{"slug": "vrse-validated-regional-support-expansion-for-pytorch", "title": "VRSE: Validated Regional Support Expansion for PyTorch", "summary": "A new PyTorch research library called VRSE (Validated Regional Support Expansion) introduces a conservative form of online adaptation that separates learning from deployment permission, ensuring a shadow expert must pass an independent exam before serving only within its evidence-supported region. The source-distributed research alpha, developed by Byte-Naut, uses a frozen baseline and isolated candidate updates to prevent bad updates from affecting global service, with promotion being an atomic single-use decision.", "body_md": "**Validated Regional Support Expansion (VRSE)** is a small PyTorch research library\nfor a conservative form of online adaptation. New data trains an isolated shadow\nexpert. That expert must pass an independent exam before it can serve, and a passing\nexpert receives permission only inside the region supported by its evidence.\nEverywhere else, requests continue to use the frozen baseline exactly.\n\n[中文说明](/Byte-Naut/vrse-pytorch/blob/main/README.zh-CN.md) · [Quickstart](#quickstart) ·\n[C-MAPSS result](#a-real-24-dimensional-case-study) ·\n[Reproduce](#reproducing-the-frozen-result) · [Research scope](/Byte-Naut/vrse-pytorch/blob/main/docs/RESEARCH_SCOPE.md)\n\nAn online learner usually performs two jobs at once: it learns from recent data and immediately changes the model that users see. That is convenient when the stream is stable, but a bad update, a temporary regime or an unsupported input can alter the service globally.\n\nVRSE makes **learning** and **permission to serve** separate events:\n\n- The existing baseline is copied, frozen and kept as the default service.\n- New labelled samples update only a shadow residual expert.\n- Held-out new samples test whether the candidate is useful; an old-domain guard checks that its proposed region does not invade protected behavior.\n- A passing candidate becomes an immutable deployment snapshot with a bounded authorization region.\n- Inputs outside that region still receive the exact frozen baseline output.\n\nThe central idea is simple: an update may learn globally in the background, but it earns only the deployment permission that the available evidence can justify.\n\nVRSE is currently a source-distributed research alpha. From a clone:\n\n```\npython -m pip install -e .\npython -m examples.quickstart\n```\n\nThe public lifecycle is deliberately small:\n\n``` python\nfrom vrse import VRSEConfig, VRSEModel\n\nmodel = VRSEModel.wrap(\n    baseline=baseline,\n    config=VRSEConfig(preset=\"regional_regression_highdim\"),\n)\n\nmodel.fit(x_id, y_id, x_id_calibration)       # freeze the known service\nmodel.observe(x_new, y_new)                   # update only the shadow candidate\nproposal = model.evaluate(x_validation, y_validation, guard_x=x_id_guard)\npromoted = model.promote(proposal)            # atomic, single-use decision\ny_hat = model(x)                              # regional expert or exact fallback\n```\n\nThe four data roles are intentionally different. `fit`\n\nestablishes the old service;\n`observe`\n\ntrains the candidate; `evaluate`\n\nuses held-out evidence; and `guard_x`\n\nprotects old behavior. Reusing the same samples across these roles weakens the meaning\nof promotion.\n\nSee [examples/quickstart.py](/Byte-Naut/vrse-pytorch/blob/main/examples/quickstart.py) for a deterministic synthetic\nwalkthrough and [examples/real_stream_cmapss.py](/Byte-Naut/vrse-pytorch/blob/main/examples/real_stream_cmapss.py) for\nthe frozen real-task lifecycle.\n\nExpected output from the deterministic walkthrough:\n\n```\nVRSE quickstart\n  isolated learning max output change : 0.000e+00\n  promotion passed                    : True\n  new-region route fraction           : 1.000\n  new-region RMSE before -> after      : 2.500 -> 0.003\n  old-region max fallback difference  : 0.000e+00\n  unknown max fallback difference     : 0.000e+00\n```\n\nWhen the documented lifecycle is followed:\n\n`observe()`\n\ncannot change the currently served snapshot.