Upload narratives
CSV, Excel, open-ended survey responses, interview excerpts, or field notes.
Research infrastructure
From qualitative
narratives to citable
research data.
NarraGrid is a collaborative platform for turning interviews, memories, and open-ended responses into structured datasets - with prompts, benchmarks, and findings you can cite.
used by 14 research labs3,184 cases scored83 external citations
prompttr_mem_valence . v1.2
I remembered the summer kitchen at my grandmother's house, the smell of pickled vine leaves, and how she would hum while folding them.
↓ claude-sonnet-4.5
{
"valence": 2.4,
"confidence": 0.88,
"rationale": "warm sensory recall"
}
The pipeline
Five steps from raw text to a dataset.
The workflow preserves the method around every result: the source text, prompt version, model provider, validation rules, and human benchmark.
Define the framework
Capture codebooks, rubrics, rating scales, and exclusion criteria as method.
Version the prompt
Build a structured instrument with schema validation and change history.
Score with models
Run GPT, Claude, Gemini, or local models against the same research task.
Benchmark with humans
Compare against expert raters and report r, kappa, F1, MAE, and n.
Public artifacts
Prompts, benchmarks, and findings that can be cited.
Treat the research instrument as a first-class object. Publish the prompt, benchmark, or result with provenance attached.
Turkish autobiographical memory valence rating
Kavdir, A. (2026). Turkish autobiographical memory valence prompt, v1.2. NarraGrid.
Rates each memory on a -3 to +3 scale with rationale. Validated against expert ratings.
- r
- .724
- kappa
- .81
- cited
- 47x
Self-defining memory specificity, three class
Narrative Identity Lab. (2026). Specificity benchmark, v1.0. NarraGrid.
Compares four LLMs against expert-coded memories using a shared codebook.
- F1
- .81
- kappa
- .76
- n
- 200
Model comparison on Turkish narrative valence
Kavdir, A. and Celik, M. (2026). Turkish valence model comparison, v1.1. NarraGrid.
A same-prompt comparison of Claude, GPT, and Gemini against human ratings.
- models
- 3
- delta r
- .14
- runs
- 18
For research labs
A workflow built for methods sections.
structure
Turn text into variables.
- Score narrative material on valence, specificity, agency, coherence, trust, or your own rubric.
- Export datasets with variable names, codebook notes, and provenance attached.
benchmark
Compare LLMs like raters.
- Run multiple models on the same task and inspect where their judgments diverge.
- Calibrate against expert ratings and report the measurement behind every claim.
publish
Cite methods directly.
- Publish prompts, benchmarks, datasets, and findings as stable versioned artifacts.
- Let other labs reuse the instrument instead of reverse-engineering your appendix.
Worked example
A prompt can carry evidence, not just instructions.
Every artifact keeps the receipt: which model, which dataset, which evaluation, and which version produced the result.
Disciplines
Where open text becomes evidence.
PsychologySociologyPolitical scienceAnthropologyEducationCommunication studiesPublic healthPolicy researchOrganizational researchHumanitiesSurvey methodologyMarket research