How Forefront assesses confidence

Study design

Each piece of evidence is placed in a design tier, and that tier is the dominant term in how much it weighs. A meta-analysis counts for more than a randomised trial, which counts for more than a cohort study, which counts for more than a case series. Trial design is read from the registry's design fields — allocation, intervention model, study type — and not from the trial's phase, because a randomised masked trial is a randomised trial even when its phase is unrecorded.

An engineering heuristic encoding the conventional evidence hierarchy. This is not a study-quality assessment: a double-blind placebo-controlled trial and an open-label randomised trial receive identical weight.

Limitation. Design tier is the only quality-like input the score has. A double-blind placebo-controlled trial and an open-label randomised trial of the same size receive identical weight, because blinding and pre-registration are never populated on production evidence.

Direction

Evidence moves an assessment only if the engine can establish which way it points. There are four outcomes, and the fourth is the common one.

• Supports — Counts for the claim. It raises the score, and it narrows the uncertainty range, because there is now more evidence that resolves the question either way.
• Contradicts — Counts against the claim. It lowers the score, and it widens the uncertainty range, because the evidence is now in dispute with itself.
• No effect — A study large enough to have detected an effect, which found none. It counts against the claim, and it does two things to the range at once: it narrows it, because the question is better resolved than before, and it widens it where the claim also has real support, because a null and a positive result disagreeing is a genuine dispute.
• Neutral — On topic, but the direction could not be read from the record. It contributes nothing at all: it does not move the score and it does not affect the range. It is kept only so the evidence counts stay honest about what was found and not used.

Limitation. Most retrieved records are neutral. A study existing is never treated as the study agreeing: where direction cannot be established from the record, the study contributes nothing to the number and appears only in the coverage figures.

Relevance to the claim

Evidence that does not quite match the claim is discounted rather than discarded. The engine compares three things: whether the study was treating or preventing, whether the enrolled population is the one the claim names, and whether it falls inside the claim's age band.

Unknown or mainstream scope receives no discount, so an unclassifiable study is never penalised for being unclassifiable.

Limitation. Relevance is coarse. Dose, duration, follow-up length, study setting, geography and ancestry are not modelled at all, and evidence is never transferred between different diseases.

Correlation and de-duplication

Repeated results from the same source are not repeated evidence. The engine handles this in two distinct steps, and claims only what it can actually detect.

First, de-duplication. Publications reporting the same underlying experiment collapse to a single strongest contribution, identified by registry number or trial acronym. Five papers about one trial count about once.

Second, correlation grouping. What survives is clustered by whatever shared influence the record reveals — a commercial sponsor, a research team, an institution, a publication venue. Inside a cluster each further piece of evidence contributes less than the last, so a single source cannot keep raising an assessment however much it publishes.

A detected correlation cluster (sponsor, research team, institution, venue, or shared trial). A floor on the correlation present, never proof of independence — undetected correlation stays undetected.

What we can detect is bounded, and it is bounded in a direction worth knowing about — every limit below causes evidence to be treated as less related than it really is, never more:

• Who funded a publication is read from grant records. Company-funded trials are rarely written up as grants, so much industry-funded literature ends up looking unattributed and is treated as its own separate source.
• The map from a trial programme to the company behind it is applied to reviews and summaries, but not to primary research papers.
• A paper that names a trial only by its study name is not matched to the same trial's registry entry, so those two accounts of one experiment are not recognised as the same experiment.
• Shared investigators, shared patient cohorts, shared contract research organisations and declared conflicts of interest are not detected at all.

Limitation. Grouping is a floor on the correlation present, never proof of independence. The limits above are the ones we know about; correlation we cannot see from the record stays undetected, and the engine never describes evidence as independent on the strength of having failed to link it.

Extraction confidence

When the engine reads a result it also records how much that one reading of that one result should count as directional evidence. This is the quantity most often misread, so it is worth stating precisely.

How much one reading of one reported result should count as directional evidence. Four things set it, and only the first is legibility in the narrow sense: where in the record the direction was read from (a title counts for less than a results section); whether the between-arm comparison was statistically significant; how probative the comparator was (placebo, standard care, none); and how many participants are known to have produced the result. Not study quality, not risk of bias, not statistical power, and not confidence that the result is correct.

A reading graded high, medium or low carries 1, 0.6 or 0.3 of its otherwise-computed weight.

Limitation. Extraction confidence says nothing about whether a study was well conducted, and it is not a statement that the extracted reading is correct. It bounds how far one reading of one reported result is allowed to carry an assessment.

Sample size

A larger study is measured more precisely. It is not thereby more supportive. The engine encodes exactly that distinction: sample size enters in two bounded places, and neither is a general multiplier on evidence weight.

First, the uncertainty range. Each resolving experiment carries a bounded size adjustment anchored at 1,000 participants, moving by 0.25 per ten-fold change in size and clamped to [0.5, 1.5]. These are mass-weighted across the resolving experiments and scale the range. Larger studies narrow it; smaller ones widen it.

Second, a ceiling on small supporting readings. A supporting read drawn from fewer than 30 participants can rise no higher than medium extraction confidence, and below 10 no higher than low. This is a ceiling, not a multiplier: it can only lower a reading that was above it, it is idempotent where a reading already sits below it, and it never touches contradicting or null readings.

Where the count is unknown it stays unknown. An absent sample size is neutral — an adjustment of exactly 1.0, and no ceiling applies. Unknown is never treated as small.

Limitation. The two places read different counts, deliberately and imperfectly. The ceiling judges the smallest population genuinely known about a reading; the well-powered-null gate still reads the recorded registry enrolment. Publications carry no structured participant count at all, so neither the size adjustment nor the ceiling can apply to them.

Uncertainty

The range is computed from the same evidence as the score, by a deterministic rule.

A deterministic heuristic half-width: the reciprocal of accumulated resolving mass — scaled by a bounded study-size adjustment — plus a contradiction term, plus a flat single-source penalty. Study size enters here and nowhere else. It has no sampling distribution and no coverage guarantee: two five-million-participant trials still yield a range of several points, where a statistical interval would be a fraction of one.

No sampling distribution, no likelihood, no posterior, no coverage guarantee. It is symmetric by construction and clamped, so a 0% assessment displays a range that extends only upward.

• Narrowed by accumulating supporting mass.
• Narrowed by accumulating well-powered null mass.
• Narrowed by evidence spread across two or more source groups.
• Widened by thin evidence.
• Widened by contradiction (support vs harm).
• Widened by all decisive evidence sitting in one source group.

Limitation. The two kinds of disagreement do not widen the range by the same amount. A study finding no effect and a study finding harm now both register as disagreement, on the same scale — but a null also counts as evidence that resolves the question, so a support-versus-null split still reads as somewhat more settled than an equal support-versus-harm one. That is deliberate rather than an oversight: a null answers the question, a contradiction disputes the answer.