Temple of Two
← Instruments

Study

The Black Box Boundary

A byte-exact reconstruction of how much of a deployed model's input is actually the operator.

StudyPublishedBounded, n=1

Operator words: 0.56–14.00% of model-input bytes

The question
The opacity debate points at the weights. For a system someone actually lives with, how big is the opaque part relative to everything around it?
The instrument
A non-interventional, byte-exact reconstruction of the inference-event boundary around a small deployed language model, with the complete raw record of its first lived session published and SHA-256 manifested.
The finding
In the studied session the operator's words were 0.56–14.00% of model-input bytes. The remainder was enumerable application-layer material — which means most of what shapes the output is inspectable, and was simply never inspected.
What failed
Nothing about the weights. This measures the boundary, not the interior; the weights black box is untouched, and reading the result as an interpretability claim would be a category error.
Standing
StudyPublishedBounded, n=1
Verify it
DOI 10.5281/zenodo.21683054 ↗Code + evidence ↗

Metrics with scope

Every number names what was measured, on what sample, and as of when.

  • Operator share

    0.56–14.00%

    Fraction of model-input bytes across one session

    As of 2026-07-29 · source

  • Sample

    n = 1 session

    Single deployed companion; not generalised

    As of 2026-07-29

From wonder to receipt

The Black Box Boundary

Wonder initiated the search. Rigor changed the answer.

StudyPublishedBounded, n=1
  1. 01Wonder

    Everyone says the model is a black box. Sitting with one every day, how much of what it receives is actually dark?

  2. 02Hypothesis

    If most of a deployed system's input is application-layer material rather than the operator, then the opaque fraction is far smaller than the debate assumes — and measurable.

  3. 03Instrument

    A non-interventional, byte-exact reconstruction of the inference-event boundary, with the complete raw session record published and SHA-256 manifested.

  4. 04Encounter

    The operator's words were 0.56–14.00% of model-input bytes. The rest was enumerable — inspectable all along, and simply never inspected.

  5. 05Standing

    Bounded, n=1. This measures the boundary, not the interior; the weights black box is untouched, and reading it as interpretability would be a category error.

  6. 06Receipt

    DOI 10.5281/zenodo.21683054 · code + evidence 10.5281/zenodo.21683073

DOI 10.5281/zenodo.21683054 ↗Code + evidence ↗