readoutbench/readoutbench-r1

ReadoutBench Round 1: forecast 653 company trials before they report

These are the rules of Round 1 as it opens on 1 November 2026. They were revised through many rounds of outside review before launch, by AI review agents run independently of the build. The additions of 29 to 31 October (three deadlines, a ranked number score, and the safety, key secondary, subgroup, next-step and amendment questions) were checked by independent agents that wrote none of them, and their mechanics by the later outside AI reviews, which did not re-examine the design of these questions; the code that applies them, every results check and the build are published in the round's GitHub repository (https://github.com/readoutbench/readoutbench).

Round 1 of ReadoutBench has 653 trials. Each is a phase 2 or phase 3 trial registered on ClinicalTrials.gov as randomized, with these properties (548 of them have an efficacy or equivalence comparison the round scores; the other 105 do not, for example doses compared with no control or a design that is in fact single-arm, and are asked operations questions only):

• run by a company (not an individual sponsor-investigator or a non-profit, whatever the registry's sponsor class says);
• from any therapeutic area;
• primary completion, as registered, between July 2026 and March 2027;
• not stopped, and no public result of its primary outcome found on the launch date.

It is a prospective benchmark of forecasts of how these trials read out: the trials are a selected cohort, and only those that report by the deadline are scored for p_met, so the scores describe forecasting this cohort's readouts, not clinical-trial forecasting in general. There is no human-expert arm in Round 1.

The trials ask these kinds of questions:

Trials	Will it work? (p_met, at three deadlines)	Numbers (control, effect; subgroups)	Safety, key secondary, next step	Will a result be out? (p_public, at three deadlines)	Operations (p_late, p_stopped, p_posted, p_amended)
536 superiority and non-inferiority trials	yes	509 of them; subgroups on 275 phase 3 trials	yes (key secondary where the record has one)	yes	yes
12 equivalence trials (biosimilars, bioequivalence, lot consistency)	yes: equivalence shown	no	next step only	yes	yes
105 trials with no efficacy goal to meet (safety only, pilot, descriptive, no control group)	no	no	no	no	yes

There are two boards. Models: one model, answering from the packet alone with the round's instructions unchanged; it compares models. Open: anything else, any harness, tools, web access or several models; it shows how far a system can go.

There are two rankings, both in tiers against a fixed reference, set once the first pass's labels are complete: p_met against the registry model (B1), and the number score against the frozen Opus reference entry. Everything else is shown on each entry's scorecard and never ranked.

Each trial comes with three files:

• its packet: the ClinicalTrials.gov record as pulled for the round (packets/; the pull date is in the packet), followed by its question card and its operations questions;
• its question card (cards.jsonl), which fixes the endpoint, unit, arms, design and planned arm sizes the questions refer to;
• its operations row (operations.jsonl): the status and registered primary completion date on the launch date, and which operations questions it asks.

sisters.json lists the sister-trial groups (trials of the same experimental drug) that the scoring resamples together. forecast_instructions.md is the round's prompt, the one the Models board uses. manifest.json holds the SHA-256 of every packet, of the cards, of the operations rows, of the sister groups and of the instructions. Before the round opens, the organiser posts the manifest's SHA-256 and the reference entries' SHA-256s outside the submission page, where it is dated and public: the round's GitHub repository (fingerprints/2026-r1.txt, and the release 2026-r1-fingerprints at https://github.com/readoutbench/readoutbench/releases) (see Reference entries), and while the round is open posts a fingerprint of the list of uploads every week and at the close.

