Consent, context, and lineage should stay attached to research materials.

In many research workflows, consent forms, transcripts, excerpts, translations, coded segments, and published quotations are stored and circulated separately. As materials move across tools, people, and publications, the conditions that governed their use can become harder to see. KNOBE approaches this as a file-level problem: a research object can pair its readable content with a sealed declaration of consent terms, source context, transformation history, and a verification hash, in one plain-text file.

• Researchers
• IRB reviewers
• Archivists
• Research staff

The lineage this profile documents: consent terms are captured in the sealed original; every excerpt or translation chains to it by hash; a downstream reader verifies the chain and reads the terms before using the material.

1. Originalconsent captured at seal
   →
2. Excerpt or translationchains to it by hash
   →
3. Downstream readerverifies, then reads the terms

The conditions that governed collection stay attached to every derivative.

What breaks in research today

The failure on this site's front page is a research scenario: an oral history whose consent terms, do-not-quote request, and contextual notes fall away as an excerpt moves through an AI abstract into a newsletter and a grant report. Every step was reasonable. By the fourth, the words still circulate and the conditions that governed them are gone.

Three familiar drifts sit underneath that decay. Consent detaches from data: terms agreed at collection are stored apart from the material and become hard to consult downstream. Derivation goes quiet: by the time a quotation reaches print, the path back through coding, translation, and summarization is reconstructable only from memory. AI-assisted steps compound both: summaries and abstracts move faster and farther than the conditions that governed their sources.

The technical question is whether consent terms and lineage can remain attached to the material itself, rather than being reconstructed later from surrounding documentation. KNOBE is one answer to that question.

What a sealed research object declares

A sealed research object is one plain-text file: the transcript, excerpt, or translation as its readable body, and a sealed payload that declares:

• Consent terms. The version agreed to, and the uses permitted and prohibited.
• Source context. Who contributed, who collected, under what instrument.
• Transformation history. What kind of derivation this object is, declared rather than inferred.
• Lineage. A parents field chaining to the hash of the object it came from.
• A verification hash. Change one character and the check fails.

What do "consent terms inside the payload" look like? These are the actual fields from the sealed example below:

The basic workflow

At minimum, sealing is a file-integrity step: it establishes that a transcript has not changed since collection. Everything beyond that is optional and builds on it.

This does not require a new platform, and it does not change what you collect or how. Materials are sealed as they are produced, and each later object chains back to what it came from. The harder questions are procedural rather than technical: when to seal materials, how to describe consent terms, and how a research team or an IRB should treat derivative objects.

1. A transcript is sealed as a .knobe.md: the words as the readable body, the consent terms inside the sealed payload.
2. A coded excerpt is sealed as a derivative. Its parents field chains to the transcript's hash, and its content_type declares the transformation.
3. A translation is sealed the same way and declares its own fidelity limits.
4. A quotation in a publication can be traced back through the chain to the consented source.
5. Anyone holding the files, including an IRB, can verify the chain locally, with no registry or account.

The range of use

Low friction

• Excerpts keep source context. A quotation leaves the transcript with its consent terms, collection context, and de-identification status attached.
• Translation declares fidelity. A translated instrument or interview chains to its original and declares its equivalence limits.
• Coding becomes traceable. Coded segments chain to the transcript and the codebook version that shaped them.
• AI assistance is declared. Transcription, translation, coding help, drafting: stated per object rather than per project, and declared-synthetic material stays separable from human corpora.

Research infrastructure

• Codebooks as versioned objects, so revisions stop being unreconstructable.
• Analytic memos that chain to the excerpts and notes they interpret.
• Instruments sealed per version, prompts included; every response chains to the exact version used.
• Replication packages in which tables and figures chain to the data and code that produced them.
• Meta-analysis extraction trails, each estimate chained to the passage it came from.
• IRB audit packets: one sealed chain from consent version to published quotation, instead of scattered files.