- A proposal binds the baseline, configuration, candidate and authorized region that were evaluated; stale or reused proposals are rejected.\n- Promotion is atomic: the tested GP posterior and region are frozen together.\n- Outside an authorized region,\n`model(x)`\n\nreturns the frozen baseline value exactly, rather than approximately preserving it on average. - One-step rollback restores the previous deployment snapshot.\n\nThese are software and routing properties. They are not a statistical certificate that the baseline is safe, that every unknown input is detected, or that an authorized expert generalizes beyond its demonstrated support.\n\nThe first representative experiment uses **NASA C-MAPSS FD002**, an industrial\nsimulation benchmark for turbofan remaining-useful-life prediction. Each observation\ncontains three operating settings and 21 sensor values. The benchmark is simulated,\nnot field data from operating aircraft.\n\nThe frozen study assigned entire engines to disjoint roles, used five seeds and paired two streams:\n\n**Stable condition:** the new operating regime remains learnable after observation.**Reversed condition:** the inputs are identical, but validation labels are inverted to create a candidate that should not be promoted.\n\nThe protocol, implementation and result artifacts are frozen under\n`phase3b-discovery-global-normalization-v1`\n\n.\n\n| Result across five seeds | Frozen baseline | VRSE |\n|---|---|---|\n| Stable new-regime RMSE, mean | 96.18 | 21.61 |\n| Stable new-regime RMSE reduction | — | 77.5% |\n| Stable candidates promoted | — | 5/5 |\n| Reversed candidates falsely promoted | — | 0/5 |\n| New-regime expert routing | — | 93.0–96.0% |\n| ID / adjacent-unknown expert routing | — | 0.0% / 0.0% |\n\nThe stable result is not produced by rejecting everything: most supported new-regime samples use the promoted expert, reducing error by roughly 4.5×. At the same time, the protected ID and adjacent unknown regime remain on the exact baseline.\n\nThe left panel shows that VRSE recovers most of the utility of globally serving the shadow expert. The right panel shows why isolation matters: under concept reversal, the ungated online learner serves a harmful update, while VRSE rejects the candidate and returns to the baseline. The reversed-label exam changes both the candidate error and the baseline denominator, so it should be read as a controlled rejection test—not as evidence that candidate degradation alone caused rejection.\n\nIn the frozen feature projection, the promoted region covers the demonstrated new condition while leaving the ID and adjacent unknown clusters on fallback. The safety–plasticity plot shows the trade-off directly: global adaptation is useful but interferes everywhere; static rejection avoids interference but never adapts; VRSE occupies the useful, low-interference corner in this experiment.\n\nFull evidence: [frozen snapshot](/Byte-Naut/vrse-pytorch/blob/main/results/PHASE3B_SNAPSHOT.md),\n[mechanical verdict](/Byte-Naut/vrse-pytorch/blob/main/results/PHASE3_RESULT.md),\n[per-seed metrics](/Byte-Naut/vrse-pytorch/blob/main/results/phase3_metrics_table.md), and\n[protocol](/Byte-Naut/vrse-pytorch/blob/main/docs/Phase3_Plan.md).\n\nC-MAPSS is not redistributed by this repository. Obtain it from the\n[NASA dataset page](https://data.nasa.gov/dataset/cmapss-jet-engine-simulated-data),\nextract it to `data/CMAPSSData/`\n\n, then run:\n\n```\npython -m pip install -e \".