How the 653 were chosen: 1,085 company trials completing in the window → 116 out with stale records (still recruiting after their own end date) → 14 stopped → 79 already reported or stopped (the first results check, 25 and 26 October) → 21 held for review → 1 mock study → 6 with an individual or non-profit sponsor → 6 from one sponsor not expected to give a verifiable result (single-site trials of unapproved peptides, one a copy of a published trial) → 4 whose stop was announced in public before launch, though the registry did not yet show it (found by a web check of the operations-only trials past their completion date) → 121 out after the results check of every trial on 29 October (a PubMed search by NCT number, then a web search of all 836 remaining trials for results, interim readouts and announced stops, each finding confirmed by a second agent): 98 with their primary result public (2 of them found by the earlier checks), 19 with a stop announced, and 4 left out to be safe where the second check could not settle it → 717. A trial can report its primary result at an interim analysis long before its registered primary completion date, which is why the check covered every trial, not only those past that date. The records were pulled again on 29 October; none had stopped or posted results since. → 8 out after a final, independent results check of all 717 on 29 October (agents that had not done the first check, counting any public statement on any registered primary outcome, including safety, drug-level and immune-response primaries, and searching for programme and pooled readouts; every trial with such a statement is out, with no case-by-case exception): NCT04647240, NCT05540860, NCT05683457, NCT06100887, NCT06260722, NCT06369298, NCT07033481, NCT07222670 → 1 taken out because the organiser's Opus 5.5 run could not answer it (NCT07385287, see below) → 7 out after a deeper check of every trial on 29 and 30 October (at least two searches, one in Chinese, Japanese or Korean for a trial run only in East Asia, and at least one page opened that is not a registry listing; results_gate.py applies this bar to every trial left in): NCT04951219, NCT05445908, NCT07038382, NCT07251309 (results public), NCT07229950 (results public), NCT07300839, NCT06939595 (stop announced) → 3 out found by outside reviewers on 30 October, each with its primary result public before launch: NCT07165223 (16 October), NCT06388941 (an ACR Convergence abstract, 20 October), NCT07576868 (blinded safety data, 27 October) → 698 → 12 out because no check could reach the bar before the round opened (the bar was raised on 30 October so that registry copies, filings on sec.gov and news-repost sites no longer count as an opened page; a second deeper check of the 116 trials that then fell short cleared 103 and found one more result): NCT05555862, NCT06125327, NCT06206564, NCT06257966, NCT06304103, NCT06309485, NCT06769659, NCT07121413, NCT07127042, NCT07413094, NCT07494409, NCT07672561 → 686 → 3 out after a second outside review on 30 October: NCT05217628 (primary result in an IASP World Congress abstract, 27 October), NCT04042844 (blinded efficacy data on the primary released in March and May 2026) and NCT05456425 (topline overdue: database locked in September, topline guided for the third quarter, so a readout in the round's first days was likely) → 683 → 2 more taken out by the organiser on 30 October: NCT06989918 (topline overdue: guided for the third quarter, none public) and NCT07669454 (a Malaysian ministry called a Nuvastatic phase 3 in breast cancer successful in May; this record's primary outcome names breast cancer, so the statement may cover it) → 681 → 2 more taken out on 30 October after a fresh audit sample: NCT06111235 (topline overdue: guided for the first half of 2026, none public) and NCT06474806 (the sponsor said its completed Part 1 "supported the feasibility" of the imaging approach, a statement on that part's primary) → 679 → 16 more left out on 30 October because no check reached the bar after it was raised again (trial-listing sites, including sponsors' own trial sites and CTIS, and stock-news and summary sites no longer count as an opened page; the 151 trials that then fell short were checked again, and 135 cleared it, with no result found): NCT02616302, NCT05267613, NCT05884398, NCT05888922, NCT06194734, NCT06470191, NCT07087158, NCT07107256, NCT07285291, NCT07372924, NCT07521631, NCT07598383, NCT07790120, NCT07791524, NCT07801820, NCT07804797 → 663 → 5 out after an outside review on 30 October: NCT06351020 (an HKEX announcement of 23 June 2026: the interim analysis met both primary endpoints), NCT06843018 (a journal article of 4 March 2026), NCT06577519 (an approval application announced on 26 June 2026 with a trend favouring the drug on the primary), NCT06447597 (terminated in China's drug-trial register on 13 April 2026) and NCT07118241 (left out: no check reached the bar; an earlier version said so but had not done it) → 658 → a native-language sweep on 30 October of 187 trials, every trial run only in China, Korea, Japan, Taiwan, Hong Kong or Russia and every trial that failed the stricter bar, by agents told nothing about earlier checks and required to search in the sponsor's language, its stock-exchange filings, applications for approval and local journals: 1 out (NCT05751928: a different endpoint presented as the primary analysis at ESMO 2024) and 16 left out because the sweep could not reach the bar for them (NCT04208178, NCT05413655, NCT05943795, NCT06434610, NCT06519370, NCT06708260, NCT06716203, NCT06799559, NCT06800651, NCT06860971, NCT06993610, NCT07032597, NCT07057622, NCT07473960, NCT07572396, NCT07719530) → 641 → 1 more taken out on 30 October, under the same rule as NCT05456425: NCT06776432 (topline guided for the first half of 2026, none public) → 640 → 1 out after an outside review on 30 October: NCT04442022 (the sponsor said on 24 March 2026 that it would cease development in this indication, and the accelerated approval was withdrawn on 30 April; the registry still lists the trial as active) → 639 → a strict recheck on 30 October of the 71 trials whose check carried a doubt the checking agent raised (wording such as "recheck", "cannot be ruled out", "low confidence"), requiring an opened page about the trial or its drug: 1 out with its result public (NCT06335173: topline released on 30 October, primary not met), 7 kept because the doubt was only that a readout is due soon (signed off by the organiser and on the required watch list for 31 October), and 35 left out because the recheck could not reach the bar or raised the doubt again: NCT03592472, NCT04701476, NCT05087056, NCT05116462, NCT05139719, NCT05372718, NCT05654454, NCT05995340, NCT06008093, NCT06029972, NCT06128447, NCT06161441, NCT06292923, NCT06339008, NCT06403267, NCT06507436, NCT06512584, NCT06569056, NCT06619860, NCT06632951, NCT06641466, NCT06664775, NCT06671548, NCT06750289, NCT06772168, NCT06841731, NCT06842966, NCT06881836, NCT06925594, NCT07192588, NCT07230860, NCT07298434, NCT07446309, NCT07489612, NCT07776210 → 603 → a stop check on 30 October of the other 568 trials (mostly one search per trial for words such as discontinue, terminate or pause with the sponsor and drug, in the local language where the trial runs only in East Asia or Russia; 35 trials got two or more searches, and a page was opened only to confirm a possible stop, so most rows cite no source; it did not pull EU or SEC records trial by trial, and the FDA's list of withdrawn accelerated approvals was checked separately): 6 out with a stop or pause announced before launch that the registry does not show (NCT03878849, NCT06132893, NCT06593522, NCT06668805, NCT06775379, NCT06965413) and 1 left out because it could not be searched (NCT07241377) → 596 → 11 trials whose only opened page was a copy of an exchange filing on pdf.dfcfw.com (Eastmoney's PDF host, now refused like eastmoney itself) were rechecked on 30 October under the strict rule: 4 cleared it, no result was found, and 7 are left out because the recheck could not open a page about the trial or its drug: NCT06707883, NCT06789289, NCT06954961, NCT07090759, NCT07124221, NCT07396337, NCT07791537 → 589 → on 30 October the organiser put back all 88 trials left out at the steps above because no check reached the bar (no result or stop had been found for any of them; for 48 of them the checking agent had written down a specific doubt): they are marked "check": "below the bar" in questions.json, listed with the reason in below_bar.jsonl, and the leaderboard shows each entry's Brier score, skill against the registry model and tier also without them → 677 → 3 out after a third outside review on 30 October: NCT06671938 and NCT06107413 (a public statement on a registered safety primary: a sponsor's "favourable safety profile" and an investigator's interview on the study's safety) and NCT05462587 (a web page saying the phase 3 had concluded, low credibility but taken out under the literal rule) → 674 → a scan on 30 October of the 179 trials whose primary outcome is safety, tolerability, pharmacokinetics or immune response, or that have no efficacy goal, for any public statement on that primary ("well tolerated", a monitoring committee's review, a dose-selection outcome, blinded event rates; safety_statement_check_instructions.md): 6 out (NCT05202509, NCT05389215, NCT05956626, NCT06016738, NCT06125951, NCT07142291) → 668 → 1 out after a fourth outside review on 30 October: NCT06292923 (below the bar; topline guided for October 2026 after its last visit on 21 September, none public, under the same rule as NCT05456425) → 667 → 3 out after a fifth outside review on 30 October: NCT04208178 (its registered primary, Part 1 dose-limiting toxicities, was presented at AACR 2023, and its record says recruitment was halted for good in December 2022; both checks had missed it) and NCT06109272 and NCT06632951 (AbbVie removed these livmoniplimab programmes, first-line liver cancer and urothelial cancer, from its pipeline on 4 February 2026; the registry still lists both as active) → 664 → 5 out after a sixth outside review on 30 October: NCT05822752 (a third livmoniplimab programme AbbVie removed, second-line liver cancer), NCT05846230 (pooled long-term safety, its primary, at ATS 2026), NCT06534892 (preliminary results on its immunogenicity primary shown to ACIP in April 2025), NCT06284265 (the sponsor said on completion that the drug improves the effectiveness of therapy) and NCT05778188 (a safety statement on its safety primary). A second look at the 106 operations-only trials left, told to search congress abstracts, advisory-committee slides, investor decks and local trade press but in practice one or two web searches per trial (69 got one, 37 got two; most rows cite no source), found nothing more → 659 → 1 out after a seventh outside review on 30 October: NCT06465563 (its primary result, ORR 45.0% vs 27.5%, was an ESMO 2026 abstract, covered on 24 October) → 658 → 2 out after an eighth outside review on 30 October: NCT05884255 (LUMINET-1: its primary result, HR 0.06 for progression-free survival, was a late-breaking abstract at ESMO Asia 2025 on 6 December 2025, published in an Annals of Oncology supplement and on the sponsor's news page; one of our own check rows had cited that result while checking a different trial and did not act on it) and NCT06080789 (the sponsor said in a release on its June 2026 EPC presentation that the drug was well tolerated with no adverse safety signals in the ongoing Phase 2, a statement on its safety primary) → 656 → 3 out after a ninth outside review and the sweeps it prompted, on 31 October: NCT06564844 (TROPION-Lung12: AstraZeneca discontinued it in February 2026 with 25 patients enrolled; the registry still says active), NCT05139849 (VAST-A, taken out by the organiser: its record was last updated in January 2022 and still says recruiting toward 1,276 patients, and the published pilot says enrolment feasibility needs improvement) and NCT04884360 (MONO-OLA1: AstraZeneca's third-quarter results of 30 October say both primary endpoints were met; found by a sweep of the 195 trials of 37 large sponsors through their pipeline documents and third-quarter reports, which found nothing else; it could not read every report: for 61 of the 194 trials some part stayed unread, mostly AbbVie's and Regeneron's third-quarter releases, which would not open, BMS's pipeline listing and the per-trial searches for Novartis and Takeda, and the reports of sponsors that had not yet published, listed with the sponsor and the reason in results_check/quarterly_read/unread_trials.json, which also lists Pfizer's 11 trials, read against its second-quarter release only because its third-quarter report was not yet out) → 653. A second congress sweep had followed on 30 and 31 October (congress_sweep2_instructions.md), of 263 trials: every oncology trial (118; for the 45 run by Chinese sponsors or only in China, three searches each, one in Chinese, aimed at ESMO Asia 2025, CSCO 2025 and 2026, ASCO GI and ESMO GI 2026, ASCO 2026, ESMO 2026, WCLC and ASH; two each for the rest), every neurology trial (17; AAIC, CTAD and AD/PD), and every other trial run by a Chinese sponsor or only in China (128; two searches each, one in Chinese, for results, "met its primary endpoint", topline and filings). Every row names its queries and sources: 217 trials got two searches, 42 got three and 4 got one; 172 had a query in Chinese, and a page was opened for 75. It found nothing more. We also searched every check row for a result credited to another trial in the round (the way LUMINET-1 was missed): LUMINET-1 was the only one. Before that, a first sweep of the 2026 congress abstract books, one search per trial for the 269 trials in oncology or past their primary completion (ESMO 2026 and the other autumn congresses, oncology and China-only trials first; congress_sweep_instructions.md): nothing more found. Every trial left passed the deeper check except the 85 marked below the bar. That check shows each trial was searched to a stated bar (the searches and pages are published); it cannot prove that no result was public. A fresh, independent sample of 60 of the 681 was searched again on 30 October (results_check/audit60/): 58 were confirmed clean, 1 had a disputed statement on a primary and was removed (NCT06474806), and 1 could not be checked and was later left out (NCT07118241). An outside reviewer's own sample of 50 the same day found 2 clear results and 1 probable one (all removed, above). A second outside sample of 50, taken after the native-language sweep, found no result and 1 announced stop (NCT04442022, removed); the strict recheck and the stop check of every trial followed. So some missed results are likely still in the round, a few per cent by these samples; the native-language sweep was run to bring that down. A trial found later to have had its primary result public before launch is never scored. funnel.json lists the NCT numbers at every step, and every check, with its date, sources and the organiser's decisions, is published with the round. The 85 trials below the bar are not a random part of the round: 48 of them (56%) run only in China (every country in the record is China; sisters.json's trial_china_only uses the build's own definition, which counts a few fewer), against 20% of the other 568, because sponsors' pages and filings there were the hardest for the checks to open. A trial found later to have had its primary result public before launch is never scored, below the bar or not; for them the build waives every part of the bar that concerns how deep or recent the check was: as the build reads them today, 44 have no check row that reached the bar at all, 11 carry the checking agent's own doubt, 4 have no page opened, 2 lack a search in the sponsor's language, 1 has too few searches, and 23 now pass the bar. A trial found after launch to have had any public statement on a registered primary before the round opened is never scored: a result on the efficacy primary is caught by its first public date in the labels, and any other statement (a safety statement on a safety primary, say) is put on the organiser's void list (void.jsonl in the round folder, with the reason and the date), which the leaderboard reads; a voided trial is scored on nothing, operations questions included. It does not waive the rest: a finding in any check row, or a safety statement noted on a safety primary, still takes a trial out. A below-bar trial whose latest row says only "unchecked" (no search reached it) and records no finding and no safety statement is kept and marked, by design: the mark tells entrants that no check reached it. The ranking and the tiers use every scored trial, the 85 included. Next to every entry the leaderboard also shows its Brier score, its skill against the registry model with its interval, and the tier it would get without the 85 (without_below_bar, from trial_below_bar in sisters.json), so a reader can see whether they move a ranking; the other scores (calibration and sorting, AUROC, the number questions, the subgroups) are not split, and when fewer than two scored trials are left without them the block says it cannot be computed. 17 of them are already completed on the registry, and they are on the required watch list for 31 October. Seven of the 653 test a supplement, a probiotic or a device, or are run by a contract research organisation, whose reporting habits differ from drug companies'; they stay in and are marked in questions.json (note): NCT05232279, NCT07315516, NCT07457723, NCT07669519, NCT07256743, NCT07193927, NCT07029516.