Originals held by participants or communities

• Oral history whose do-not-quote and attribution terms travel with every excerpt.
• Community-based research in which the community holds the corpus and licenses researcher access.
• Data donation as purpose-limited license objects rather than surrendered raw exports.
• Long-term AI-interaction panels, the study described below.

The last group pairs with the custody pilot, next. Nothing in the first two groups requires it.

An advanced pilot: the participant keeps the original

Everything above leaves custody where it already is. For some studies, the same mechanics could support a stronger model: the participant retains the sealed original, and researchers work from verified derivatives whose lineage points back to it.

The participant's copy is plain text. It lives in their email or their drive, opens in any editor, and verifies locally years later, with no account or platform required. Because the consent terms travel inside it, the participant keeps proof of what they agreed to rather than a photocopy of a form. Years later, they could grant a new use by issuing a new license object chained to the original, or a revocation object that cooperating tools would cite when declining further use.

This model is not required for any of the workflow above. It is a promising fit for oral history, community-based research, data donation, and long-term panels, and an IRB-governed pilot would test it first.

Limits

Can a participant withdraw?

In the custody model above, withdrawal could work like this: the participant deletes their original and issues a revocation object chained to it, and cooperating tools would decline further use and cite it. No protocol can claw back copies already made elsewhere. This would make withdrawal visible and citable, not enforced.

Does the seal prove when something was created?

No. created_date is a declared field; the seal proves content, not time. Preregistration and priority claims need trusted timestamping, which is application-layer work (for example, ledger anchoring in a credentialed environment) and never a requirement of the open format.

What about re-identification?

A sealed pseudonymous object is still a disclosure-risk object; sealing does not anonymize. De-identification remains the researcher's methodological responsibility. What KNOBE adds is that the de-identified derivative declares its transformation and chains to a consented source an IRB can audit.

At larger scale: a study we have not run

The workflow works at any scale. A fifteen-participant qualitative study is a complete use on its own. What follows is the far end of the range.

In research on what AI does to cognition and culture, the stimulus and the response both need provenance, yet human-AI interaction is often thinly provenanced: the model version, the prompt, the output, and the human's edits can separate almost immediately. A panel study built on participant-retained originals would look like this: participants donate sealed monthly slices of their AI interactions. Each donation records which model produced the stimulus, what the human did with it, and the participant's terms. Each participant keeps their own longitudinal corpus, so the panel's value accrues to the panelists and not only to the lab.

• Contamination would be inspectable by declaration. Every object states its AI involvement, so corpora could be built AI-labeled or AI-excluded without relying on detection heuristics.
• Replication would inherit materials, not methods paragraphs. Instrument, exact stimuli, and responses, all sealed and byte-verifiable years later.
• Waves could not be quietly revised. Wave-one objects stay sealed against a frozen specification.
• Analysis tools would consult each object's terms before use, honoring aggregate-only conditions per object, per participant, with a citable basis for every refusal.

This study has not been run.

For your IRB

KNOBE does not change what you collect or how you collect it. It changes how data is stored and transmitted: a plain-text file with the consent terms embedded, instead of a transcript in one folder and a consent form in a drawer. Integrity can be checked in a browser with the Lens, with no technical training. Whether sealing counts as a protocol amendment is your board's call. Sealing is not anonymization; de-identification remains a methods responsibility.

How to cite this

A preprint DOI will be added when available.

What is built, and what is proposed

The participant who keeps their interviews is the same person as the student who keeps their learning record and anyone who keeps their AI conversations. A lifetime of sealed objects is a personal corpus, and platforms become interchangeable lenses over it. That argument is larger than research and belongs to the white paper.

Pilot studies

A pilot could take several forms: an existing consented dataset resealed to test the workflow, a new interview study designed around it, an archive accession, or a course-based methods project. It need not adopt participant-held originals; it could begin with a single sealed transcript or a short excerpt chain, evaluating what the format adds, what it complicates, and what an IRB would need to know.