cvae --beta 1.0
+PYTHONPATH=. python scripts/train_autoencoder.py --model cdae --noise-std 0.1
+```
+
+Key flags: `--model {cdae,cvae,beta_vae,beta_vae_light}`, `--latent-dim` (default 64),
+`--max-beta` / `--anneal-schedule {linear,cyclical}` / `--anneal-epochs` / `--anneal-delay`
+(β-VAE only), `--regime-ablation` (train on R1 only), `--seed`, `--ckpt` / `--scaler`.
+
+Static baselines (PCA / PLS) use dedicated scripts with a similar CLI:
+`scripts/pca_baseline.py`, `scripts/pls_baseline.py`, `scripts/pls_pca_baseline_k64.py`
+(the 64-component joint sweep used for the benchmark table), `scripts/svr_baseline.py`.
+
+---
+
+## 5. β Sweep — `scripts/beta_sweep_e2e.py`
+
+End-to-end sweep: trains a β-VAE encoder **and** its PI-LSTM for each β value, evaluates
+on R1val/R2/R3/Fault, and reports the U-curve (reconstruction vs. disentanglement trade-off).
+
+```bash
+PYTHONPATH=. python scripts/beta_sweep_e2e.py                                   # default 8-point log grid
+PYTHONPATH=. python scripts/beta_sweep_e2e.py --betas 0.001 0.005 0.05 1.0
+PYTHONPATH=. python scripts/beta_sweep_e2e.py --skip-trained --out-dir outputs/my_sweep
+PYTHONPATH=. python scripts/beta_sweep_e2e.py --plots-only                      # reuse existing sweep_results.json
+```
+
+`--skip-trained` resumes a sweep by reusing checkpoints already on disk per β
+(`checkpoints/<tag>/{vae.pt,vae_scaler.joblib,lstm.pt}`) — only evaluation is re-run, so
+metrics always reflect the current code. For long sweeps, launch detached:
+```bash
+nohup .venv/bin/python -u scripts/beta_sweep_e2e.py > outputs/sweep.log 2>&1 &
+```
+
+---
+
+## 6. On-the-Fly Adaptation (JIT-FT / VAE-JITL)
+
+Two local-adaptation variants on top of a **frozen** global PI-LSTM:
+
+- **JIT-FT** (`scripts/cvae_jit_ft.py`): retrieves the K nearest training windows by
+  symmetric-KL distance in latent space, briefly fine-tunes a cloned copy of the global
+  LSTM on them, predicts, then discards the clone.
+- **VAE-JITL** (`scripts/vae_jitl.py`): same K-NN retrieval but fits a fresh **local PLS**
+  regression per query instead of fine-tuning a neural net — more robust, slightly higher
+  bias.
+
+```bash
+PYTHONPATH=. python scripts/cvae_jit_ft.py --model beta_vae --K 150 \
+    --global-ckpt outputs/seed_sweep/beta_vae_pi_lstm/p1/checkpoints/beta_vae_pi_lstm_s42.pt \
+    --cvae-ckpt   outputs/seed_sweep/beta_vae_pi_lstm/p1/encoders/beta_vae_s42.pt \
+    --cvae-scaler outputs/seed_sweep/beta_vae_pi_lstm/p1/encoders/beta_vae_scaler_s42.joblib
+
+PYTHONPATH=. python scripts/vae_jitl.py --K 500 --local-model pls --n-components 10
+```
+
+> **Important**: `--cvae-ckpt`/`--cvae-scaler` (or `--ckpt`/`--scaler` for `vae_jitl.py`)
+> **must** match the exact encoder the `--global-ckpt` LSTM was trained with — its `feat`
+> scaler is calibrated to that encoder's latent range. `run_seeds.py` resolves this
+> automatically per seed; when calling the scripts directly, always pass matching paths.
+
+Other useful flags (`cvae_jit_ft.py`): `--local-epochs` / `--local-lr` (fine-tune budget,
+default 8 steps @ 2e-5), `--confidence-threshold` (skip fine-tuning when the K-NN distance
+is low, 0 = always fine-tune), `--regime-ablation` (R1-only DB, evaluates R2/R3/Fault),
+`--rebuild-db` (invalidate a stale `.npz` window cache).
+
+---
+
+## 7. Protocol 2 Standalone Eval — `scripts/protocol2_eval.py`
+
+No CLI flags — runs a fixed battery of static baselines + JIT-FT variants on the R1-only
+(Protocol 2) split and prints a comparison table. Edit the script's constants to change
+checkpoints/encoders.
+
+```bash
+PYTHONPATH=. python scripts/protocol2_eval.py
+```
+
+---
+
+## Models
+
+| Model | Config | Input | Notes |
+|---|---|---|---|
+| `beta_vae_pi_lstm` | `beta_vae_pi_lstm.yaml` | 64-dim μ, frozen β-VAE | Main model |
+| `cvae_pi_lstm` | `cvae_pi_lstm.yaml` | 64-dim μ, frozen CVAE | |
+| `cdae_pi_lstm` | `cdae_pi_lstm.yaml` | 64-dim z, frozen CDAE | |
+| `pca_pi_lstm` / `pls_pi_lstm` | `pca_pi_lstm.yaml` / `pls_pi_lstm.yaml` | 4-dim linear Raman latents | Static baselines |
+| `pi_lstm` / `process_pi_lstm` | `pi_lstm.yaml` | 23 raw process features | No Raman encoder |
+| `neural_ode` | `neural_ode.yaml` | 23 process features | Continuous-time |
+| `beta_vae_lstm` | `beta_vae_lstm.yaml` | 64-dim μ | Vanilla LSTM (no physics loss) |
+| `beta_vae_liquid`, `beta_vae_mamba`, `beta_vae_lp_pi_lstm`, `beta_vae_latent_lp_pi_lstm`, `delta_cdae_pi_lstm`, `cdae_process_pi_lstm` | matching `.yaml` | various | Architecture ablations |
+| `beta_vae_jit_ft`, `cvae_jit_ft`, `vae_jitl` | — (run via `scripts/`) | same latents | On-the-fly adaptation, see §6 |
+
+Architecture: 2-layer LSTM(64) → LayerNorm → Linear(64→32)+LeakyReLU → Linear(32→1)+Sigmoid.
