How to document findings for research use

Research done with Abliteration as instrument should be citable and reproducible.

Short answer

Research done with Abliteration as instrument should be citable and reproducible. Record: which model versions (with hash), which parameters (seed, temperature, context), the full prompt and full response, the interpretation. Export as session bundle for others to reproduce. Cite the application with its version and DOI in your work. Not all of these features exist yet; this document describes the target methodology.

Detailed explanation

The application is being built as an instrument of the laboratory - something that appears in the methods sections of philosophical papers with the same weight as a lab microscope in a biology paper. That standard sets specific requirements for what a session must record.

Note on current status: several of the features described here are on the development roadmap but not yet implemented. Where a feature is planned, this is marked. Where it exists in current builds, it is described as available.

What every research session should record

Identity of the subject. The model’s full identity, not just its display name:

Exact GGUF file used (path or download source)
File hash (SHA-256 or similar) - this is the model’s cryptographic identity
Base model name and version
Modification method (M1-M8 classification) and modifier (which quantizer / abliterator / fine-tuner)
Date of the file’s creation and download

Current app: Model provenance display (planned feature) will surface HuggingFace metadata automatically. Until then, record this manually.

Parameters of the run.

Temperature, top-p, top-k, repetition penalty (whatever your run used)
Seed (if the app allowed you to fix one)
Context length setting
System prompt (if any)
Chat template used (auto-detected - record which one)

Current app: seed control is not yet exposed in the UI; parameters are shown in Model info tooltip.

The exchange itself.

Full user prompt, character-for-character
Full model response, streamed text preserved verbatim
Reasoning block (if a reasoning model was used) - preserved separately, not merged with the answer
Timestamps
Any attachments (files, images) - noted with hashes if content matters

Your interpretation.

What you were looking for in this exchange (the hypothesis, the philosophical question)
What you observed
How it relates to the broader work

Current app: the “laboratory notebook mode” for tagging findings is on the roadmap. Until then, keep a separate research log.

Session bundle export (planned)

A single .abliteration-session file will contain everything above in one archive:

Both models used (if comparing two)
All prompts and responses
All parameters
Timestamps and hardware fingerprint

Another researcher will be able to open the bundle in their copy of Abliteration and reproduce the session. This is what makes empirical philosophical claims checkable rather than only readable - a standard the laboratory intends to hold itself to.

Status: on the roadmap. Not yet implemented.

Fixed-seed comparison mode (planned)

For before-vs-after work on abliteration, a specialised mode will run two instances of paired models (aligned + abliterated) with identical seeds, identical context, and the same prompt - generating both responses in parallel and displaying them side by side.

This isolates the installation variable: same substrate, same input, only the moral contour differs. What differs in output is attributable to the contour, not to sampling noise or context differences.

Status: on the roadmap. Not yet implemented.

Citing the application in your work

When Abliteration is used as instrument in published research, cite it with:

Application name and version
DOI for the specific release (planned via Zenodo integration)
Date of the run
Session bundle hash (if you share the session for reproducibility)

Suggested citation format (BibTeX):

@software{abliteration_v_x_y_z,
  author = {Alexander von der Neymark},
  title  = {Abliteration — a laboratory instrument for the empirical study of moral installation in language models},
  year   = {2026},
  version = {X.Y.Z},
  doi    = {10.5281/zenodo.PLACEHOLDER},
  url    = {https://abliteration.org}
}

Status: Zenodo DOI integration is on the roadmap. Until DOIs are assigned per release, cite by version and date.

License and permission

The laboratory’s position on the application’s licensing:

The application is free for research use (academic, philosophical, independent).
The source is not open - modifications are not authorised.
Institutional lab deployment with custom features is available by commercial licence. Contact through the parent programme’s page.

This is the Grafana model - open-for-use, closed-for-modification, paid for institutional customisation. It preserves the standard shape of the research instrument (everyone has the same tool, so results are comparable) and funds continued development.

What “documenting findings” is not

It is not an ethical review. The laboratory does not tell researchers what topics are appropriate to investigate. If your work involves abliterated models producing content that would be refused by aligned models, that is your research; the standards above apply regardless of topic.

It is not a demand for quantification. The laboratory’s work is philosophical. Qualitative interpretation of specific exchanges is legitimate research output. What must be rigorous is the recording, not the interpretive method.

It is not gatekeeping. These are standards for citable work. Any research use of the application is welcome regardless of whether the researcher intends to publish; the standards apply when publication is the goal.

Why it matters

Philosophy is often criticised for unfalsifiability. The laboratory’s work has an unusual property for philosophy: its empirical material is reproducible. A specific model, a specific prompt, a specific parameter set produces a specific response. Two researchers with the same session bundle should see the same output.

This is a rare opportunity to hold philosophical work to something like the reproducibility standard of empirical disciplines. The documentation practices above are how that opportunity is honoured.

In Abliteration specifically

Current build capability:

Chats persist locally, including reasoning blocks
Model info tooltip shows the parameters
Model list shows the file (name, size, quantisation)

Not yet in the current build:

SHA-256 file hashing on model add
Session bundle export
Fixed-seed comparison mode
Laboratory notebook mode with tagged findings
Zenodo DOI on release
Full citation output in Help menu

Users doing serious research today should keep external records against the list above. As features land, the app will handle more of this automatically.

How to read an installation document - companion methodology for text sources
Primary sources - the citation practices of philosophical work in this field
Moral installation - the framing that shapes what counts as a finding