Back to Blog

Why 3-Layer Provenance Matters for AI Study & Research Notes

Provenance4 min read

A fluent note can hide a fragile chain of decisions. Speech recognition mishears a term, a language model repairs it, and a summary repeats the repair as fact. If those steps overwrite each other, you can no longer tell what the source said. TomeSift keeps that chain inspectable.

The cost of silent rewriting

Consider a hypothetical lecture transcript: “The intervention did not improve recall.” Removing one small word changes the conclusion. An assistant trying to make a sentence sound more natural can introduce exactly that kind of change. A polished paragraph gives no clue that the meaning moved unless the earlier version is still available.

Provenance is the record that lets a reader distinguish an observation from an interpretation. For study material, that record needs to survive extraction, cleanup, and summarization. Keeping only the final answer makes every later check depend on reconstructing a history that was never saved.

Layer 1: preserve the original & the raw extraction

TomeSift fingerprints the uploaded source with SHA-256 and retains raw ASR or OCR separately from later edits. The source file and the machine’s first extraction have different roles: the file is the material you supplied; the extraction is a machine reading of it. “Raw” means untouched by subsequent repair, not guaranteed accurate.

A SHA-256 fingerprint identifies the exact bytes used. It can help detect whether a file changed; it does not certify the speaker’s claims, the author’s identity, or the correctness of the transcription. That boundary matters. Cryptographic integrity and factual accuracy answer different questions.

This layer gives later review a stable reference. If a specialist term looks wrong in a brief, you should be able to return to the raw text and the relevant source location, rather than trusting a second generated explanation of the first generated answer.

Layer 2: make repair auditable

Repair has a legitimate job. It can restore punctuation, fix a likely recognition error, or make a sentence readable. TomeSift stores that interpretation in Layer 2 and exposes word-level additions and deletions. The original extraction remains available alongside it.

  • Check meaning-changing edits first: negation, quantities, units, names, and technical terminology.
  • Compare a proposed repair with the source audio or page. Grammatical fluency alone is weak evidence for a correction.
  • When the source is ambiguous, preserve the uncertainty rather than completing a plausible sentence as if it were observed.

A word diff makes changes visible; it does not approve them. The reader still decides whether a proposed repair is justified. Separating this layer also makes disagreement useful: you can reject an interpretation without losing the source or the history of the proposed change.

Layer 3: distill with citations & limits

Distillation makes a different promise from transcription. It selects ideas, compresses explanations, and relates claims to evidence. TomeSift keeps this work in Layer 3, with citations and explicit evidence verdicts rather than presenting a concise paragraph as an unquestionable record.

A source anchor tells you where a claim came from in your material. An external citation helps you examine supporting or conflicting literature. Those are complementary links, not substitutes. A paper discussing the same topic does not automatically support the exact claim in a lecture.

Preprints from arXiv, medRxiv, and bioRxiv carry a mandatory “Preprint · Not Peer-Reviewed” label. A “Supported” verdict describes a relationship to retrieved evidence; it does not erase publication status, establish clinical advice, or make a legal conclusion authoritative.

Use the layers as a review workflow

Start with the distilled brief to orient yourself. Follow a consequential claim to its source anchor. Inspect any repair that changes the wording, then return to the raw extraction and original material when the distinction matters. Read external evidence in its own context and check its publication status.

The architecture supports that separation: the browser workbench presents the layers, the Rust API coordinates jobs and credit accounting, and the Python processing pipeline produces distinct artifacts. Private R2 object storage and Neon Postgres hold sources, artifacts, and their records. The design does not require the final prose to carry the entire history in its sentences.

Sources & further reading

Provider documentation and research infrastructure behind the discussion. Pricing and terms can change; check the linked provider before choosing a service.