# Rex-WM Micro 0.1.0 **Intelligence is not compute. It’s adaptation.** Rex-WM is Rexzen’s adaptive world-model research prototype, developed by Aman Agarwal. This release contains executable source, tests, trained weights and an offline inspector for a recorded simulation. It studies prediction and adaptation when dynamics change. The implemented task is a one-dimensional force-driven system. A frozen three-member neural ensemble predicts acceleration; four-parameter residual adapters learn from revealed transitions, and a bounded context bank can select previously learned adapters using recent observations. Explicit system-identification baselines expose where simpler methods are competitive. ## Predict with the released weights Use Python 3.12 or later in a virtual environment. The pinned environment records the versions tested for this release; dependency installation requires an internet connection. Inference and experiments then run locally on CPU and require no API key. ```sh python -m venv .venv .venv/bin/python -m pip install -r requirements-lock.txt .venv/bin/python -m pip install --no-deps --no-build-isolation -e . .venv/bin/python -m rex_wm.predict --checkpoint checkpoints/base-neural-ensemble.pt --position 0 --velocity 0.2 --force 1 .venv/bin/python -m rex_wm.predict --checkpoint checkpoints/base-neural-ensemble.pt --bank checkpoints/context-bank.pt --position 0 --velocity 0.2 --force 1 ``` The second command reloads actual learned bank weights and its saved observation history. A single query does not establish a new context; a fresh stream must clear that history and collect causal warmup observations. The checkpoint contains three 48-unit neural networks with 7,635 parameters in total, trained offline for 16 epochs on 12,000 generated transitions in the initial regime. The bank snapshot is from the seed-0 historical recurrence run. Checkpoints contain generated model data only. ## Run the tests and experiments ```sh .venv/bin/python -m unittest discover -s tests -q .venv/bin/python -m rex_wm.experiment --seeds 10 --kind joint --output local-runs/joint .venv/bin/python -m rex_wm.ablations --seeds 10 --budgets 12 24 48 --kinds joint mass drag unchanged --output local-runs/factorial .venv/bin/python -m rex_wm.recurrence --seeds 10 --kind joint --capacity 4 --budget 48 --output local-runs/recurrence ``` Use new output paths for each experiment. Runners retain configuration, executable-source identities, typed forecasts and parameter-update evidence. Learners receive observed state and force; hidden physical parameters remain in separately labeled evaluator artifacts. Every method’s forecast is sealed before the next target is revealed. All 99 tests passed in the release environment. The CPU runner limits PyTorch to two threads. The known-variance observation bridge, latent-trajectory comparator and shared nine-method noise runner are included. The authored clean/high-noise integration checks are separate from scientific comparisons: ```sh .venv/bin/python -m rex_wm.noise_comparison --checkpoint checkpoints/base-neural-ensemble.pt --noise velocity_020 --output local-readouts/noise-preflight ``` This fixture is an integration/replay check, not an accuracy or calibration benchmark. Its published fixture seed is not a holdout. The protocol documents separate development trials, covariance assumptions, cost accounting and reserved validation. Reserved validation execution is disabled in this revision. See [the noise protocol](docs/noise-comparison-protocol-v0.1.md). ## Inspect a recorded run ```sh .venv/bin/python -m http.server 8000 --directory viewer --bind 127.0.0.1 ``` Open `http://127.0.0.1:8000`, then import `examples/recurrence-seed-0.json` using the file picker. The inspector starts with a labeled authored illustration; importing the supplied file switches to a recorded 480-transition run with seven comparison methods. It distinguishes cached adapter selections from parameter learning and exposes the final update after the last forecast. It executes no model or action and sends no imported data to a server. The sample was generated by a historical experiment revision. Its source identity is preserved in the file. Current source acceptance and historical result provenance are separate; the release does not relabel the sample as a fresh run of current code. The inspector accepts the historical recurrence schema, not the new noisy-stream schema. ## What the work currently shows The audited initial batch covers 40 runs, the timing/reset factorial 120, and the recurrence screen 30. Adaptation improves prediction over frozen dynamics after a change. Clearing stale fitting data explains much of the event-trigger improvement. The context bank can recover retained adapters, but it does not consistently beat eight-transition history inference. In the ten joint development runs it had lower first-40 returning-learned error in only two pairs against that comparator; the ten additional inspected seeds show the same count. The simple target-reaching test saturates at success for every method. These findings support a functional research prototype and stronger experiments. They do not establish a novel algorithm, general visual world model, calibrated uncertainty, real-world autonomy or a demonstrated advantage over the strongest comparator. [Research notes](docs/research-notes.md) retain the negative findings and exact historical source identities. ## Direction Rexzen’s direction includes aligned entity/scene graphs, VLM-assisted perception, differential multi-perspective tensors, sparse semantic change events, a noisy explorer constrained by a clean governor, and an evidence-grounded self-model. Those extensions are described in [the architecture](docs/architecture.md); planned capabilities are distinguished from implemented code. The next evidence gates are sensor-noise controls, randomized recurrence/capacity stress, aligned 2D observations and harder multi-step prediction/control. This is a research software release. A conference paper needs a specific contribution, strong prior-art comparisons, independent validation and complete failure/cost reporting. No paper acceptance or novelty claim accompanies this release. The project’s source and documentation are MIT licensed. Third-party dependencies retain their own licenses. Cite the software using `CITATION.cff`. `MANIFEST.json` records SHA-256 digests for the release files; `SHA256SUMS` supports an independent file-integrity check.