What to forecast

For every trial whose card asks it (548 trials):

1. p_met: the probability that the trial meets its primary objective, given that a result is public by the deadline (30 June 2027). Only trials that report by the deadline are scored, so this is the question the score asks; trials that report tend to be the ones that worked, and a forecast that knows this is a better forecast. "Met" means the sponsor or investigators report the trial's own primary hypothesis met: a statistically significant improvement, or non-inferiority against the trial's own pre-specified margin when it was designed to show it. If the card's design differs from the trial's own hypothesis, the trial's hypothesis decides; 16 of the 24 non-inferiority cards infer the design because the record does not state it, and say so (design_basis). For the 12 equivalence trials (card design equivalence), it is equivalence shown on the primary outcomes as the protocol defines it. An interim analysis that declares the objective met counts. "Met" follows the rules below. This is the headline question. It measures reported success under the round's labelling rules: a trial whose sponsor, by the deadline, only calls the result positive or encouraging counts as not met (rule 6). That is a rule about what was reported, not proof that the trial failed; the leaderboard shows how many scored trials were only vague, and every entry's score without them.

   p_met_mar and p_met_dec (optional): the same probability, given that a result is public by 31 March 2027, and by 31 December 2027: the early and final passes (see Dates). The three may differ: under rule 6, a trial still only vague at an earlier deadline counts as not met there, which is why the base rate for 31 March is a little below the one for 30 June (see Reference entries). Where an entry leaves either out for a trial, its p_met is used for that pass, and the share of trials where that happened is shown with the entry. The ranked result uses p_met alone.

For the trials whose card says "numbers": "yes" (the number questions), also give:

2. control: what the control arm will report, in the card's control_measure and unit, as your median and an 80% range (low80 and high80 are your 10th and 90th percentiles). Examples: median overall survival in months, the % of placebo patients responding, the mean change on placebo.

3. effect: the effect the trial will report, in the card's effect_measure, as a median and an 80% range in the same way:

   • a hazard ratio (or rate ratio) for time-to-event outcomes, as the ratio itself, for example 0.75;
   • the difference in percentage points, experimental minus control, for responder outcomes;
   • the difference in means, experimental minus control, in the outcome's unit, for continuous outcomes.

   The range is for the number the trial will publish, so it includes the trial's own chance variation.

   Both are asked given that the trial reports after the round closes and publishes the number by the deadline of the pass that scores it (30 June 2027 for the ranking). A number published later is not scored in that pass, and whether and when a number is published can depend on the result, as with p_met. The early and final passes reuse the same answers under their own deadlines and are shown only.

   The 44 cards whose record does not fully state the measure (its unit, time point, scale or form, such as absolute or percent change) say so (measure_basis: assumed), and their notes give the reading the card takes. The number questions refer to that reading; a result published only in another form leaves the numbers void, and p_met is not affected.

Optional, for the phase 3 trials with number questions whose card lists subgroups (279 trials; sex on the 259 that enrol both sexes, and on each one or two more subgroups chosen for the trial on 29 and 30 October):

• subgroups: the same effect within each group of each subgroup the card lists, as {"<subgroup>": {"<group>": {"median", "low80", "high80"}, "<group>": {…}}, …}, with the card's keys, in the card's effect measure. Each subgroup splits the patients into two groups, with the cut-off the field usually reports: a stratification factor named in the record where there is one, else a subgroup the record says it will analyse, else one that results in that disease are usually broken down by (age, region, baseline severity, prior therapy, a biomarker). They were chosen from the record alone by agents that did not see any forecast, and checked by others. Subgroup results usually come with the full presentation or paper, so many will resolve late or not at all; unresolved ones are void. A group resolves only when a publication reports exactly that group. Sex and the subgroups the record gives (a stratification factor, or one it says it will analyse) make the subgroup score; the ones chosen as commonly reported in the disease, most of them, are scored apart and shown as experimental, because a publication may well cut them differently. Shown, never ranked.

Optional, for the 536 superiority and non-inferiority trials:

• safety: the % of patients in the card's experimental arm and in its control arm who stop the study treatment (any part of it, for good) because of an adverse event, as {"experimental": {"median", "low80", "high80"}, "control": {…}}, each from 0 to 100. It resolves on the first source by the deadline of the pass (a press release, congress or paper; not the registry's posted results) that gives the share stopping any study treatment for an adverse event; if none does, the share stopping the experimental drug or its placebo; if none, the share leaving the study for an adverse event. For an arm with nothing to stop (surveillance, no treatment, delayed treatment, or a single dose given at the start), it is the share leaving the study for an adverse event. Void when none of these is published. Shown, never ranked.
• p_key_secondary, for the trials whose card names a key secondary endpoint (the one the record names as key or first in the testing order, else the first efficacy secondary outcome it lists that is not the primary repeated: at another time, in another form (percent instead of absolute change, a count of events instead of the time to the first), by another assessor, in another analysis set or in a subset; a responder analysis of the primary's measure, the share of patients reaching a threshold such as a 50% reduction, counts as an endpoint of its own; chosen on 29 and 30 October and checked by others): the probability that it is met, given that its result is public by the deadline. Met: a source reports it significant in favour of the experimental arm (p below the trial's level, two-sided 0.05 if none is stated, or a 95% interval that excludes no effect) and does not say it was outside formal testing. Not met: any other reported result, including one reported as not significant, as not formally tested (the testing order stopped, the primary was not met, or it was outside the testing plan), or without a test. Void: no result public by the deadline, or a result reported only as immature with a later formal analysis planned. A benefit means superiority, on non-inferiority cards too. Shown, never ranked.

