Independent data work
NostalgAI turns historical public discussion archives into carefully documented, machine-readable datasets. The work keeps the context needed to examine what was said, when it was said, how replies connect, and where uncertainty or disagreement remains. Provenance stays attached.
Questions answered
Documented research datasets reconstructed from historical public discussions, with schemas, provenance pointers, quality evidence and explicit uncertainty. Current work covers beekeeping and neural-network archives; exact contents depend on dataset and edition.
Use NostalgAI when research needs chronology, contribution type, concepts, references, reply relationships or auditable provenance. A raw archive or ordinary search is better when reconstructed context and record-level analytical fields are unnecessary.
NostalgAI is an independent data project operated by Sarah van Oorsouw. There is no self-service purchase or public price. Sarah van Oorsouw reviews the research question, rights and privacy constraints, and suitable scope case by case.
Current work
This dataset structures discussion from the archived Usenet group
sci.agriculture.beekeeping. It connects each reconstructed
thread to the question or issue that started it and links each extracted
knowledge item back to its source.
The package comes in JSONL, CSV, and SQLite. It includes the schema, provenance pointers, quality evidence, and explicit records of where the extraction abstained.
Current work
This dataset structures technical discussion from the archived Usenet
group comp.ai.neural-nets. It records bounded contributions
by date, detected concept, technical reference, contribution type, and
how the writer frames what is known or uncertain.
Two editions support different kinds of analysis. The analytical edition works at contribution level. The pseudonymized provenance edition adds reconstructed threads, reply relationships, chronology, confidence, and stable contributor pseudonyms.
Two commercial editions
Primary research product
Use this edition when the research depends on context. It links contributions to reconstructed threads, chronology, stable contributor pseudonyms, pseudonymous message pointers, reply relationships, and record-level reconstruction confidence.
Minimized analytical view
Use this edition for contribution-level analysis. It contains the same analytical records, without customer-visible person-level or thread-level provenance.
What this enables
Frozen Holdout 03
The method was frozen before evaluation on a fresh 100-thread, 730-message holdout. Manual adjudication reviewed a stratified frame of 260 records. These figures measure product performance; they are not inter-annotator agreement statistics.
Reconstruction limits
Of the 34,029 reconstructed threads, 33,838 are marked reliable, 180 partial, and 11 unresolved. Reconstruction confidence and semantic confidence remain separate. Missing replies stay missing: silence is never counted as agreement or resolution.
The archive records what participants wrote at the time. Their statements may be incorrect, outdated, disputed, incomplete, or later corrected. Sarah van Oorsouw reviews technical questions and sample requests case by case.
Ask Sarah van Oorsouw about a research use case: hello@nostalgai.us
Working principles
Records keep the source pointers and pseudonymous attribution needed for audit and review.
The data keeps competing statements, uncertainty, reconstruction limits, and explicit extraction abstentions visible.
These records show historical discussion. They are not current professional guidance or verified ground truth.