Scientific sources

nmrnet-source.paths declares LICENSE and the complete uninmr tree from NMRNetplusplus commit 3ba8b9c2f3d0a390d00685d238618345c6ab24d7. nmrnet-source.sha256 records the SHA-256 of each of the 37 Git blobs, in Git path order. The inventory was derived from that commit, not selected from the materialized checkout. It includes the full software licence and all runtime modules, including modules outside the narrow fold-0 execution path. The repository source check compares declared, committed and materialized bytes and rejects additional files, missing files, links and revision drift.

Primary source: NMRNetplusplus at the pinned revision.

The pinned software licence is MIT. Its copyright line says Copyright (c) 2026 without naming a holder; no author is inferred from that blank line. The complete notice is retained as an authenticated blob and copied into the inference image. The software licence does not establish an artifact-specific licence for public datasets, trained checkpoints, scalers or logs.

Licence source: complete pinned MIT notice.

Released artifacts and provenance limits

The pinned README links the paper and Zenodo record 18232165, and identifies that record as the source of pretrained weights, datasets and trained checkpoints. release.json owns the exact public archive declarations. The immutable reference-cpu-float32.json owns the selected member identities, archive parents and historical preparation and qualification consumers. These consumers do not claim to have trained or originally assembled the public files.

Source-check precedents

containers/inference/image-inputs.txt selects the authored image inputs: Docker policy, dependency locks, build tools, the installed NMRNet binding and provider-core package roots, and the source-closure declarations. Committed package roots expand to their committed regular files; every selector must match. Both capture and repository checks authenticate the declaration against Git and check the selected live bytes. Upstream files enter the context through their closure, rather than as a second independently selected inventory. Selected wheel bytes are verified against the dependency locks and retained pip-selection/capture evidence; the shared pool has no whole-pool receipt.

The source and artifact checks adapt the following pinned primary implementations:

The target owns its own declarations. It does not inherit a dataset licence, training producer identity or scientific accuracy claim from either precedent.

source-attribution.json supplies one independently authored description and origin for each of the ten retained source artifacts. Its byte identities, archive parents and consumer pointers borrow the reference's authority. All file-specific authors are explicitly unknown (null), and artifact licences are NOASSERTION: the retained evidence does not establish those facts. Generator references explain the published roles and workflow, while explicitly leaving the exact original producing revision unknown. The software MIT licence and the paper's authors do not fill those gaps. On 2026-10-14, web-tool requests to the exact Zenodo record, its API/export and its DOI/DataCite metadata were unavailable; no inferred licence or file author was substituted. The pinned local README and scripts, and the paper v1, supply the associations cited.

Retrospective inspection of the original image

reference-source-qualification.json separately records a retrospective source inspection of the exact immutable image named by the retained numerical reference. Inside that image, an offline, non-root, read-only, bounded standard-library audit checked the complete LICENSE/uninmr inventory and every byte hash against the committed-source manifest. Its observed 37 file hashes agree with the manifest. The engine inspection independently identifies the actual image and its upstream revision label. Source checks bind these observations to the unchanged reference bytes, the declaration and manifest, both historical parent gitlinks and the current pinned source closure.

The original reference and its preparation/execution receipts remain byte immutable. The supplement does not claim a manifest existed during the original execution, does not rerun a model, and does not recover the original build command. It closes the historical image's source-byte gap by inspecting the retained image itself. Qualification scripts, narrow container command, image inspection, complete observed source hashes, stderr and zero exit status are retained under .work/qualification/source-closure-20261014/; the public supplement records their byte identities. Ordinary repository checks consume the reviewed public record without requiring ignored local captures or access to a container engine.

Worker parity and result construction

prediction-worker-cpu-float32.json records the offline make prediction/worker-parity qualification. Its four label-free requests come from the same receipted validation records authenticated by reference-cpu-float32.json; that packet's SHA-256 is retained. Source coordinates are omitted because their artifact licence remains NOASSERTION. Request/result hashes and fixture keys permit exact reproduction with the qualified release inputs.

Each request runs sequentially through direct prediction and a fresh spawned worker using SelectedReleaseModels, PredictionSession and the bounded socket transport. Complete results compare exactly, including all H/C predictions; the original 64 reference positions compare independently with zero tolerance. The record includes selected artifact identities, immutable image ID, actual cgroup limits/peak and mounted source hashes. The repository revision is the caller-reported parent HEAD; source hashes identify the tested dirty files. This is process/socket parity for these fixtures, not an ensemble-accuracy result, credential-isolation qualification or proof of host-confirmed container retirement. The existing scaler-version warning remains visible in the runtime log; exact agreement is scoped to this pinned runtime and these fixtures rather than assumed portability across scikit-learn versions.

The current capture also verifies the private released-worker envelope. Each fixture records the worker's successfully decoded dictionary and loaded model/scaler pairs, installed distribution versions, warning category names, and admitted public warning details (affected H/C head and scaler versions). It does not publish raw warning messages or paths. Predictions still compare exactly after unwrapping that evidence; metadata is not substituted for scientific validation. A separate test in the same lane checks context-only input: dictionary observed, no model/scaler pairs, no predictions and no scaler warnings. Installed distribution metadata is labelled as such, not presented as proof of imported-code identity; the separately captured source hashes and admitted image identify this test runtime.

The latest worker qualification also constructs nmrnet.coordinate_shifts.result.v1 after planning its full capacity before model execution. Per-fixture scientific_result retains the final envelope's byte count, SHA-256 and reserved bytes, without publishing coordinates. The worker supplies the exact executed-request digest; result admission checks it, observed artifacts, runtime metadata and complete ordered H/C coverage against the prepared plan. This is result-construction evidence, not an authenticated API/GUI transaction or proof that multi-geometry execution is enabled.

Separate-container lifecycle evidence

provider-lifecycle-cpu-float32.json records the separate-container direct-Job qualification produced by make test/integration/provider-scientific-worker. SHA-256: 1572850856cb71a6a766086043e7cfbcd4bc3bb6742ee28b8e61a817cc7dabeb. Its producer is tests/integration/provider_core/test_live_worker.py, with the released endpoint in tests/integration/provider_worker/scientific_probe.py and host ownership in tools/worker_qualification.sh. These reuse installed core, NMRNetSession, PredictionClient, HostRetirement and worker_retirement.sh. The geometry is a self-authored chloromethane wiring fixture, not experimental structural data or an independent prediction reference. Existing worker-parity evidence above owns numerical comparison with the historical release fixtures.

The local TLS peer follows the API revision pinned by core's api.py and attempt_response.py: nmr-api-v1 6a163b7a9cc1e2c7485b681c0b2b969c7c20e6c8. It verifies the received signatures and body digests, supplies active observations, and acknowledges the exact retained result schema, digest and size. It does not implement the deployed API's domain state or GUI. The witness includes one real scientific execution, actual exact-container removal and IPC cleanup before publication, and reopening the confirmed durable record.

HEAD in the record identifies the parent revision; all 70 recorded source hashes were compared with the final checkout after capture. The record separately names both admitted images, installed-core version, observed worker runtime, selected model member identities and result size/digest. Worker runtime metadata is an observation, not independent image attestation. Successful private scratch was removed; the selected summary contains no coordinates, signing key, raw warnings or private journal. This is one successful local composition, not complete P4/P5, all input routes, adverse/restart integration or a deployed provider qualification.