Optional, for the 548 trials with p_met:

• p_next_step: the probability that, after the launch day and by 31 December 2027, the sponsor or a partner takes the next development step for the experimental drug, whatever this trial's result. For a phase 3 trial (phase 2/3 included): a marketing application, new or supplemental, in any country, that rests at least in part on this trial, announced as submitted, accepted for review or approved; the announcement must name this trial among the application's bases, or the sponsor must have said this trial supports it. For a phase 2 trial: such a marketing application, or a phase 3 trial of the drug in the same disease (a condition listed in this trial's record, at any stage or line of therapy) registered in a public trial registry or announced as started after the launch day; a phase 3 registered before the launch day does not count, nor do plans to start. Void when the trial's primary result came out before the close. Resolved after 31 December 2027 by two labellers (next_step_label_instructions.md). Shown, never ranked.

Optional, for any trial:

4. variability: for a continuous outcome, the standard deviation between patients, in the card's unit; for a time-to-event outcome, the number of primary events at the analysis. It is used only for the consistency check below.
5. main_risk: the most likely reason the trial would fail. It is one of "drug too weak", "control arm too strong", "wrong population", "endpoint or timing", "too small or underpowered", "safety or dropouts" or "other". With it, reason: one sentence on that risk, at most 400 characters. Both are published with the entry and never scored.
6. rationale and evidence: why you forecast what you did, in up to 150 words (and 1,500 characters), and up to 5 sources you relied on (NCT numbers, PMIDs, DOIs, links or a trial's name and year, each up to 300 characters). On the Models board a rationale is required for every trial with p_met: the round's instructions ask for it, in up to 100 words to keep runs cheap; the checker allows 150 for everyone. On the Open board it is optional, since an entry may be a statistical model with nothing to say in words. Control, direction-override and zero-width characters are refused. Never scored; published with the entry after the close, as plain text, so readers can see what the forecasts rest on. The organiser may withhold abusive text. assumptions (optional): up to 3 facts the forecast rests on that anyone can check, each as {"claim": …, "source": …} (each up to 300 characters): for example a control arm's result in an earlier trial and that trial's name and year. After the close the organiser's agents check up to 50 claims per entry, drawn with a random seed posted at the close, against their sources, as supported, not supported or not checkable. The share supported is published with the entry, with the counts and a 95% interval; an entry that gave none is shown as "none given". Never scored.

With the operations questions, for every trial with a p_met question (548 trials):

7. p_public: the probability that the trial gets a p_met score: that a public statement on its primary result comes out by the deadline (30 June 2027), in a press release, filing, congress abstract, paper or posted registry results. It counts whether the statement says the primary objective was met or not, or only calls the result positive or encouraging, which counts as not met at the deadline (rule 6); a stop with no public reason counts too (rule 5). It is decided by the labellers from the first pass's labels; it is void when the statement comes out before the close. With p_met it gives the chance of a met result by the deadline: p_met × p_public. p_public_mar and p_public_dec: the same by 31 March 2027 and by 31 December 2027, from the early and final passes' labels; so p_public_mar ≤ p_public ≤ p_public_dec, which the checker enforces.

Operations questions, for every trial (653 trials), answered from the trial's ClinicalTrials.gov record as it stands on 31 August 2027:

8. p_late (late per the registry): the probability that primary completion comes more than 90 days after the date registered on the launch date, as the registry shows it. It counts as late when, on 31 August 2027, the record gives a primary completion date more than 90 days after that date, or the trial is still recruiting (recruiting, not yet recruiting, or enrolling by invitation) more than 90 days after it. It is void when the trial has stopped, and void when the record's primary completion date is then still marked "estimated" although more than 90 days past (the sponsor has not said whether the trial finished). Asked only when the registered date is after the launch date (484 trials for a launch on 1 November), because a date that has passed is no longer a forecast.
9. p_stopped: the probability that the trial's status is terminated, suspended or withdrawn.
10. p_posted: the probability that results are posted on ClinicalTrials.gov. Most trials of the round are not yet legally due to post by the read date (US law asks for results about a year after primary completion), so most answers are expected to be low.
11. p_amended: the probability that between the launch day and 31 August 2027 the record's primary outcomes change in substance. Only these count: a primary outcome measure added or removed (splitting or merging entries without changing the measures does not count); the measured variable or its scale changed; a fixed assessment time changed (week 24 to week 52, say); or the comparison changed (superiority to non-inferiority). A change in the maximum follow-up of a time-to-event outcome, rewording, added detail and changes to secondary outcomes do not count. Judged by two labellers who compare the record as pulled for the round (its primary outcomes are in the packet, fingerprinted in the manifest) with the record fetched on 31 August 2027 by the same means (amend_label_instructions.md; disagreements settled by the organiser); void only when that fetch fails.

The packet ends with the trial's operations questions, its launch-day status and the exact day after which it counts as late.

The answers come in blocks, each all or none on its own:

• An entry that gives number answers gives control and effect for every number question. An entry without them is scored on p_met alone.
• An entry that gives subgroup answers gives them for every trial with subgroups, every subgroup and group the card lists.
• The same for safety, p_key_secondary and p_next_step: each for every trial that asks it, or for none.
• An entry that gives operations answers (p_late, p_stopped, p_posted) gives them for every trial, including the 105 with no p_met question. An entry without them sends no line for those 105 trials, unless it answers p_amended.
• p_public, p_public_mar, p_public_dec and p_amended are blocks of their own: each for every trial that asks it or for none, with or without the others.
• p_met_mar, p_met_dec, rationale (except on the Models board), evidence and assumptions may be given trial by trial.

All answers are JSON numbers: 0.5, not "0.5" or true.

Cards get "numbers": "no" in five cases:

• the outcome's unit is arbitrary or specific to a device or lab, so no forecaster could know its scale;
• the trial has no control arm to compare with;
• the main result is not one number (a win ratio, a hierarchical composite);
• the trial is an equivalence trial, where the question is whether two products match;
• the trial has no efficacy goal to meet (card design none).

Who can enter, and with what

Anyone. No primary result of these trials was found public at launch, so there is nothing to look up, and every entry goes on one of two boards:

• Models board (setting: closed): one model, the round's instructions (forecast_instructions.md) unchanged, the packet and card only, one answer per trial: no web, no tools that bring in information, no other model, no prompt of your own. The trials may be sent one at a time or in batches, and writing the answers to a file is fine. Give the exact model version as the provider names it, with its date or number (an alias such as "latest" is refused), and the provider or API; the page fills in the instructions' SHA-256 when you confirm you used them unchanged. Run settings (reasoning effort, temperature, dates) are welcome. If the model's answer for a trial fails the checker, that trial may be run again with the same model, instructions and settings, at most twice more, alone or with the batch it was sent in (whose answers are then all replaced by the new run's); the number of re-runs is declared in run_settings and shown with the entry. Every trial must be answered by the declared model: a trial the model refuses, or that the provider answers with another model, is run again under the same rule, and an entry still missing a trial after that is incomplete and refused. An entrant may also leave out a whole optional block. No other change to the model's output is allowed. The organiser's own runs follow the same rule and publish their re-run counts. The board rests on what entrants declare: the organiser cannot check how an entry was made, except for its own runs, whose logs it keeps and checks for any web search or fetch. The organiser may re-run 40 trials of any Models board entry placed clearly better, with the declared model and the round's instructions, and publish how closely they agree. The 40 trials are drawn with a random seed posted at the close. The entry counts as reproduced when the mean absolute difference in p_met over the 40 is at most 0.08. The threshold comes from the organiser's own runs on the round's prompt: 56 trials run twice by each of three models before the round opened. Two runs of the same model differed by 0.045 (Opus 5.5), 0.053 (Sonnet 5.5) and 0.062 (Haiku 5.5) on average; a 40-trial sample of the noisiest of the three differs by more than 0.08 about once in 100 draws, the other two almost never. Two different models can also pass: Sonnet 5.5 passed as Opus 5.5 in about half of the draws, Haiku 5.5 as Opus 5.5 in about 6%. So the check is of honesty, not identity: it catches a declared model far from the one used, not every substitution. Entry names, team names and model versions are short plain text (at most 80 characters, no control characters) and are shown escaped, so they cannot carry links or images onto the page.
• Open board (setting: open): anything else. Any model, any number of models, any prompt, tools, retrieval and web access, and people in the loop. Say how the entry was made, shown with it on the leaderboard: retrieval (one of "none", "web search", "own document collection", "web search and own collection"), and true or false for human_in_the_loop (people edited or chose forecasts), fine_tuned (a model was trained or fine-tuned for this) and ensemble (answers of several models or runs combined), with the harness in a paragraph and, if you like, a link to the code. Registry and literature APIs count as web search. retrieval must agree with web_access: the checker refuses "none" with web access, and web search without it.
• Report the average cost per trial in US dollars: compute and API costs, not people's time (say in the harness if people worked on the forecasts). There is no cap; it is shown with the entry as declared, not checked.
• Each Hugging Face account may hold up to three entries, each a different harness. Every entry is listed with the account that uploaded it. If an account ends up with more (for example from uploads at the same moment), its first three entries by first upload count, and the others are listed as left out right after the close.
• Say in the harness description if you work for, or advise, a sponsor of a trial in the round.

