It began with one molecule.
The question was narrow — does protect the failing kidney? — and the answer was built not by mining PDFs, but by reading the trial registry itself, then rebuilt each time the evidence moved. That one living review became a method. The portfolio now runs to 947 living reviews — of which 617 are provenance-verified (only the curated 17 benchmark apps are externally validated) — each pools registered trials traced to their publications (PMID + registry), and the curated core is benchmarked against published meta-analyses and R metafor. The remainder are live but provenance-pending, shown in the dashboard without a pooled estimate. Open any one, and follow the evidence to its source.
From registries, not PDFs
Trials, arms and outcomes are read from structured registry data — so a review is refreshed as the evidence grows, not frozen at the date of publication.
Cochrane-grade, in the browser
REML and HKSJ pooling, prediction intervals, network meta-analysis, GRADE — the full Cochrane v6.5 apparatus, with nothing to install.
Auditable to the trial
Every estimate is cross-checked against R metafor and the published record — and every number traces back to the registration it came from.
Twenty open-access comparators, scored on a rule frozen before the first one was retrieved
The target was 40. The honest yield was 20. The standard was not moved. Both routes were on the table and fully costed. The selection criteria were frozen at commit fe1f2fd — before the first comparator was retrieved — and were never once amended: not for a new specialty, not for a target, not for a disappointing result.
This replaces a Cochrane comparison that could not be scored. In a first run, one condition failed unanimously across both judge families: we cannot enumerate Cochrane’s included studies, because Cochrane does not publish them under any free route. Against an open-access comparator that condition becomes satisfiable, so publishing an enumerable included-study list is a hard inclusion criterion here, not a quality bonus.
6,182 candidate meta-analyses screened (PubMed; no date, language or open-access filter) 1,105 full text retrieved 442 EXAMINED = retrieved AND an included-study list recovered 20 ELIGIBLE COMPARATORS 20 comparators / 14 independent topics / 10 drug families / 24 paper-topic pairs frames: cardiology 802 -> 12 · infectious disease 788 -> 2 · k>=4 1,998 -> 1 · remaining k>=2 2,594 -> 6 155 topic objects disposed: 49 framed + 65 framed + 4 excluded on their own declarations + 37 at k<2
The 37 are excluded by arithmetic: the match rule requires at least two overlapping trials, so a review with one contributing trial can never be matched by any comparator. The 4 are excluded on the objects’ own text — two say “DUPLICATE PAGE” and “RETIRED”, two are near-duplicates of framed topics.
Three caveats that a reader would otherwise find first
- Summing the frames gives 21; the union is 20. One comparator is eligible in two frames. The sum is the wrong number and the union is the right one.
- PMID 40998847 is eligible against three of our ablation topics — 3 of the 24 pairs, one paper, one drug family. It is closer to one demonstration than three. Two further comparators each span two topics inside a single family, so 20 comparators produce 24 pairs and 17 of them produce one each. Every figure here is reported as comparators / independent topics / drug families and never as a pair count.
- A mechanism this project asserted, wrote into a frozen rule, and then measured to be false. The orchestrating instruction argued that at k ≥ 4 the overlap threshold “becomes reachable at half”, and that sentence was written into the rule. The run disproved it. Half of a large k is a large absolute number: at k = 26 the threshold demands 13 overlapping trials and the best any comparator supplied was 4. The threshold is easiest at k = 4–6 and degrades in both directions — it demands 100% overlap at k = 2. The rule was not amended. It was frozen with its justification, the justification was wrong, and the refutation is published while the rule stands as run. That frame screened 1,998 candidates and returned one comparator.
Why the third caveat is the point. A protocol that records a mechanism its own run disproved — and does not quietly edit the rule to match the new understanding — cannot simultaneously be a protocol that was tuned toward a target. It costs a reader nothing to check, and it cost this project its most promising-looking route to 40.
What was frozen, and when
Selection rule fe1f2fd · join ruled at 5a42600 · infectious-disease inputs 7ad2538 · k ≥ 4 rule 05642cc · consolidated statement 10a5e74. Every frame’s inputs and every prediction were committed and pushed before that frame ran.
The join is the strongest single check. All three join options were computed and published before the choice, from one artefact: 22 comparators, 12, or 8. The chosen join gives 12 — smaller than the alternative — so the choice went against our own headline count. The evidence for it was a hand-read of three papers, not a look at which number was larger: a frozen trial acronym turned out to find mentions rather than inclusions, and two of the three papers read named the trial as ongoing.
One correction to this page’s own numbers, kept visible. A draft of this block said 451 examined. Two frames had counted rows whose retrieval succeeded rather than rows that reached examined, and one figure had never been measured at all. An independent audit, given 451 as a known-answer control, failed it and reported 442. The figure above is 442.
Quality here means prospectively registered in PROSPERO and nothing more — a procedural marker, not a measurement of conduct. It is the largest single filter on matched comparators, and it excludes some well-matched Cochrane reviews. Recorded, never adjusted.