DeCiphr is designed to strictly apply the analytical framework you give it. This is a step-by-step guide for developing a coding scheme the AI can apply consistently.
A framework is a self-contained coding scheme: a set of codes with precise definitions, an optional controlled vocabulary of values for each code, and a master prompt that situates the analysis within your analytical tradition. It is stored as a .sigframe file — plain, human-readable JSON — so it can be reviewed, edited, shared with collaborators, and reported in your research output.
DeCiphr runs in strict coding mode only. The AI assigns values from the codes you define; it does not introduce categories, terminology, or observations beyond your framework. Where the visual evidence does not warrant a confident assignment, it returns Uncodable rather than guessing. You review and can edit every result.
Every framework can be organized into two layers that do different work. The AI layer holds deductive codes assignable from what is visually present in the image: languages, layout, materiality, composition. The researcher layer holds interpretive questions that require fieldwork presence, community knowledge, and ethnographic judgment. The AI never answers researcher-layer questions, by design.
Name, a short description of what the framework analyzes, and the field or discipline it belongs to.
The framing instruction the AI receives before your codes: the analytical role it takes, how to treat text and image, and the required output discipline (values only, Uncodable when evidence is insufficient).
The heart of the framework. Each code has a name, a definition written as a coding instruction, and optional sub-categories that act as a controlled vocabulary of permissible values.
Each code belongs to the AI layer Researcher layer or both. Layer assignment is set per code in the web app editor; codes created without a layer setting are treated as AI-layer codes.
The published sources your framework operationalizes, in your citation style, plus creator details (name, institution) so shared frameworks remain attributable.
Frameworks are private by default. You can share them as .sigframe files directly, or publish them to the in-app Framework Library for other researchers to browse and import.
In the web app (app.deciphr.tech) or the mobile app, go to Frameworks and create a new framework. The web app editor includes the per-code layer picker, so it is the better place to build two-layer frameworks.
Give the framework a name, a one- or two-sentence description of what it analyzes, and its field or discipline. Name deliberately — include a version (for example, "v1") so revisions stay distinguishable.
State the analytical role, the tradition the codes come from, whether the AI should attend to text, image, or both, and the output rules: assign only defined values, no prose, Uncodable when the image provides insufficient information.
One code per analytical question. Each needs a name, a definition written as an instruction, and — wherever the theory provides them — sub-categories as the controlled set of permissible values. On mobile you can also import codes from a CSV file.
In the web app editor, mark each code AI, Researcher, or Both. Ask one question of every code: can this be answered from the image surface alone? If it requires being there, knowing the community, or interpreting significance, it belongs to the researcher layer.
Run the framework on a small batch of five to ten items before committing to a corpus. Where the AI's assignments diverge from your intent, refine your codes and definitions. When stable, export the .sigframe or publish to the Framework Library.
The definition is the instruction. The AI sees only what you write; it is instructed by your framework at the moment of analysis, not trained on it, and it has no access to the source text behind your codes. Definitions therefore need to be self-contained: state what to look for, in what order, and what form the answer takes. Write in the imperative ("Identify every language present…", "Assign one of the following…") and give decision rules for hard cases rather than leaving them to judgment.
Constrain values wherever possible. Sub-categories turn a code from an open description into a controlled assignment, which is what makes coding consistent across hundreds of items and comparable across coders. Reserve open-ended codes for the researcher layer, where your own interpretive writing belongs.
Draw the layer boundary clearly. The AI layer is for structural, surface-readable features; it returns coded values only, not interpretive prose. Anything that depends on fieldwork presence, community knowledge, policy context, or reflexive judgment belongs to the researcher layer. The division is what preserves your interpretive authority.
Plan for Uncodable. A well-built framework tells the AI what to do when evidence is insufficient. A high Uncodable rate in your pilot is diagnostic: either the definition demands information the images cannot supply, or the code belongs in the researcher layer.
The example below is taken from the Geosemiotics framework (Scollon & Wong Scollon, 2003) used in the developer's own fieldwork. It shows how a single construct, code preference, is split into an AI-layer structural code and a set of researcher-layer interpretive questions.
Note what the prompt does: it names the analytical role, fixes the scope to what is visible, forbids interpretation, and defines the insufficient-evidence behavior.
Everything here is answerable from the image surface: which codes appear and how they are spatially arranged. The definition carries the framework's own terms and its citation, so the coded output remains traceable to the source.
None of these can be read off the image. Whether a dominant language on a shopfront indexes the surrounding community or symbolizes something else entirely is knowable only through presence at the site — which is precisely why these questions are reserved for the researcher.
Questions about framework design, or a framework you would like to share? Get in touch.