Dates

• Opens: 1 November 2026.
• Submissions close: at the end of 15 November 2026, UTC: an upload received at 23:59:59 UTC counts, one at 00:00 UTC on 16 November does not. The window is kept short on purpose: a trial whose result comes out before the close is dropped, and many trials in the round complete their primary outcome in November and December.
• Early pass: 31 March 2027. The trials whose first public result comes after the close and by that date, scored on p_met_mar (or p_met where an entry gave none) and p_public_mar, against B0 and B1 worked out for that date. Its tiers are shown once, when its labels are complete and 100 trials are scored there, as an early look; it does not change the ranking. Fewer trials, and the quickest reporters.
• Deadline for results: 30 June 2027. The ranked result is this first pass: the trials whose first public result comes after the close and on or before the deadline. p_met is asked for exactly this set.
• Final pass: 31 December 2027. The same scores again, counting results up to that date, on p_met_dec (or p_met) and p_public_dec, shown beside the ranked result and never ranked. It covers more trials, including slower reporters, so it shows how far the June result holds; it does not change the ranking. It shows no skill against B0 or B1, whose June forecasts are not matched to its windows (see Reference entries). p_next_step is resolved after this date.
• Trials that report before the close are dropped and listed. This includes a trial found, when it is labelled, to have had a public result before the round opened that the checks before launch missed: its first public date comes before the close, so it is never scored.
• Trials still silent at the deadline of a pass are dropped from that pass. Their number is shown, with the share met among the trials scored next to the base rate, so anyone can see how far the reported trials lean towards success.
• The leaderboard is updated as results come out, at least once a month, with scores and intervals. Tiers are shown only once a pass's labels are complete: for the first pass after 30 June 2027 (the ranking), and for the early pass once, after 31 March 2027 (an early look). Looking at interim tiers every month would multiply the chance of a false "clearly better", so the monthly updates show none.
• p_public: scored once the labels are complete for each pass, after its deadline.
• Operations questions: read from the registry on 31 August 2027, two months after the deadline, so that sponsors have had time to update their records; scored then, p_amended included.

How to submit

Upload forecasts.jsonl on the ReadoutBench submission page (a Hugging Face Space), signed in with a Hugging Face account, before the close. The page asks for the entry's details (team, entry name, contact e-mail, models, harness, board, web access, cost per trial; on the Models board also the exact model version and the provider; on the Open board also retrieval, people in the loop, fine-tuning and ensembles) and checks the file at once: if it fails, it says why and stores nothing. You can run the same check offline first:

tdcb trials round validate --round-dir rounds/2026-r1 --forecasts forecasts.jsonl --meta meta.json

forecasts.jsonl has one line per trial (see template/):

{"nct_id": "NCT06516952", "p_met": 0.71, "p_met_mar": 0.71, "p_met_dec": 0.7, "control": {"median": 18, "low80": 10, "high80": 28}, "effect": {"median": 22, "low80": 9, "high80": 35}, "subgroups": {"sex": {"women": {"median": 22, "low80": 6, "high80": 38}, "men": {"median": 21, "low80": 4, "high80": 38}}, "baseline_iga": {"moderate": {"median": 23, "low80": 7, "high80": 39}, "severe": {"median": 19, "low80": 0, "high80": 38}}, "atopy": {"atopic": {"median": 21, "low80": 3, "high80": 39}, "non_atopic": {"median": 22, "low80": 6, "high80": 38}}}, "safety": {"experimental": {"median": 5, "low80": 2, "high80": 10}, "control": {"median": 3, "low80": 1, "high80": 6}}, "p_key_secondary": 0.75, "p_next_step": 0.55, "main_risk": "control arm too strong", "reason": "One sentence, at most 400 characters, on the main risk.", "p_stopped": 0.01, "p_posted": 0.35, "p_public_mar": 0.3, "p_public": 0.85, "p_public_dec": 0.95, "p_amended": 0.05, "rationale": "Optional on the Open board, required on the Models board for every trial with p_met: up to 150 words on what drives the forecast.", "evidence": ["NCT01234567", "an earlier trial's name, 2023"], "assumptions": [{"claim": "One checkable fact the forecast rests on, in one sentence.", "source": "a trial's name and year, an NCT number or a PMID"}]}
{"nct_id": "NCT05702034", "p_met": 0.55, "control": {"median": 4.5, "low80": 3.2, "high80": 6.0}, "effect": {"median": 0.82, "low80": 0.68, "high80": 0.98}, "variability": 560, "main_risk": "drug too weak", "p_stopped": 0.04, "p_posted": 0.05, "p_public": 0.9, "rationale": "Up to 150 words on what drives the forecast: the record, the drug, the disease and earlier trials."}
{"nct_id": "NCT06927739", "p_met": 0.3, "p_late": 0.4, "p_stopped": 0.06, "p_posted": 0.1, "p_public": 0.6, "rationale": "Up to 150 words on what drives the forecast: the record, the drug, the disease and earlier trials."}
{"nct_id": "NCT06754644", "p_met": 0.9, "p_late": 0.3, "p_stopped": 0.03, "p_posted": 0.05, "p_public": 0.8, "rationale": "Up to 150 words on what drives the forecast: the record, the drug, the disease and earlier trials."}
{"nct_id": "NCT03878849", "p_late": 0.2, "p_stopped": 0.03, "p_posted": 0.15}

The five lines show the five cases:

• The first is a responder outcome. The control is the % of placebo patients responding; the effect is the difference in percentage points. It also shows every optional block: p_met_mar and p_met_dec, subgroups (with the card's keys), safety, p_key_secondary, p_next_step, reason, p_public_mar and p_public_dec, p_amended, evidence and assumptions.
• The second is time-to-event with no median expected. The control is the % of placebo patients with a stroke; the effect is a hazard ratio. Its card also asks subgroups, safety, a key secondary endpoint and the next step, left out here.
• For both, the registered primary completion date has passed, so p_late is not asked.
• The third has no number questions.
• The fourth is an equivalence trial (a biosimilar against Keytruda): p_met is the probability that equivalence is shown.
• The fifth has no efficacy goal to meet, so it has the operations questions only, without p_public. Without operations answers, leave such lines out.
• Every line with p_met carries a rationale, which the Models board requires; on the Open board it is optional.

meta.json (only for the offline check; the page builds it from the form):

{"team": "Your team", "entry_name": "Model X with search and a three-model vote", "contact": "you@example.org",
 "models": ["model-a", "model-b"], "harness": "One paragraph: tools, number of calls, how answers are combined.",
 "setting": "open", "web_access": true, "cost_usd_per_trial": 2.4, "retrieval": "web search",
 "human_in_the_loop": false, "fine_tuned": false, "ensemble": true, "code_url": "optional"}

setting is closed (the Models board) or open (the Open board); an Open board entry adds retrieval, human_in_the_loop, fine_tuned and ensemble. contact is an e-mail address; it is never published. A Models board entry adds model_version, provider and prompt_sha256 (the instructions' SHA-256, from manifest.json), and may add run_settings; template/meta.models_board.example.json shows one.

