NextConsensus
Start with one claim

Authority transitions leave tracks in the public record

That is the premise, and the research program tests it: estimates for named regulatory, guideline, and coverage decisions, registered before the outcome is known and then checked against what happened. The method is public so the answer does not depend on trusting us. The registry stands whichever way the answer comes out; About separates what is contributed now from what is under test.

Start with one claim Read the method

Registered questions, outcomes still pending

Registry snapshot: . Several records concern the same asset, with overlapping outcomes.

Registered
29
Resolved yes or no
0
Unresolved
29
Expired
0
Every registered forecast, by class

NCCN guideline reclassification

2 registered; 0 resolved, 2 unresolved, 0 expired.

FDA label revision

2 registered; 0 resolved, 2 unresolved, 0 expired.

FDA action by goal date

14 registered; 0 resolved, 14 unresolved, 0 expired.

NICE guidance by scheduled date

11 registered; 0 resolved, 11 unresolved, 0 expired.

  • Unresolved 29
  • Resolved 0
  • Not scored 0
  • Expired 0

One cell for each registered forecast; open one to read its record. A cell fills in when its forecast resolves, whichever way it came out. A forecast that could not be scored stays in its class count as a hatched cell.

Registration fixes what was estimated and when. Each forecast is scored once its deadline passes and the outcome is adjudicated, against the three baselines registered beside it.

Verify a record · Read the research questions

The questions the program asks

Can defined public medical transitions be forecast prospectively?

in progress

Do the numbers mean what they say? Does a 70% forecast happen about 70 times in 100?

planned

Does NextConsensus outperform base-rate, heuristic, expert, and LLM baselines?

planned

Which evidence and institutional features improve forecasts?

planned

Does an action by one authority make an action by another more likely, including across markets?

planned

Which authorities move first, and which follow?

planned

How far ahead can a forecast be issued before its false alarms cost more than the lead time is worth?

planned

Which target rungs are forecastable, and which are not?

planned

How many resolved forecasts does one kind of decision need before its record means anything?

planned

The transition model

Evidence, expert recognition, and the decision itself move at different speeds. A decision at rungs 5–7 of the 12-rung ladder is one of three public actions: a guideline body changing a recommendation, a regulator changing a label, a payer changing coverage. The model estimates the probability of one of those happening by a stated deadline, from what was observable at registration.

Illustrative sequence
Current stage
Evidence emerges
Registry state
Not registered

An observable upstream event enters the time-bounded evidence record.

How the program runs

Public research tests whether the method forecasts. Program work tests whether a forecast is useful in a decision. Neither result stands in for the other.

Retrospective backtesting

Reconstruct forecasts from historical evidence. Hold out future information. Score against outcomes. Control for temporal leakage.

Prospective registration

Register forecasts before outcomes are known. Freeze the evidence state. Publish registration timestamps. Adjudicate outcomes independently.

Calibration analysis

Compute Brier scores, reliability curves, and baseline comparisons per cohort, event class, and method version.

Boundary determination

Identify which target rungs can be forecast. Report verdicts: detection, nowcasting, forecasting, or unsupported.

Ranking which claims to review first

Between reconstruction and forecasting sits a step that ranks which claims to look at first, by how much their evidence has moved. It exists as a review workflow; the validation below decides whether it becomes a program.

Claim prioritization

The direction and strength of movement on rungs 1–4 (support attaching, qualifiers falling away, spread across sources, or regulator language arriving early), ordered into a review queue with the record and its limits.

What would settle it

Whether the queue ranks claims that later moved on rungs 5–7 higher than those that did not, measured on held-out history with the cutoff enforced. An earlier proof cycle that appeared to show this was invalid; the corrected rebuild found no lift in general revision activity, so the test now runs on specific claims and specific patterns.

Method demonstrations

Two reconstructions built to test the method, not an asset: whether one claim can be held across years of public edits, and whether a second analyst working independently arrives at the same claim families. Both use a public encyclopedia's revision history, the largest openly versioned medical text available.

What the program publishes

  • Registered forecast specifications with defined resolution rules
  • Retrospective backtests that control for temporal leakage
  • Prospective forecast cohorts with independent adjudication
  • Calibration analyses by event class and authority
  • Baseline comparison results
  • Open-source tooling via Refract Refract on GitHub

Method basis

Reading

The works this program leans on. They frame the problem — evidence records age, guidelines lag, and citations make claims look more settled than they are.