Neural Data Preservation Framework (NDPF), Version 0.9 — Preliminary

1Status of this document

Pending issue of Version 1.0, Neural Continuity Research Program (NCRP) projects are asked to apply the requirements of this draft on a provisional basis and to record any requirement they cannot meet. Those records will inform the final text.

2Purpose and scope

The NCRP produces data that cannot be regenerated. Synaptic-resolution imaging destroys the tissue it records, so a lost raw dataset is lost permanently. Reconstruction products embody years of human proofreading. Model parameter sets may be the only record of how a given result was obtained. The NDPF exists to ensure that these data remain intact, interpretable, attributable and appropriately controlled for periods measured in decades, and in some cases for a century or longer.

The Framework applies to all data created, received or held under the NCRP, at any Departmental facility or at any partner institution operating under an NCRP agreement. It applies to non-human and human-derived data alike, with additional requirements for the latter. It does not govern data held solely for administrative purposes, and it does not authorize any acquisition, experiment or computational model. Authorization of research remains with the program review process and, for any protocol involving human tissue, with the Research Ethics Review Board (RERB).

The NDPF builds on three Departmental documents: the assessment of archival media in DSC-TR-2026-005, the data-volume and pipeline requirements in DSC-TR-2026-009, and the discussion of identity and consent in DSC-WP-2026-006.¹ It does not restate their content.

Conformance language

Requirements are numbered NDPF-R-01 to NDPF-R-42. The word SHALL denotes a mandatory requirement. SHOULD denotes a recommendation from which a project may depart with documented justification approved by the NDPF Technical Committee. MAY denotes a permitted option. A project conforms to the Framework when it meets every SHALL requirement applicable to the data classes it holds.²

3Definitions

Data object

The smallest unit of data managed under the Framework, carrying its own persistent identifier and fixity value. A data object may be a single file or a defined, versioned set of files.

Preservation copy

A complete, verified copy of a data object held for long-term retention under the replication requirements of this Framework. Working copies used for active processing are not preservation copies.

Fixity

The property of a data object being unchanged from a recorded reference state, demonstrated by recomputing a cryptographic hash and comparing it with the stored value.

Offline copy

A preservation copy held on media that are not connected to any network or powered system, and that cannot be altered or deleted by any remote action.

Provenance

The documented history of a data object: its origin, the processes and agents that produced or altered it, and its inputs.

Custody

Responsibility for the safekeeping and controlled use of a data object. Under this Framework the Department holds human-derived data as custodian on behalf of donors and the public, not as owner.

Human-derived data

Any data object produced, in whole or in part, from human tissue or from data about an identifiable human donor.

Specimen

A discrete piece of tissue, with a single origin, from which data are acquired.

Emulation Fidelity Level (EFL)

The NCRP classification of computational models by the level of biological detail represented, from EFL-0 (statistical or behavioral model) to EFL-6 (validated individual-equivalent, hypothetical).

Retention class

An approved category that determines how long a data object is kept, how many preservation copies are required and whether lossy reduction is permitted.

4Data classes

Every data object belongs to one of four classes. The classes differ in volume, in whether they can be regenerated, and in sensitivity. These differences determine replication, format and access requirements.

5Metadata and provenance

A preserved dataset that cannot be interpreted has little more value than a lost one. Connectomic data are especially dependent on context: the meaning of a gray value depends on staining and instrument settings, and the meaning of a segmentation depends on the software and parameters that produced it. Provenance must therefore be captured at the moment each object is created, not reconstructed afterward.

NDPF-R-09 is a terminological requirement with a substantive purpose. The question of whether any computational model could stand in a relation of identity to a person is unresolved, and metadata that presumes an answer would outlast the documents that qualify it.³

6Integrity and fixity

Storage media degrade silently. Studies of long-term digital storage show that undetected corruption, operator error and correlated failures across supposedly independent copies are more common than device specifications suggest. The Framework therefore treats fixity checking as continuous work rather than an occasional audit.

7Format requirements

A preservation format must remain readable after the software that wrote it is gone. The Framework does not mandate specific file formats in Version 0.9. It sets properties that preserved formats must have, and the Technical Committee will publish a list of conforming formats with Version 1.0.

8Replication and geographic distribution

The Framework requires at least three preservation copies, on at least two media types, at no fewer than two facilities, with at least one copy offline. The reference configuration places preservation copies at Continuity Research Facility Bravo, in the Colorado Front Range, and Continuity Research Facility Charlie, in interior Alaska. The two sites differ in geology, climate, utility supply and hazard profile, which reduces the likelihood that a single event affects both. Working data at Facility Alpha, where most NCRP processing takes place, are not preservation copies.⁴

Capacity is the principal constraint on NDPF-R-25 for D1. Bravo Archive Level 1 holds approximately 120 PB, and the Charlie pilot vault 1.5 PB pending completion of the main vault, with a 60 PB long-duration target, in 2028. A whole-mouse-brain acquisition in Phase II would produce on the order of 1 EB of raw data. Three full copies of that volume cannot be held within currently planned capacity. Version 1.0 will therefore need to define retention classes for D1 that balance the irreproducibility of raw data against physical limits. Comment is specifically invited on this question.

9Consent and custody

Human tissue enters the NCRP only post-mortem, through donation programs operating under RERB-approved protocols. The Framework does not regulate those programs; it governs what may be done with the data afterward. Six principles apply.

1. Scope limitation. Data are used only within the scope of the consent under which the tissue was donated.
2. No representation of survival. Neither consent materials nor any Departmental communication may state or imply that preservation of data constitutes or enables the survival or continuation of the donor.
3. Custodianship. The Department holds human-derived data as custodian. Custody confers obligations, not ownership.
4. Separation of preservation from use. Consent to preservation is not consent to every future use, and in particular is not consent to the construction of models of increasing fidelity.
5. Withdrawal. Withdrawal by an authorized representative is honored to the extent technically possible, and the limits of that possibility are disclosed before donation.
6. Non-commercialization. Human-derived data are not sold, licensed for commercial gain or transferred to commercial custody.

NDPF-R-35 resolves one difficulty and creates another. Write-once offline media cannot be selectively erased; destruction of a donor-specific key makes the corresponding data unreadable without altering the media. But the same mechanism means that loss of a key is loss of data, and the keys must therefore themselves meet century-scale preservation requirements. The Technical Committee regards this as the least settled requirement in the draft.

10Access control

11Review and revision

The NDPF is maintained by a Technical Committee chaired by DNC, with members from CIA-DSC, the Long-Duration Information Preservation Program and the Office of the Chief Ethics and Research Integrity Officer. The Committee is responsible for interpretation of the Framework, approval of departures from SHOULD requirements, and maintenance of controlled vocabularies and the list of conforming formats.

Handling of public comments

Comments on this draft are invited from researchers, archivists, ethicists, tissue donation organizations and members of the public, and should be submitted by 15 December 2026 through the process described on the Department’s Contact page. Comments are most useful when they cite the requirement number to which they refer and, where they propose a change, include suggested text.

• Each comment will be logged and assigned to one or more requirements or sections.
• The Technical Committee will adjudicate each comment as accepted, accepted in part, or not accepted, with a stated reason.
• A disposition summary will be published with Version 1.0. Commenters will not be identified by name without their permission, and personal information will not be published.
• Comments received after 15 December 2026 will be held for consideration in the first biennial review.

Version 1.0 is planned for issue during fiscal year 2027. Until then, Version 0.9 applies provisionally as described under the status of this document.