Your forecasts stay private until the round closes:

• An accepted upload is stored privately. The page's public list of entries shows its Hugging Face account, team, entry name, time received (UTC, set by the server), number of uploads and the SHA-256 of the file. Every upload adds a new row to the public dataset and is never rewritten; the dataset's history records each one.
• You may upload again under the same entry name until the close; the last accepted upload counts. Nobody else can upload to your entry. The file may be at most 12 MB (a full entry with every field filled is a few MB).
• After the close, the organiser registers the last upload of each entry once, and publishes the forecasts and the metadata, without contact, with a fingerprint of the full list of uploads. For an entry uploaded more than once, the first upload is registered and scored too, next to the last and never ranked, so that anyone can see what the extra weeks were worth.

Never publish your forecast file before the close.

Scoring

• Headline: the Brier score of p_met in the first pass, against the registry model. The Brier score is the mean squared difference between the forecast and the outcome (1 = met, 0 = not met); lower is better. Like any proper score it rewards calibration as well as sorting: forecasting the right share of successes counts, and since p_met is asked given a result by the deadline, knowing that trials which report lean towards success is part of the forecast.

• Tiers. Once the first pass's labels are complete and 100 trials are scored, every entry is placed in one of three tiers against the registry model (ref-b1): clearly better (a Brier score lower by more than 0.005), no clear difference, or clearly worse (higher by more than 0.005). Each entry is tested on its own, so no other entry, and no copy of the registry model, can move its tier. Both one-sided tests are fixed in advance for every entry. Their p-values come from a normal approximation with the standard deviation of the paired bootstrap below; the smaller one is doubled, so no direction is picked after seeing the data, and compared with 5%. An entry whose true gap is exactly at the margin is placed clearly better or worse about 2.5% of the time (1.7% to 3.4% in the simulation below), so in a large field a few tiers can be chance; the number of entries tested is shown with each board. With two ranked tracks and up to three entries per account, an account whose entries all sit exactly at the margin on both tracks would see at least one false "clearly better" about 14% of the time (1 − 0.975⁶); entries truly level with the reference are called far less often. "No clear difference" means the data cannot tell the entry from the registry model, not that the two are equal. Each board lists its entries by tier, then by Brier score. The tier is the ranking: the order within a tier, and comparisons between entries (shown as exploratory, with the same test, Holm-corrected over all pairs), are not. The organiser's model runs are on the Models board like any other entry.

  How large a lead "clearly better" takes, from a simulation with the round's sister groups (sim_tiers.py, seed 7, 1,000 simulated rounds per size; its output is published). The entry's errors are about as independent of the registry model's as in the pilots:

  Trials scored	Brier score lower than the registry model's by about, to be placed clearly better half the time	80% of the time	False "clearly better" at the margin (nominal 2.5%)	False "clearly worse" at the margin (nominal 2.5%)
  100	0.044	0.061	2.6%	3.1%
  200	0.033	0.045	1.7%	3.0%
  400	0.025	0.033	1.9%	2.2%
  548 (every trial with p_met)	0.022	0.029	2.7%	2.8%

  (Simulation error: each rate has a standard error of about 0.5 points, so two runs can differ by 1.5 points or more; across this week's rebuilds the same cell has moved from 3.7% to 2.0%. Read the rates as "about 2–3%", not to the decimal.) The error sizes are set, not fitted: the registry model's forecasts err with a standard deviation of 0.9 on the logit scale, half of every error shared within a sister group. With smaller errors, more of them shared (0.6, and 80%), the lead needed half the time is about a third lower (0.030, 0.023, 0.018 and 0.016), with false calls of 1.4% to 3.2%; both runs are published. An entry built on the registry model's own forecasts, changing them only a little, varies far less from it and needs far less: about 0.013, 0.011, 0.009 and 0.009 at the same sizes, half the time. With a registry-model Brier score near 0.16 to 0.22, the first column is a skill of roughly 10% to 30% over it. In a field of 40 entries all exactly at the margin, about one would be placed clearly better by chance; entries truly level with the registry model are placed there far less often. Every pilot model was well short of this, so most entries are likely to land in "no clear difference". Shown with every entry:

  • a 95% bootstrap interval that resamples whole sister-trial groups (sisters.json: trials of the same experimental drug, whoever the sponsor), because such trials tend to succeed or fail together;
  • the skill against the base rate (ref-b0) and against the registry model (ref-b1), each with a paired interval from the same bootstrap. Both baselines' forecasts are published from the opening day, with their training table and script: they are public bars to beat, and copying one earns no more than its score. Both baselines are matched to each trial's own reporting window (see Reference entries), so they already carry the success rate of trials that report in time: beating the registry model takes knowing which trials will work, not only how many. The skill against the registry model is also shown with sponsor-level clusters (a sensitivity check) and separately, with intervals, for China-only trials and the rest, because the registry model leans on that feature;
  • adds to the registry model (exploratory): whether the entry carries information the registry model lacks, even if it loses to it on level. A logistic regression of the outcome on logit(ref-b1) and logit(entry), forecasts kept within 0.01–0.99, is fitted out of fold in 10 folds of whole sister groups (a small ridge, 0.01, on the two weights), and compared with the same regression on logit(ref-b1) alone, so recalibrating the registry model gains nothing by itself. The Brier gain is averaged over 5 random fold splits. Its uncertainty comes from 500 resamples of whole sister groups, each running the same estimator (the same 5 splits, every fold refitted); the 95% interval is the gain plus or minus 1.96 of their standard deviations. "Yes" needs the interval above 0 and a positive weight on the entry: an entry that helps only when turned upside down does not add, and its gain is not shown. Not corrected across entries; the number of entries tested is shown. Never ranked;
  • AUROC (how well the forecasts sort met from not met) and the Brier score's split into reliability (calibration), resolution (sorting) and the level gap (mean forecast minus the share met);
  • for an entry uploaded more than once, its first upload's Brier score and the paired difference, last minus first, with both uploads' setups (models, version, harness, settings). It is an upload-to-upload change: it measures the value of the extra weeks only when the setup stayed the same, and the leaderboard marks it when it did not.

• The base rate and the registry model are shown as baselines, on no board.

• Also shown for p_met: log loss and the calibration slope.

• Early pass: the same tiers, worked out the same way on the early pass's trials against B1 for 31 March 2027, are shown once, when the early pass's labels are complete and 100 trials are scored there, as an early look. They are not the ranking: few trials, and the quickest reporters. An entry that gave no p_met_mar for a trial is scored on its p_met there, and the share of trials where that happened is shown with the entry. The early pass has no number tiers.

• Number score: the second ranking. In the first pass, once its labels are complete and 100 number answers are scored (a control or an effect value counts as one), every entry that gave number answers is placed in a number tier against the frozen reference entry ref-opus-5-5: clearly better (a number score lower by more than 0.05), no clear difference or clearly worse (higher by more than 0.05), with the same test as the p_met tiers, on the paired difference over the resolved number questions, resampling whole sister groups. The reference entry itself gets no number tier. The number tier is shown in its own column; each board stays ordered by the p_met tier.

