Recon — Digital Preservation for the 20-Year Horizon (LoC + NDSA)

Collaborative Mess

25-Aug-2026

Recon pass on the question of ground truth that outlives its own technology. It derives from a conversation with Georgia Tech’s Tech AI and diverted from it.

The Question:

The design horizon is roughly twenty years. Nothing in today’s stack — vector database, embedding model, file format, the whole lot — will still be running at the end of it. So, the durable artifact can’t be the index; it has to be something from which the memory can be rebuilt with whatever exists in 2046. What does the archival or digital-preservation literature say about representing personal memory for reconstruction rather than retrieval?

The reframe that collapsed the question

The original plan pointed at archaeologists and architectural historians — Mamoli (reconstruction of ancient libraries under evidential constraint), Willkens (heritage documentation, HBIM, reality capture).

Wrong tool, and the reason is structural.

	Works with	Controls the deposit?
Archaeologist / heritage historian	Involuntary residue — what someone left without meaning to	No. The discipline exists because the depositor wasn’t trying.
Pepa	Deliberate, controlled writes, now	Yes. Entirely.

A theory of coping with someone else’s carelessness is not what’s needed. What’s needed is the rubric for not being careless. That reframe collapses most of the Question.

Mamoli isn’t wrong — she’s filed under the wrong question. Reconstructing a life from residue someone else left is what happens at Pepa deployment for a new senior: eighty years of photos, documents, and other people’s recollections that Pepa had no hand in creating. That belongs to the cold start question, and it’s the intake interview — the “geriatric experts” are conducting an excavation. Which is why that framing went in another direction instead.

What survives the collapse

One thing, and it isn’t a format problem:

A JSONL file in 2046 is readable but not necessarily interpretable. Someone opens it and finds {"trust": 0.82, "supersedes": "mem:4471", "source": "explicit_user"}. Readable, yes. Meaningful only if a plain-prose document beside it explains what trust means, what supersession implies, and what explicit_user was distinguishing itself from.

That is the only real work in the Question, and nobody else can do it — not a library, not a professor. It has to be written while the reasons are still remembered.

LoC already has a name for it: “Self-documentation.” It is one of their seven formal evaluation factors, not an afterthought.

Library of Congress — Recommended Formats Statement (RFS)

In its second decade, first launched 2014. Current edition 2025–2026. Comments to rfs@loc.gov.

Honest assessment: as a document it is mostly useless here — nine of eleven sections cover books, microforms, vinyl, motion pictures, video games. The seven sustainability factors are named in the introduction but never defined. And the RFS is explicitly geared toward published content, not personal papers.

Two sections are directly relevant.

VI.i Datasets — confirms JSONL

Platform-independent, character-based formats are preferred over native or binary formats as long as data is complete and retains full detail and precision. Line-oriented, e.g. TSV, CSV, fixed-width. Character encoding, in descending preference: UTF-8, UTF-16 (with BOM).

JSONL is line-oriented and character-based. Instinct confirmed.

The metadata requirements under Datasets are the self-documentation gap made concrete:

data dictionaries, schemas, technical specifications
a key or reference to each data field ← this is the trust: 0.82 problem, named
checksums
permanent version specifiers (date, version number)
information about how the data was collected and any sampling or post-processing applied ← provenance, stated as a deposit requirement

VI.ii Databases — names JSON specifically

Files in formats which support linking or embedding external resources (e.g. XML, JSON, Excel) should be self-contained to remain useful in the event of external service changes. The raw data is available without executing code.

An argument against storing references to rows in another system.

IV.ii Audio — corrects the WAV plan

WAVE file with embedded metadata (Broadcast WAVE) rather than without. File in native resolution rather than up-sampled. Uncompressed rather than compressed.

Action item: BWF, not plain WAV. Broadcast WAVE is a standard chunk inside an ordinary WAV file carrying origination date, time reference, and description. Any WAV reader still opens it; a reader that doesn’t know BWF ignores the chunk. Timestamp and context live in the file rather than in a filename or a separable sidecar. One library’s cost.

The Format Evaluation Matrix template (XLSX) — the actually useful artifact

This is what the PDF withholds: the seven factors with their defining questions. Responses are Yes / Maybe / No.