[benchmark,test]\"\npython -m experiments.phase3_prepare_data \\\n  --source data/CMAPSSData \\\n  --source-origin user-attested-original-extraction\npython -m experiments.phase3_preconditions\npython -m experiments.phase3_matrix\npython -m experiments.phase3_verdict\npython -m experiments.phase3_visualize\n```\n\nDo not continue past preconditions unless the verdict is `READY_FOR_MATRIX`\n\n. Detailed\nchecks, Windows commands and provenance notes are in\n[docs/Phase3_Runbook.md](/Byte-Naut/vrse-pytorch/blob/main/docs/Phase3_Runbook.md). The exact frozen files are identified\nby [results/phase3b_snapshot.sha256](/Byte-Naut/vrse-pytorch/blob/main/results/phase3b_snapshot.sha256). The public\nrelease includes the frozen matrix and checkpoints named by that manifest; the\nprepared C-MAPSS array is excluded and must be regenerated from the source dataset.\n\n```\nVRSEModel.wrap(...)    create a private frozen baseline and empty shadow service\nmodel.fit(...)         establish features, fit-time GP and ID uncertainty threshold\nmodel.observe(...)     update only the isolated shadow posterior\nmodel.evaluate(...)    create an auditable, single-use promotion proposal\nmodel.promote(...)     atomically authorize the tested regional snapshot\nmodel.revoke()         restore one previous deployment snapshot\nmodel(x)               route to the expert only where authorization holds\nmodel.route_mask(x)    expose routing decisions for audit\n```\n\nVersion `0.1.0a1`\n\nsupports supervised scalar regression, one active regional residual\nexpert, CPU execution, SNGP-style distance-aware features and either a 1-D observed\nspan or a high-dimensional KNN feature region. It is a research prototype, not a\ndrop-in safety layer for arbitrary models.\n\n- It does not solve continual learning or catastrophic forgetting in general.\n- It does not guarantee that SNGP uncertainty detects every distribution shift.\n- It does not certify closed-loop, clinical, financial or other high-stakes safety.\n- It has not yet validated classification, delayed labels, adversarial streams, multiple simultaneous experts or repeated region composition.\n- The Phase-3 result is a mechanism study, not a C-MAPSS leaderboard claim.\n\nThe negative and invalid intermediate experiments are retained because they explain\nwhy the final design looks the way it does. The concise\n[research scope](/Byte-Naut/vrse-pytorch/blob/main/docs/RESEARCH_SCOPE.md) separates the frozen evidence from open\nquestions; the public release keeps the evidence needed to audit the final claims.\n\nThe core lifecycle and the Phase-3B case study are frozen. Version `0.1.0a1`\n\nis the\nfirst public research software release corresponding to that result. It is a research\npreview: the API is not promised stable, and the package is not production ready.\nThe immutable version archive is available from\n[Zenodo](https://doi.org/10.5281/zenodo.21653095). See\n[CHANGELOG.md](/Byte-Naut/vrse-pytorch/blob/main/CHANGELOG.md), [CITATION.cff](/Byte-Naut/vrse-pytorch/blob/main/CITATION.cff) and the\n[release checklist](/Byte-Naut/vrse-pytorch/blob/main/docs/RELEASE_CHECKLIST.md).\n\nPublished by **Byte-Naut**, an independent open-source research identity.\n\nByte-Naut. (2026).\n\nVRSE-PyTorch v0.1.0a1: Validated Regional Support Expansion for Safe Online Adaptation[Computer software]. Zenodo.[https://doi.org/10.5281/zenodo.21653095]\n\nVRSE is released under the [Apache License 2.0](/Byte-Naut/vrse-pytorch/blob/main/LICENSE). The project name and\npackage identify this research implementation; they do not imply a safety\ncertification or endorsement by NASA or any other organization.", "url": "https://wpnews.pro/news/vrse-validated-regional-support-expansion-for-pytorch", "canonical_source": "https://github.com/Byte-Naut/vrse-pytorch", "published_at": "2026-07-28 21:02:17+00:00", "updated_at": "2026-07-28 21:22:55.574141+00:00", "lang": "en", "topics": ["machine-learning", "ai-research", "developer-tools"], "entities": ["VRSE", "PyTorch", "Byte-Naut"], "alternates": {"html": "https://wpnews.pro/news/vrse-validated-regional-support-expansion-for-pytorch", "markdown": "https://wpnews.pro/news/vrse-validated-regional-support-expansion-for-pytorch.md", "text": "https://wpnews.pro/news/vrse-validated-regional-support-expansion-for-pytorch.txt", "jsonld": "https://wpnews.pro/news/vrse-validated-regional-support-expansion-for-pytorch.jsonld"}}