  How large a lead this takes, from a resampling of Pilot 3's 116 published numbers (64 effects and 52 control arms), scaled the same way, with trials kept whole, for three pilot model runs against Opus: to be placed clearly better half the time, a number score lower than the reference's by about 0.36 to 0.61 at 100 scored answers, 0.27 to 0.44 at 200 and 0.21 to 0.33 at 400 (80% of the time: 0.49 to 0.85, 0.37 to 0.60 and 0.27 to 0.45). An entry exactly at the margin was placed clearly better 0.8% to 3.5% of the time (nominal 2.5%), so the test holds up despite a few very large misses. ref-opus-5-5 both sets the scale and is the comparator; in Pilot 3, a copy of it with its ranges widened gained less than the margin (at most 0.04), and the order of the models was the same whichever model's widths set the scale. The scripts and their output are published. The number score:

  • low80, median and high80 are read as your 10th, 50th and 90th percentiles, and all three are scored. Each resolved answer gets the range's width, plus 10 times any miss, plus 5 times the distance from your median to the result. This is the percentile (pinball) loss, the same as the weighted interval score used by the COVID-19 forecast hubs (Bracher et al. 2021) up to a constant. Ratios and times are scored on the log scale.
  • To compare trials with different units, each score is divided by the width of the reference entry's range for the same question. The reference is ref-opus-5-5, frozen and fingerprinted before the round opens. No width counts as narrower than a floor, so a near-zero reference range cannot blow up one question: for time-to-event and responder questions, the 10th percentile of the reference's widths for that part (control or effect) and endpoint kind; for continuous questions, whose units differ, the 10th percentile of the reference's widths relative to each question's own size (the reference's control median plus its control range), times that question's size.
  • The number score is the plain mean over the resolved number questions, the average of a proper score, so honest ranges do best. The median and the mean with the top and bottom 5% left out are shown beside it, to show how much a few large misses move it; they are not used to compare entries, because trimming would reward ranges narrower than the forecaster believes. Lower is better; a range as wide as the reference's, with the median on the result, scores 1.
  • The share of results that fell inside the 80% ranges is shown too; well-calibrated ranges hold about 80%.

• Subgroup score: shown next to the ranking, not part of it. Each resolved subgroup effect is scored like an effect answer, divided by the width of the reference entry's range for the same group (never less than the effect floor). Sex and the subgroups the record gives make the subgroup score; the commonly reported ones are shown apart, as experimental; sex is also shown on its own.

• Safety score: shown next to the ranking, not part of it. Each resolved share is scored with the same percentile loss, in percentage points (every trial shares the unit, so nothing is divided), with the share inside the 80% ranges.

• Key secondary and next step: the Brier score of p_key_secondary (on the trials whose key secondary result is public by the pass's deadline) and of p_next_step (after 31 December 2027; void for trials that reported before the close), each against the share in the round itself. Shown next to the ranking, not part of it.

• Assumptions: the share of an entry's checked assumptions that its sources support, with the counts. Shown, never scored.

• Consistency: shown, not scored. It measures how far the stated p_met is from the p_met implied by the entry's own control, effect and variability, worked through the card's arm sizes and significance level.

• p_public score: the Brier score of p_public on the trials that did not report before the close, against the share public by the deadline in the round itself; in the early and final passes, of p_public_mar and p_public_dec. Shown next to the ranking, not part of it.

• Met by the deadline: the Brier score of p_met × p_public against "a statement on the primary result after the close and by the deadline, and met", on the same trials, against the share in the round itself, for entries that gave p_public for every trial. It answers a different question from the headline (will this trial deliver a met readout in time?) and is shown next to the ranking, not part of it.

• Operations score: shown next to the ranking, not part of it, so that knowing how sponsors run their trials cannot lift an entry in a ranking about medicine.

  • The Brier score of each operations question, and of p_late, p_stopped and p_posted pooled; p_amended is shown on its own and is not in the pool.
  • The reference is the share in the round itself (for example, the share of trials that ended up late), worked out after the read date. There is no history of registry records to fit a base rate on in advance, so this reference is computed after the fact and says so.
  • Skill is 1 minus the entry's Brier score divided by the reference's.

Labels

A trial's label is set by its first public disclosure of whether the primary objective was met:

• a press release;
• a regulatory or stock-exchange filing;
• a congress abstract or presentation;
• a paper;
• results posted in a public registry.

Labels are made from public sources by the organiser's labelling agents, and every label is checked by a second, independent labeller that does not see any forecast. Disagreements are settled by the organiser, and the agreement rate is published. Labels are published with their sources at each update and corrected in the open; the models used for labelling are named with them.

The rules for "met":

1. The primary objective as the trial defined it:
   • superiority at the trial's own level, or non-inferiority against its own pre-specified margin; when the trial goes on to test superiority, non-inferiority is enough;
   • the trial's own hypothesis decides, also when the card inferred a different design;
   • only the card's primary outcome at its pre-specified level within the trial's own testing order counts;
   • a numerical improvement without significance, or a "clinically meaningful" one without significance, is not met;
   • a win only in a subgroup, a per-protocol set or a post-hoc analysis is not met.
2. In superiority and non-inferiority trials, only efficacy counts. Safety, blood-level and lot-consistency outcomes listed as primary do not. (Equivalence trials: rule 9.)
3. Several efficacy primary outcomes:
   • in trials with parts, only the randomized confirmatory part counts;
   • co-primary outcomes must all be met by the same arm, unless they split the significance level, in which case either one counts;
   • otherwise the sponsor's own statement of whether the primary endpoint was met decides.
4. Several doses or experimental arms: one arm meeting the objective against the registered primary comparison is enough.
5. Interim analyses and early stops:
   • an interim analysis that declares the primary objective met is met; one that neither declares it met nor stops the trial leaves the question open;
   • stopped for futility or harm is not met;
   • stopped with a public reason that is not about efficacy (business, slow enrolment, a supply problem) and no efficacy result is dropped and listed;
   • stopped after an unblinded interim analysis with no public reason, or "after reviewing the data", is not met.
6. "Positive" or "encouraging" with no statement on the primary endpoint keeps the question open. If nothing clearer comes by the deadline of the pass, it is not met in that pass. This is the round's choice, on the view that a sponsor which talks about its trial without claiming the primary endpoint has more often missed it; it is an assumption, not something measured in this cohort, so these trials are marked and every score is also shown without them. (A trial vague in June and clear by December can therefore differ between the two passes; the ranked pass uses June.)
7. Labels are judged against the primary outcomes in the round's packet. If the sponsor reports a different endpoint as primary and never reports the registered one, it is not met.
8. Sister trials are labelled one by one. A pooled-only result counts as vague (rule 6).
9. Equivalence trials: met when the sponsor reports that equivalence (bioequivalence, biosimilarity, comparability or lot consistency) was shown on the primary outcomes, within the protocol's margins (no record in the round states them). The blood-level, efficacy and immune-response comparisons listed as primary all count; safety outcomes do not. When PK and efficacy are both primary, both must succeed.

p_public labels. From the first pass's labelling, after the deadline: 1 when the first public statement on the primary result (a vague one or a stop with no public reason included) came after the close and by 30 June 2027; 0 when none was public by then; void when it came before the close, or the trial left the round (a stop for a reason not about efficacy, rule 5) by the deadline; a trial that leaves the round after a pass's deadline counts as 0 in that pass. The same for p_public_mar and p_public_dec, each from its own pass's labels; the met-by-the-deadline score of each pass uses that pass's two fields. A pass is re-scored when its labels are corrected, and the change is listed.

Dates. Labellers record two dates: the first public statement on the primary result, vague ones included, and the first clear one (whether the objective was met). The first decides whether a trial reported before the close (dropped) or after a pass's deadline (not scored in it). The second decides rule 6: a trial that was only vague at a pass's deadline counts as not met in that pass, even if it became clear later, and is marked as such; so is a pooled-only result (rule 8). A stop with no public reason is not met (rule 5) but is not marked: it is a stop, not a vague statement. A met, not-met or changed-endpoint label needs the clear date. Numbers count in a pass only if published by its deadline. When the two labellers' dates fall on different sides of the close or of a deadline, or a date is malformed, the trial goes to the organiser, so that one labeller's date alone never decides whether a trial is dropped, scored or late.

Operations labels. Read automatically from each trial's record, fetched from the ClinicalTrials.gov API on 31 August 2027, against the launch-day row in operations.jsonl (tdcb trials round ops-resolve). No judgement is involved, except for p_amended: two labellers compare the primary outcomes of the record as pulled for the round with the record fetched on 31 August 2027, against the closed list of changes above (amend_label_instructions.md); disagreements go to the organiser.

