A public standard for reviewable forecasts
Each forecast specifies who decides, the action, the deadline, and the evidence cutoff before the outcome is known. The registered terms remain available for review after the decision. FF-001 is the first record. Whether a claim's public history improves these forecasts is a separate research question.
What a forecast is made of
"Will payers cover our drug?" cannot be scored. A forecast is that question rewritten until a named regulator, guideline body, or payer either takes a defined action by a date or does not, and every part of it is fixed before the outcome is known.
Specific enough that an independent reader can tell whether it happened: NCCN lists the agent first-line; FDA approves the supplemental application; CMS issues the determination.
The date by which the action must happen, and the evidence cutoff: the last date a source can carry and still be used. Both are written into the record at registration, so the forecast reflects only what was public by the cutoff.
Which document settles it and how to read it, written before the window opens. Listing confined to a later line of therapy does not count; a withdrawn agent resolves NO rather than being voided.
A base rate from comparable cases, a model that knows only the authority's calendar, and a reconstruction of what an analyst reading the same record would say — each frozen at the same cutoff.
How a claim keeps its identity across revisions
A claim is rarely deleted. It is reworded, moved to another section, qualified, broadened, narrowed, given a new citation, and sometimes put back after being taken out. The claim keeps one identity across all of those steps. A sentence that says "reduces" where it used to say "may reduce" is the same claim with a qualifier gone. A sentence that now names a different population is a different claim, and the history says so rather than carrying the old identity forward.
Keeping the claim linked across revisions lets us distinguish changes in evidence from rewording and citation cleanup. We label evidence as strengthened only when new evidence supports that finding. The illustration below shows a qualifier added, removed, and restored by different authors while the claim retains the same identity.
Keeping hindsight out
NextConsensus freezes every source at the evidence cutoff date; the probability, proposition, baselines, and sources are locked at registration and never retroactively modified. If new trial readouts or policy revisions occur later, they go into a new, distinct forecast; the earlier one stands on its original record, and both are in the registry, each with its hash.
Nobody grades their own work
We read the evidence and make the estimate. An adjudicator with no stake in the forecast determines whether the specified action occurred, applying the registered rule to its named source. Their job is narrow on purpose: they do not weigh evidence and they do not touch probabilities.
Being named as a reader means applying the frozen rule to the public record; it is not support for the estimate, the method, or NextConsensus.
- Step 1 The deadline passes The record stays as registered.
- Step 2 The window closes Nothing is adjudicated before it does.
- Step 3 The rule is applied The rule written at registration, to the source it names.
- Step 4 What is recorded
Scored
- Occurred on time
- Did not occur
- Occurred after the deadline
- Occurred in part; the rule was not met
Not scored
- Ambiguous, handled by the rule
- Could not be resolved
- Cancelled by an outside event
No finding
Posted with no finding. It stays in the class count. - Step 5 It is posted In the same layout whether it occurred or not, with the count for its class beside it.
- Step 6 The class count updates Registered, resolved, unresolved, and expired.
How a record is scored
Each resolved forecast is scored against its outcome using a proper scoring rule, declared in advance, beside the baselines it was registered against. Results are published per reference class with the denominator, the unresolved and expired records, the method version, and the date range needed to read them. Lead time is never reported on its own: any system can look early by calling everything early, so a lead-time figure comes with the precision and the false alarms it cost. A forecast that does not resolve is classified by why, not scored as a miss.
How results will be reported
Written down on 2026-09-26, when the registry held 22 registered forecasts and none had resolved, so the way results are shown was not chosen after seeing one.
Every forecast that reaches its deadline is posted, whether it occurred or not, in the same layout.
A posted resolution shows who adjudicated it, the sources they read where the adjudication lists them, the finding, and the registered probability beside the baselines it was registered against, all scored by one rule: squared error.
Each posted resolution carries the count for its class: registered, resolved, unresolved, and expired. One forecast is one instance, not a rate.
NCCN guideline reclassification, FDA label revision, FDA action by goal date, and NICE guidance by scheduled date are counted apart.
No figure across a class, such as the share that occurred or an average score, appears until the class has 15 admissible resolved forecasts. Forecasts that turn on the same decision count once.
A forecast that cannot be resolved is posted with its status, and stays in the class count: ambiguous, could not be resolved, or cancelled by an outside event.
The same two reader roles are assigned to every forecast registered so far.
Within each class, the forecasts whose registered probability sits 15 points or more from their base-rate baseline are counted and scored as one group, and the rest of the class as another, by the same rules. Which group a forecast is in is fixed by those two numbers at registration, before any outcome. Neither group shows a figure until it has 15 admissible resolved forecasts of its own.
“Departures of 15 points are shown apart” was added on 2026-10-06, when the registry held 22 registered forecasts and none had resolved.
Where a chart groups forecasts by registered probability, it uses five bins of equal width, each 20 points wide, fixed here before any outcome: 0–20%, 20–40%, 40–60%, 60–80%, 80–100%. Bins chosen after the results are a way to tune a picture.
When we decline to forecast
When the public evidence cannot support an estimate, we decline to forecast. We publish the reason and date beside the scoring record so readers can see which questions were declined.
The public documents are too thin to support an estimate.
The documents support two readings, and the rule cannot choose between them.
Credible public sources disagree, and the record preserves the disagreement.
The question needs something the declared record does not contain.
The works this standard leans on are listed once on the research page, with the contribution of each.