Methodology
Objective
The private project explored how to build a source-grounded assistant across three distinct layers:
- supervised examples that teach answer behavior;
- retrieval records that preserve source identity and claim boundaries;
- sealed evaluations that measure retrieval, reasoning, and unsupported-claim behavior.
The public toolkit documents this architecture without distributing the private source corpus.
Provenance model
Every private source was assigned a namespace, a stable source identifier, a source type, and a confidence or claim-status field. The design deliberately separates:
- direct source evidence;
- implementation metadata;
- secondary summaries;
- theories and fan interpretations;
- independently authored material.
An implementation artifact can support a claim about observed behavior without automatically supporting a claim about narrative intent.
Split construction
The final SFT release used train, validation, and test splits. Release checks rejected:
- exact prompt or answer-pair overlap across splits;
- normalized overlap after case and whitespace normalization;
- semantic families crossing split boundaries;
- leakage from training prompts into sealed evaluation questions.
The public synthetic records use the same broad chat structure but are newly authored and unrelated to the private source text.
Evaluation design
The sealed private benchmark covered source-grounded answering, cross-scope reasoning, exact retrieval, chronology, false-premise correction, implementation interpretation, visual identification, citation precision, and missing-source behavior.
Each evaluation record separated required claims, forbidden claims, source references, difficulty, and scoring method. Evaluation files were excluded from training.
Release validation
The private blocking validator checked structural integrity, source resolution, image validity, manifest consistency, hashes, dataset counts, split isolation, and local-path redaction. Aggregate results are preserved in this public repository; protected inputs and generated content are not.