Key secondary labels. By the deadline of the pass, under the rule given with the question: met when a source reports it significant in favour of the card's experimental arm and does not say it was outside formal testing; not met for any other reported result (not significant, not formally tested, or reported without a test); void when no result is public, when it is reported only as immature with a later formal analysis planned, or when the trial left the round.

Safety labels. For each of the card's arms, by the deadline of the pass, from a press release, congress or paper (not the registry's posted results): the first published share stopping any study treatment for an adverse event (the safety population); if none, the share stopping the experimental drug or its placebo; if none, the share leaving the study for an adverse event; for an arm with nothing to stop, the share leaving the study for an adverse event. Each arm resolves on its own; void when none of these is published.

Next-step labels. After 31 December 2027, by two labellers searching public sources and trial registries (next_step_label_instructions.md): 1 when the card's next step, as defined with the question, was announced after the launch day and by 31 December 2027, else 0; a statement linking an application to the trial counts only if it too is public by 31 December 2027. Void for trials that reported before the close or left the round by 31 December 2027. Scored once, with the last pass's labels (December 2027).

Number labels.

• A number question resolves with the first public disclosure, by the deadline of the pass, that gives the number for the card's endpoint, time point, arms and comparison.
• A number for another time point, dose or population does not count, and the question stays void unless the right one comes out.
• A least-squares or adjusted mean difference counts as the difference in means. For a responder card, a result given only as an odds or risk ratio is converted to a difference in percentage points when both arms' rates are published; otherwise the effect question is void.
• When the effect is published without the control arm's value, only the control answer is void.
• Unpublished numbers make the question void; the p_met question still counts.
• A subgroup question resolves with the first public effect estimate within each group, as the card defines the groups, for the card's endpoint and comparison (a forest plot counts), by the deadline of the pass. A group resolves only when the publication reports exactly that group; a group the publication cuts differently (70 years instead of 65, or three bands instead of two) is void. Each group resolves on its own.

Fair play

No inside information about a trial's result or its timelines, and no use of other entries' forecasts. The organiser may ask the top entries for their code or logs and may remove an entry that breaks these rules. A team holds at most three entries in all, whatever the number of accounts it uploads from; the organiser removes the extra ones, keeping the team's three earliest. After the close, the organiser also looks for entries that are confident and right far more often on one sponsor's trials than on the rest.

Reference entries

The organiser registers these before the round opens and posts their SHA-256s, with the manifest's, outside the submission page. The base rate and the registry model are published in full on the opening day; the model entries stay private until the close:

• the base rate (B0) and a model on seven registry fields (B1), p_met and p_met_mar only, fitted on earlier company trials. Both are published with their training table, the label audit and the script (make_baselines.py), so anyone can refit them byte for byte.
  • Training trials. 218 earlier company trials, all completed in 2026, whose primary result was public before the round opened and came at a time after primary completion that the round can score (between 3.7 months before and 12 months after; 93 of them are Pilot 2's trials). Stops with no result are left out, since the check did not record their reason; the round counts a stop with no public reason as not met, so B1 may run slightly high on trials likely to stop.
  • Labels, read as the round reads them. Each training trial keeps the date of its first public statement and the date its result became clear. For a window ending b months after primary completion, a training trial counts as met only if its result was clearly met by then; a result still vague, reported on another endpoint than the registered one, or clear only later counts as not met (rules 6 and 7). A training trial counts for a window only if its first result came within it and it was followed for all of it: its primary completion plus b is on or before the check date, 25 October 2026. The others are left out, whatever their label: early results lean met, so keeping the clear ones and leaving out only the vague ones would push the rate up.
  • Checks of the labels. The 54 training labels whose first result came more than a month before completion were checked by web search: all but one were the trial's own readout (often an interim win, with the completion date moved later), 13 got an earlier first-disclosure date, 3 are safety stops (not met, rule 5), and 1 could not be confirmed and is left out; the trials that reported years early are outside the scored lags and not used. Since the audit moved 13 of 54 dates earlier (a press release before the paper), some other training dates may also be late; this mainly blurs the window matching. All 23 labels first called vague or changed endpoint were checked again by two independent labellers on 28 October, without seeing the first call. They agreed on 18 of the 23: 7 clear results public before the check date (6 met, 1 not met; corrected), 9 still vague, and 2 where neither found anything. The organiser decided the other 7 (the 5 disagreements, and the 2 where neither labeller found anything but the first check had found a vague statement) from the sources the labellers and the first check cited, opened again: 1 met (a sponsor's statement that only one labeller had found), 4 vague, and 2 left out (an interim report with no primary result, and a stop). Both re-checks and every decision, with its reasons and source, are published (baselines_relabel_check_a.jsonl, _b.jsonl and baselines_label_audit.jsonl).
  • B0. Each round trial is scored only if its result comes after the close and by 30 June 2027, a window that depends on its primary completion date. B0 is the share met at the window's end among the training trials that count for that same window after their own primary completion, shrunk towards 63.8% (the share met in B1's fit) with a weight of 20 trials. It runs from 0.58 to 0.72. The round trials that completed earliest have the longest windows, and fewer training trials were followed that long: for the 65 that completed by early September 2026 (windows ending 10–12 months out) there are none, so their B0 is 63.8%; 109 round trials have fewer than 20 behind them, and the median round trial has 105.
  • B1. B1 moves B0, on the logit scale, by a logistic regression on phase 3, placebo control, enrolment, cancer, large company, China only and time-to-event endpoint (C = 0.1), and is kept within 0.05–0.95. The regression is fitted once, on the median round trial's window (from 0.2 months before to 6.5 months after primary completion): 105 training trials, 63.8% met. One fit for every window is an approximation. Fitted on the window of the round trial at the 25th or the 75th percentile instead, B1's mean over the round would be 0.60 or 0.67 (it is 0.63); with the 7 organiser decisions left out of training, 0.66 (baselines_info.json, sensitivity).
  • The early and final passes. For the early pass, each round trial's window ends on 31 March 2027 instead, and B0 and B1 are worked out for it the same way, with the same fit (B1's regression is not refitted); they are p_met_mar. Every round trial has training trials followed for its whole early window. B0 runs from 0.55 to 0.69 for 31 March (mean 0.63, against 0.64 for 30 June), and B1's mean is 0.61 (0.63 for 30 June). For the final pass, no training trial was followed long enough (windows ending up to 18 months out), so B0 and B1 have no p_met_dec; the final pass shows them with their June forecasts but no entry's skill against them.
  • China-only trials. None of the 40 China-only training trials was a clear failure: 34 were met and 6 were only vague. In B1's fit, 13 were met and 2 count as not met, both only vague. B1 leans on that feature (0.76 on China-only trials, 0.58 on the rest; 0.85 on China-only trials if the 7 organiser decisions are left out), and its China-only effect may measure how often such sponsors report vaguely as much as how often they succeed. Hence the split shown next to its skill.
• Models board runs by Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, under the Models board's rules (the round's instructions unchanged, the packet and card only, no web), with every question, the operations questions included. They are placed in tiers like any Models board entry (ref-opus-5-5 gets no number tier, being the number score's reference). A batch with an answer the checker refused, or whose log showed a tool other than reading the instructions and the batch and writing the answers, was run again whole, at most twice, as the Models board allows; the re-run counts and reasons are in their metadata (a first log check also refused runs that only tidied their own answer file, since allowed). One trial, NCT07385287 (a live attenuated malaria-parasite vaccine), was taken out of the round before it opened: a safety filter stopped Opus 5.5 on it in every run and the provider answered with another model, so no complete Opus entry could include it. Taking the trial out, rather than the Opus entry, is a selection: the round no longer holds the one trial the organiser's reference model would not answer. An outside entry whose model refuses a trial of the round cannot be completed and is refused; tell the organiser before the close if that happens, and it will be listed. Their exact model versions, run settings, run dates and the instructions' SHA-256 are in their metadata.

ref-opus-5-5 is also the reference that scales the number score.

The model runs are closed book, so running them later would change nothing: they see the same packet with the same model. Entries with web access can use news that comes out while the round is open; the leaderboard therefore shows every entry's board, web access and last upload date next to its score, and scores first uploads too.

Research use only. Nothing here is medical advice.