Industries / Health / Pharmaceuticals
Clinical trials
Testing a candidate medicine in people, in planned studies, to learn whether it is safe and whether it works.
6
Milestones achieved, from 1747 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
In depth
The same ground is charted in more depth, as an industry of its own.
The road
AchievedWatch forOpen
Achieved
1747
Six scurvy treatments compared
James Lind, a Scottish naval surgeon unsure which of many proposed scurvy treatments to use, compared six of them in a planned, controlled test. He published the work in 1753.
Credit
James Lind
1948
Keeping assignments hidden
The UK Medical Research Council's trial of streptomycin for lung tuberculosis is a landmark in method, the James Lind Library says, because its report described the steps taken so that no one could know in advance which treatment the next patient would get.
Credit
UK Medical Research Council
1964
Ethics for research on people
The World Medical Association adopted the Declaration of Helsinki, its ethical principles for medical research on human participants, written in response to unethical research by physicians in the Second World War. It has been amended seven times, most recently in October 2024.
Credit
World Medical Association
1997
A public register of trials
The US Food and Drug Administration Modernization Act of 1997 required the National Institutes of Health to set up a public registry of clinical trials, whether funded by government or by companies. It became ClinicalTrials.gov.
Credit
US Congress; the National Library of Medicine at NIH
2022
Smaller trials, regulator-backed
EMA's main scientific committee (CHMP) qualified PROCOVA, a method that predicts each participant's outcome on placebo with a model trained on past trial data, and uses that prediction to sharpen the estimate of a treatment's effect. EMA said this could raise a trial's power and, under stated conditions, reduce the number of participants needed in Phase 2 and 3 trials with continuous outcomes.
Credit
Unlearn, the company that developed the method; EMA's CHMP, which adopted the opinion on 15 September 2022
2025
Trial rules rewritten
The International Council for Harmonisation (ICH), where regulators and industry agree shared rules, rewrote its good clinical practice guideline to cover newer trial designs and technology; the new version, E6(R3), took effect in the EU on 23 July 2025. In November 2025 ICH adopted M11, a common structured template for trial protocols, the documents that set out how a trial will be run. M11 took effect in the EU on 11 June 2026.
Credit
ICH; EMA's CHMP
Watch for
Forecasting which trials succeed
Bristol Myers Squibb (BMS) uses Claude to assess the plans for trials of an early-stage drug and to estimate how likely the trials are to succeed. A BMS executive said a first analysis came from a single prompt and "would otherwise take days".
Achieved whenforecasts made before results are published and then scored against the actual results.
Claimed by
Anthropic and Bristol Myers Squibb, October 2026
Open
01
Score trial forecasts in public
A public log of probability forecasts for Phase 2 and Phase 3 trials due to finish soon, each posted before results are known. Each forecast is the chance that the trial meets its main goal. After results come out, each forecast is scored with the Brier score, the average squared gap between forecast and outcome, where lower is better. Forecasts only: no betting or trading.
The briefCredit · unclaimed02
Old protocols into structured data
A converter that turns a trial protocol PDF into the ICH M11 structured format, plus a public answer key of protocols converted by hand to score it.
The briefCredit · unclaimed
Open 01 · the brief
Score trial forecasts in public
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public log of probability forecasts for Phase 2 and Phase 3 trials due to finish soon, each posted before results are known. Each forecast is the chance that the trial meets its main goal. After results come out, each forecast is scored with the Brier score, the average squared gap between forecast and outcome, where lower is better. Forecasts only: no betting or trading.
Why it is still open
Companies now make such forecasts with AI models in private. The published account of BMS's work with Claude gives no scored track record.
Built with what exists
the ClinicalTrials.gov API to list trials by expected completion date; company results announcements; Claude, or Gemini Decision 1, which returns probabilities for fixed-answer questions in place of text; a timestamped public record such as a Git repository. The note "A thousand questions, one pass" covers Gemini Decision 1.
First step
List trials with a primary completion date in the next three months and post the first 20 forecasts with timestamps.
Done looks like
At least 100 forecasts posted before their results, at least 50 of them resolved, and a Brier score reported beside that of a simple forecaster that uses past success rates for each phase.
Open 02 · the brief
Old protocols into structured data
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A converter that turns a trial protocol PDF into the ICH M11 structured format, plus a public answer key of protocols converted by hand to score it.
Why it is still open
M11 took effect in the EU on 11 June 2026, but protocols written before it are documents, not data. ClinicalTrials.gov held 53,284 study records with a protocol document attached when I queried it on 17 November 2026. I found no published test of how accurately software converts such protocols.
Built with what exists
the ICH M11 template and technical specification; the HL7 FHIR Clinical Study Protocol guide, which carries all of M11's structured elements; the ClinicalTrials.gov API; a language model that reads long PDFs, such as Claude or Gemini; a schema validator.
First step
Download 10 protocol PDFs from ClinicalTrials.gov and fill the M11 template by hand for two of them.
Done looks like
50 protocols converted by two people working separately, with disagreements settled, form the key. The converter fills at least 95% of required M11 fields correctly across the 50, and every output passes validation.