Historical source state
What was publicly knowable at a declared cutoff, with source versions and retrieval timestamps preserved.
Talent
Most AI systems explain what happened. We are building one that can be scored on what happens next.
NextConsensus reconstructs upstream medical evidence states, registers probabilities for defined authority actions, and resolves every forecast against a public outcome.
Forecasting under moving uncertainty.
A new trial is published. A guideline shifts. A competitor changes the relevant standard of care. A subgroup becomes clinically important.
Yet guideline bodies, regulators, and payers move on different schedules. NextConsensus exists to model those transitions without pretending institutional movement is the same thing as scientific truth.
Given the upstream evidence states visible at this cutoff, what is the probability that this named authority takes this defined action by the deadline?
That problem is not reducible to search, RAG, summarization, or workflow tooling. It requires normalized propositions, authority-specific histories, temporal leakage control, registered estimates, and prospective scoring.
The data asset is resolved temporal judgment, not a pile of papers.
The learning loop is: historical source state × proposition × registered probability × outcome or non-event × evaluation.
Normalized proposition
A target rung, named authority, defined action, scope, deadline, and independent resolution rule.
Probability history
Dated, method-versioned estimates that preserve every update instead of overwriting the record.
Outcome and non-event
What occurred, the highest rung attained, and where a transition stalled when the target action did not resolve.
Evaluation
Calibration, scoring, useful lead time, precision, recall, and alert burden for a declared cohort.
What makes this hard.
Literature search and summarization are becoming commodities. The hard problem is reconstructing what was knowable, defining what resolves, and evaluating probabilities without hindsight.
-
Proposition identity across sources and time
Authorities, actions, populations, deadlines, and resolution sources must remain comparable across historical reconstruction and prospective registration.
-
Temporal integrity
Every forecast is pinned to an evidence cutoff. Later information cannot enter the registered state. Hindsight does not leak in.
-
Resolution and evaluation
A correct-looking narrative is not enough. Outcomes need independent rules, non-events need structure, and performance needs calibration, baselines, and cohort metadata.
normalized transition propositionsauthority and source ontologieshistorical evidence-state reconstructionscientific entity, population, comparator, and endpoint matchingcalibrated probability estimationexplicit provenance and temporal replayindependent outcome adjudicationtemporal leakage controllearning from resolved transitions and structured non-events
What we believe about the work.
The hard problem is relational
The question is not whether a model can summarize a paper. It is whether observed upstream states change the probability of a defined authority action by a deadline.
We build scoreable forecast records
The core object is not a document or prompt. It is a frozen proposition, evidence state, probability history, resolution rule, outcome, and score.
We earn the long term
Resolved forecasts and structured non-events create the temporal corpus. Expansion follows demonstrated calibration by rung, authority, and therapeutic area—not ambition alone.
We build for expert judgment
NextConsensus does not replace medical, regulatory, legal, or compliance judgment. A probability is an input to a decision, not the decision itself.
We separate change from authority
Evidence state, forecast, outcome, and enterprise decision remain separate objects with separate authorities.
Who this work attracts.
People who are skeptical of thin copilots but interested in durable systems for evidence, time, provenance, institutional context, and accountable human judgment:
- You think medical AI should be evaluated prospectively, not trusted because its explanations sound plausible.
- You care about proposition design, temporal evidence states, transition hazards, calibration, and provenance.
- You are comfortable building systems where human adjudication is a designed component, not a gap to be engineered away.
- You can distinguish evidence, probability, outcome, and decision authority without collapsing them.
- You can write clearly about uncertainty without hedging into meaninglessness.
- You want to build the dataset that does not exist yet: how upstream evidence states become—or fail to become—defined authority transitions.
Applied AI and ML engineers
Provenance-aware retrieval, scientific document understanding, calibrated classification, temporal reasoning, graph retrieval, uncertainty estimation, and evaluation under expert disagreement.
Knowledge-graph and ontology engineers
Authorities, propositions, populations, interventions, evidence states, target rungs, resolution rules, and outcomes need durable representation.
Clinical informaticists and evidence scientists
Systematic review, HTA, pharmacoepidemiology, medical information, guideline methodology, clinical evidence synthesis, and regulatory science.
Regulated product and design leads
Expert workflow, audit trails, uncertainty communication, decision-state design, trust calibration, and interfaces that make reasoning inspectable.
Medical affairs and regulatory operators
People who understand how evidence progresses through expert recognition, procedural movement, guideline action, regulatory action, and coverage policy.
Infrastructure and systems engineers
Source observation, event pipelines, versioned state, tenant-specific logic, provenance preservation, auditability, permissions, and security.
Depth over coverage. Precision over generic summaries.
What we choose to optimize — and what we leave to others.
- We are not building "chat with PubMed." Search and summarization are basic features, not the product.
- We are not automating scientific truth or replacing expert review.
- We are not building approval or document-control infrastructure. Customers own governance and action.
- We are not presenting all 12 rungs as equally mature forecast classes. Initial formal targets stay at rungs 5–7.
- We choose depth over coverage: narrow propositions, temporal integrity, independent resolution, and prospective scoring before broad expansion.
We release infrastructure, not just records.
Sourced through Refract, our open-source developer SDK. It turns raw public revision histories into structured timelines. Anyone can inspect it, run it, or build on it.
Enterprise teams get access to demonstrated transition forecasts and private application support. Developers and researchers get the observation layer for structuring public knowledge change from version histories.
Not hiring now — but building the pipeline.
We open roles in bursts tied to product milestones. The best way to work with us today:
- Contribute to Refract — the open core. PRs, issues, and docs are the best way to demonstrate fit.
- Join the talent pool — we'll email when roles open (quarterly at most, no spam).
- Partner inquiry — if you're an org wanting to embed Refract or co-build the healthcare layer.
Method basis
Research threads behind the work.
The literature defines the problem space, not the finished system. The engineering work is temporal evidence reconstruction, proposition normalization, independent resolution, and prospective evaluation.
- The FAIR Guiding Principles for scientific data management and stewardship
Supports the need for durable, inspectable, reusable evidence and provenance records.
- How citation distortions create unfounded authority: analysis of a citation network
Shows how citation patterns can make a claim appear more settled than the source record supports.
- Guidelines for Human-AI Interaction
Supports interface patterns that keep uncertainty, user control, correction, and human authority visible in AI-assisted systems.
- Living systematic review: 1. Introduction-the why, what, when, and how
Frames evidence synthesis as an updating problem, not a one-time search problem.
Stay close to the work.
If building systems for temporal evidence, registered forecasting, calibrated uncertainty, provenance, and institutional accountability is more interesting than optimizing existing tooling, join the pool.