Skip to content

Temporal Model

Refract is designed around time — not just "what is the latest version?" but "how did this claim change across revisions?"

Most document tools answer from the present corpus. Refract answers across time:

  • Rewind: What did this claim say two years ago?
  • Replay: How did this claim drift across revisions?
  • Compare: What changed between the 2024 and 2025 guideline versions?
  • Audit: What was supportable at the time of a specific decision?
  • Forecast: What would trigger re-review?

A vector database can find similar claims. It cannot do any of the above. Refract fills that gap with deterministic, provenance-backed temporal infrastructure.

Refract emits structured events that capture how claims change over time:

Event Meaning
sentence_first_seen Claim appeared in the source for the first time
sentence_modified Claim wording changed (with edit magnitude)
sentence_removed Claim removed from the source
citation_added New supporting citation attached
citation_removed Citation removed or superseded
revert_detected Edit summary matches a revert pattern

Each event carries deterministic enrichment fields:

  • editMagnitude — minor / moderate / major
  • contentChange — introduction / removal / expansion / compression / refinement / rewrite
  • directionSignal — strengthening / weakening / neutral
  • certaintyProfile — counts of hedging and certainty markers
  • quantitativeFindings — extracted p-values, hazard ratios, and sample sizes

All fields are byte-reproducible. Same source, same events, every time.

Refract provides the timeline. Downstream systems provide the judgment.

NextConsensus, which maintains Refract, reads these timelines as one input to its own work on how claims move across institutions; what it concludes from them is its judgment, made downstream and described on its own site. Refract itself remains domain-neutral — it works on any versioned text source.