Pipeline
From raw revision history to a ranked list of claims.
Four stages, each one auditable: extract the trajectory, classify what kind of change it was, rank what it means for you, deliver it with the evidence attached. Forecasting is a research program built on top — not yet a proven capability.
The pipeline
Extract, classify, prioritize, deliver
The classify step is the one that carries the weight. Without it, a burst of editorial housekeeping looks identical to a claim genuinely gaining ground — and a system that cannot tell those apart will confidently rank the wrong things.
Extract
Refract reconstructs claim trajectories from public revision history. Every wording change, citation, certainty shift, and propagation event is dated and verifiable.
Classify
Each event is labeled: real evidence movement, public assimilation, authority-language importation, editorial maintenance, or noise. This prevents attractive but misleading outputs.
Prioritize
Direction and strength of movement — gaining stronger support, becoming harder to dismiss, spreading across influential sources, losing qualifiers, appearing in regulator or guideline language — rank what deserves attention now.
Deliver
Prioritized claim intelligence arrives with full trajectory evidence — early enough to investigate, brief, or escalate before the decision is obvious.
Three layers of movement
Evidence, public representation, and institutions move at different speeds.
That gap is the commercial opportunity. Customers want to know where evidence is ahead of policy, where public representation is ahead of institutions, and where authority language is spreading after formal action.
What studies, reviews, labels, guidelines, and safety communications say over time. The primary scientific record.
How public and professional knowledge sources describe and support the claim. Wording, citations, certainty, prominence, propagation.
How regulators, guideline bodies, payers, and other authorities formally act. Label changes, guideline updates, coverage decisions.
Architecture
The temporal ladder
Medical evidence climbs a ladder of recognition — from first publication through replication and expert consensus to formal institutional action. NextConsensus tracks how claims move through these rungs. Refract observes rungs 1–4. NextConsensus is testing whether that observation predicts rungs 5–7.
Reconstructs claim trajectories from public revision history: emergence, wording evolution, evidence attachment, persistence, spread across influential sources. The method.
Tests whether trajectory features — persistence, citation quality, velocity, propagation — predict institutional transitions. Prospectively registered, not yet calibrated. The hypothesis.
Calibration, baseline outperformance, and decision-useful lead time. These must be earned through prospective scoring, not asserted. The future.
For a complete breakdown of each state's observable signals, roles, and validation criteria, read the Evidence-to-Action Ladder reference.
Provenance
Every event is traceable to a public revision
Every claim trajectory event — emergence, wording change, citation attachment, certainty shift, persistence, removal, propagation — has a dated public source and a reproducible hash. You can recompute any trajectory from the public record and check it yourself.
The system maintains the same claim across paraphrase, wording changes, broadening, narrowing, qualification, relocation, removal, reintroduction, and citation replacement.
A public source may change because of new evidence, delayed editor attention, imported regulator language, citation cleanup, editorial judgment, or routine maintenance. The system labels each — it does not automatically claim "evidence strengthened."
Every trajectory is byte-for-byte reproducible from the public record. Registration hashes let anyone verify that the trajectory has not been altered.
Research program
What NextConsensus is testing — and what it has not yet proven
The forecasting layer is a hypothesis under prospective test, not a product capability. An earlier proof cycle appeared to show that public revision activity predicted authority actions, but the result was invalid: positive cases had full revision histories while negatives had truncated data. With correctly ingested negatives, no signal remained in revision volume, section activity, content changes, safety-term mentions, or composite scores.
The signal, if one exists, is more likely at the level of specific claims and specific patterns of claim movement — not general activity. That is the next experiment. Until it is proven, forecasting stays in the research program.
Prospective calibration. No forecast is scored yet — numbers appear only after a registered forecast resolves. FF-001 is registered, first in the ledger.
Baseline outperformance. Whether trajectory features improve over base-rate forecasts is the research question — not a claim we make.
Decision-useful lead time. Whether the system identifies movement earlier than existing workflows is unproven.
Generalization across markets. Results from one therapeutic area, authority, or event class may not transfer.