Factor	The question it asks
Disclosure	Is technical information available through complete and open documentation and specifications?
Adoption	Widely used, especially by peers? Integrated into multiple toolsets, not locked to one vendor? Are community user groups available for advice and support?
Transparency	Can it be analyzed with basic tools? Standard character encoding supported? Is lossy compression or encryption enforced?
Self-documentation	Can a file in this format describe its own content and structure with embedded metadata? (Plus accessibility support.)
External dependencies	Free of dependence on particular hardware, OS, or software for rendering or use?
Impact of patents	Free from patent terms that might impede long-term use?
Technical protection mechanisms	If DRM/encryption is required, can custodians reliably maintain future access?

Preferred = meets or exceeds benchmarks on all relevant factors. Acceptable = minimum acceptability, or misses some.

Two findings from the worked sample

1. LoC’s own sample row is WAV/LPCM, rated Preferred on all seven factors. Links to their format description fdd000002. The exact format already chosen.

2. The transcript can live inside the WAV. From the Self-documentation cell on that row:

Closed captions and transcriptions can be embedded within the Labeled Text chunk (ltxt) and identified with a ‘Purpose’ label — capt for closed-caption text, tran for transcription. The optional Associated Data Chunk provides context for the audio data along the timeline.

For a voice memory archive this is significant: one self-describing file, timeline-aligned, no join to maintain for twenty years. No sidecar JSON that can get separated from its audio.

Adapting the second half

The matrix’s back half is LC-local — staff expertise, software availability, whether they hold 100,000+ files of it, repository tooling, whether it renders on loc.gov. Not directly applicable, but the shape adapts cleanly to Casa Delta: can José maintain it, will it run on hardware he’ll still have, is it already in the workflow.

NDSA — Levels of Digital Preservation v2.1

Released March 2026. The operational complement.

LoC answers “what format do I write.” NDSA answers “what do I have to keep doing, forever.” Fixity checking, multiple copies, storage refresh, access control, metadata upkeep. Format choice is made once; the Levels are the chores that never stop. Functional areas include Storage, Integrity, and Content.

Tiered matrix — you can sit at different levels in different functional areas according to resources. The right shape for one person with a homelab rather than an institution.

Resources available:

Matrix (color + B/W)
Assessment Tool — downloadable spreadsheet template to record and visualize results, with guidance and two worked case studies. The usable artifact.
Implementation Guidelines and Working Definitions
Digital Curation Decision Guide (visual + narrative)
Environmental Sustainability Guide (new in 2.1)
Training slide deck (v2.0)

All hosted on OSF (osf.io/rcs38). Site content is CC BY-SA 4.0 — reusable in the Pepa blueprint with attribution, share-alike on that portion. Worth noting against MIT.

Spanish translation of the Matrix v2.0 exists via APREDIG (Ibero-American association for digital preservation). Not incidental — an existing Spanish-language preservation community and a vocabulary that doesn’t have to be invented.

Open doors found (none pursued)

Door	Cost	Note
NDSA Levels open sessions	Free	Quarterly open sessions, anyone welcome; bi-monthly discussion sessions via NDSA-All listserv. A room of working archivists vs. one cold email to one professor.
DigiPres Conference 2026	CFP open, virtual	A one-person twenty-year personal preservation program is unusual for that audience.
APREDIG	Unknown	Spanish-language preservation community.
`ndsa.digipres@gmail.com`	—	General NDSA contact.
`rfs@loc.gov`	—	RFS comments.

Where this leaves the Question

Largely answerable without outreach. The reframe removed the need for a theory of reconstruction; LoC supplies the format rubric; NDSA supplies the operational program.

Not yet done:

Fill the LoC evaluation matrix for JSONL and BWF (~1 hour; produces a defensible answer in a form that can be handed to anyone)
Run the NDSA Assessment Tool against current Casa Delta practice
Write the data dictionary — the plain-prose “key or reference to each data field.” The one irreducible piece of work.
Verify where JSON/JSONL currently sits in LoC’s format descriptions (their file-format research posts show active work on YAML, so the dataset category is moving)

Not decided here

Recon only. No OBOSBS requested or produced. BWF-over-WAV is the only concrete change proposed, and it is not yet accepted.