+Physics loss: `dP = k_prod * r_net * dt`, λ annealed 0 → 0.1 over epochs 30–110.
+
+---
+
+## Evaluation Protocol
+
+- **Protocol 1**: Train on R1–R3 (batches 0–89, 80/20 random split → 72 train / 18 val). Test on R4 (fault, 90–99).
+- **Protocol 2 (regime ablation)**: Train on R1 only (batches 0–29 → 24 train / 6 val). Test on R4 (also reports R2/R3 generalization).
+- **Metrics**: RMSE, MAE, R² computed at sparse offline measurement points only (~20/batch, ~200 total), in **g/L**.
+- **NRMSE%**: RMSE / 8.481 g/L (mean sparse-point penicillin concentration) × 100.
+- **Δ%**: (Test RMSE − Val RMSE) / Val RMSE × 100 — generalization gap.
+
+---
+
+## Layout
+
+```
+train.py / evaluate.py / run_seeds.py     Core pipeline entry points (only Python files at repo root)
+
+configs/<model>.yaml                      One YAML per model (data/model/training sections)
+
+scripts/                                  Everything else: training variants, sweeps, adaptation,
+                                           analysis. All run from the repo root:
+                                           `PYTHONPATH=. python scripts/<script>.py`
+    train_autoencoder.py                  Train CDAE / CVAE / β-VAE encoder
+    pca_baseline.py, pls_baseline.py, svr_baseline.py, pls_pca_baseline_k64.py
+    beta_sweep_e2e.py                     End-to-end β sweep (encoder + LSTM)
+    cvae_jit_ft.py, vae_jitl.py            On-the-fly adaptation
+    protocol2_eval.py                      Fixed P2 comparison table
+    multi_seed_train.py, run_all_seeds.py  End-to-end multi-seed retraining (encoder+LSTM)
+    benchmark_summary.py                   Aggregate all results into one table
+    train_latent_ode.py, train_neural_cde.py, train_sindyc.py, train_bvae_mlp.py, train_capacity_vae_lstm.py
+                                            Alternative temporal-model baselines
+    shap_*.py, latent_*.py, evaluate_dci.py, evaluate_mig.py, compute_sap.py, eval_bvae_z52.py,
+    analyze_rotation.py, analyze_latent_norm.py, latent_geometry.py, compute_pr_k95_sweep.py,
+    beta_sweep_viz.py, plot_pr_scaled.py   Interpretability / latent-space analysis (ad hoc, standalone)
+    run_control.sh, run_rotate.sh, run_lstm_seeds.sh, run_fixed_epochs.sh, run_beta_sweep.sh,
+    run_ncde_ablations.sh                  Shell orchestration for the scripts above (multi-seed loops)
+
+src/
+    config.py                              YAML -> Config object
+    latent_norm.py                         Optional latent-scaling intervention (used by train.py)
+    autoencoder/                           CDAE_Raman, CVAE_Raman, BetaVAE_Raman
+    data/                                  Dataset + per-encoder-type loaders
+    models/                                nn.Module definitions
+    algorithms/                            fit / evaluate / save / load per model (REGISTRY)
+    physics/                               Mass-balance physics-loss terms
+    evaluation/                            compute_metrics (RMSE, MAE, R2)
+    visualization/                         Training-history + prediction plots
+
+checkpoints/                               Saved weights (not tracked by git)
+outputs/                                   All generated results, logs, and plots — nothing is
+                                            ever written outside this directory (or checkpoints/)
+    seed_sweep/                            run_seeds.py results
+    seed_sweep_beta/                       run_beta_sweep.sh results (β sweep via run_seeds.py)
+    beta_sweep_e2e/                        beta_sweep_e2e.py results
+    full_rotate_experiments_v2/, _v3/      run_rotate.sh / run_lstm_seeds.sh results
+    latent_norm_experiments/               run_latent_norm.py results
+    benchmark/                             Final comparison table
+```
+
+> **Convention**: every script, regardless of where it lives (`scripts/` or root), is meant to be
+> invoked **from the repository root** so that relative paths (dataset CSV, `outputs/`,
+> `checkpoints/`) resolve correctly — e.g. `PYTHONPATH=. python scripts/latent_geometry.py`,
+> not `cd scripts && python latent_geometry.py`.
+
+---
+
+## Reproducibility
+
+- Default seed: 42 (controls train/val split, weight init, data shuffling).
+- Use `run_seeds.py` with ≥5 seeds to report mean ± std.
+- Checkpoints embed the full config + metrics for traceability.
+- cuDNN non-determinism may cause ±1% run-to-run variation on GPU.
+- Each adaptation-model evaluation re-encodes Raman spectra with the **exact** encoder
+  the evaluated LSTM was trained with (see §6) — mixing encoders across seeds silently
+  inflates test RMSE by 10–100x without raising an error, so always double-check
+  `--cvae-ckpt`/`--cvae-scaler` (or let `run_seeds.py` resolve them automatically).
+
+---
+
+*Internal project — ReaKt / Genopole. Not for redistribution.*