BookProof by BedVibe • Powered by LongBook Verifier
BookProof is a document-grounded verification and retrieval-evaluation tool: it checks whether an AI-generated summary, answer, report, or adaptation is actually supported by your manuscript. It helps authors, editors, audiobook creators, researchers, and RAG developers find missing story details, unsupported claims, and hallucinations before the output is trusted.
BookProof audits AI summaries and answers against the source manuscript, then separates confident language from supported evidence.
Compare AI summaries or answers against the source manuscript, detect missing evidence terms, flag weakly supported or unsupported claims, and generate audit reports for long-book QA.
Separate answer-term coverage from evidence-term coverage so short answers and manuscript grounding can be reviewed independently.
AI summaries can sound confident while missing key plot evidence, chronology, character motivation, object continuity, or chapter-specific details. BookProof treats AI output as something that must be checked against the manuscript, not trusted because it sounds fluent.
BookProof is backed by the LongBook Verifier research package, published on Zenodo. The full benchmark results — 5 retrieval strategies compared and 9 AI systems scored — are presented in the LongBook Verifier case study.
Approximately 64k-word single-book benchmark with a 40-question pilot benchmark, retrieval baselines, evidence-term scoring, answer-term scoring, and model-output evaluation.
240,767-word / approximately 320,220-token stress-test corpus with an 80-question extended benchmark, reports, plots, and derived evaluation artifacts.
Research package published. Local verify-only MVP prototype exists. Public upload demo is planned after authentication, rate limits, file-retention controls, and privacy protections are added.
The private manuscript corpus is not publicly redistributed. The public research package contains code, schemas, metrics, reports, plots, paper drafts, and derived evaluation artifacts.
BookProof is designed for people who need long-form narrative AI output to be checkable against source evidence.
Verify-only mode requires no paid model API. The user provides the manuscript and the AI output, and the verifier audits whether the output is grounded in the source.
Optional questions JSONL enables stronger scoring with expected evidence terms, gold-answer terms, QID detection, and answer-detection behavior.
Future generation mode may use local models or paid APIs only for paid/protected users, with privacy controls and file-retention policies in place first.
BookProof exposes a token-gated JSON API so agents and scripts can verify claims or questions against a document directly — no browser, no file upload. Deterministic local retrieval only; no external model calls.
Free public demo: deterministic retrieval verification, limited to one run per day — try it here. Developer API: token-gated for higher-limit scripted access. Future generation mode: may use local or paid models only for protected users.
Machine-readable spec: https://bookproof.bedvibe.studio/api/bookproof/v1/spec
curl -s -X POST https://bookproof.bedvibe.studio/api/bookproof/v1/verify \
-H "X-BookProof-Token: <token>" \
-H "Content-Type: application/json" \
-d '{"title":"My doc","text":"...document text...","claims":["a simple claim string",{"claim":"a richer claim","gold_answer":"expected answer","evidence_terms":["term1","term2"],"expected_chapters":["ch01"]}],"method":"flat_chunk_rag","top_k":5}'
Access is token-gated and limited to one active run at a time (HTTP 429 when busy). Contact BedVibe for an API token.