Trial Run

How this is built

Methodology

Every number here is computed from public data by a deterministic pipeline. Same snapshot in, same numbers out. This page is the contract. Open any section.

SourcesClinicalTrials.gov v2 + registry history
SourceUsed forRows
ClinicalTrials.gov API v2Interventional studies, field-projected. Same registry data AACT mirrors.454,031
ClinicalTrials.gov registry version historyPer-study module change history; used for the eligibility-amendment proxy on a sample.8,919

Snapshot as of 2026-07-21. The registry data is the same corpus AACT mirrors, pulled directly from the official ClinicalTrials.gov v2 API.

The pain indexthree signals, ranked and averaged

Each component is percentile-ranked across all populated cells (higher rank = more pain); a cell's index is the equal-weighted mean of whichever components it has, and is null below 2 present components.

Three signals, equal-weighted. A cell needs at least 2 present to score; otherwise it reads insufficient, never zero.

  • Eligibility amendments. Share of sampled trials whose registry history shows a change to the Eligibility module. A disclosed proxy for protocol amendments, following AMEND++ (arXiv 2601.06300), recomputed here for transparent coverage.
  • Enrollment failure. Of trials that concluded, the share terminated or withdrawn rather than completed. Ongoing trials excluded.
  • Startup lag. Median days from registry submission to study start, over trials registered before they began.
Value at stakethe amendment proxy, priced conservatively

Where a cell has an amendment proxy, it is priced in dollars. The anchor is deliberately conservative: Tufts CSDD puts a substantial amendment near $141K in Phase II and $535K in Phase III, averaging about $454K, so this uses $450K. Delay, six or more weeks per amendment, is not priced in, which makes every figure a floor.

The math is simple: of 100 industry trials, the proxy rate is how many amend at least once; times the anchor gives the dollars at stake per 100 trials, scaled to the cell. The sensitivity row shows a 10, 30 or 50 percent reduction. It counts only trials with at least one eligibility change, so it understates the true exposure.

Anchor: Tufts CSDD, New Benchmarks on Protocol Amendment Practices (2024).

External benchmarkshow this sits next to industry-cited figures

This map measures the amendment tax from the public registry independently. For context: Biorce cites research that 85 to 87 percent of trials are amended, roughly 45 percent from human error and 40 percent from poor design. Those are company-cited, not verified here, and describe amendments broadly, not the eligibility-only proxy computed here. Quoted only to show this looks at the same tax from a disclosed angle.

Thresholdsminimum samples before a figure shows

A cell must clear a minimum sample before a figure is reported, so tiny denominators do not drive the map.

Minimum concluded trials for a failure rate20
Minimum trials for a startup median15
Minimum sampled histories for an amendment rate15
Limits, stated in the openwhat this does not claim
Limit 1The amendment figure is a PROXY. It is the share of sampled trials whose ClinicalTrials.gov registry history shows at least one change to the Eligibility module. A trial can amend its protocol without changing registry eligibility text, so this understates total amendments; it is a reproducible, disclosed signal, not a formal amendment count.
Limit 2Amendment rates are computed on a stratified per-cell sample of registry histories, not the full population. The sample size is shown as coverage next to every amendment figure.
Limit 3Contract research organization (CRO) involvement is not reliably recorded in public registry fields, so this instrument does not report CRO concentration. It deliberately makes no claim it cannot evidence.
Limit 4Sponsors register under several legal-entity names. Unambiguous name variants of one company are folded together for the account ranking (for example Novartis Pharmaceuticals and Novartis), via a curated, disclosed map. Parent-subsidiary and acquisition relationships are NOT merged, since those are strategic groupings rather than spelling variants.
Limit 5Therapeutic areas are assigned by a deterministic keyword heuristic over sponsor-entered condition text, not a clinical ontology. Condition names the keywords could not place were then sorted once by an AI classification into the same buckets, frozen and committed and read as a lookup table rather than a live model. Pain and anesthesiology and healthy-volunteer studies are broken out as their own rows; the remaining General and cross-cutting row holds sparse areas, cross-cutting quality-of-life and device studies, and single-trial rare conditions, which are kept there rather than forced into a disease area.
Limit 6The stated-reason categories come from a deterministic keyword pass first. Free-text reasons the keywords could not place were sorted once by an AI classification into the same fixed categories; that output is frozen and committed, and the build reads it as a lookup table, so it is a disclosed, auditable supplement rather than a live model in the pipeline. Text that fit no category, or only pointed to an unavailable document, stays in Other.
Limit 7"No enrollment recorded (inferred)" is not a sponsor-stated reason. It is inferred from a zero enrollment count on a stopped trial, which means the trial never enrolled a participant, and it is labeled as inferred so it is never counted as a stated reason.
Limit 8This instrument maps market-level pain from public data. It does not model or predict any individual protocol, which is the work of clinical-trial design tools, not of this analysis.
Reproducibilitysame snapshot in, same numbers out

Deterministic: no wall-clock reads in the transforms, the snapshot date is an explicit input, and every figure carries its numerator, denominator and exclusions. Same snapshot in, same numbers out. Therapeutic-area and stop-reason labels use an ordered keyword pass plus a frozen, committed AI classification for the residual; the keyword pass stays authoritative.

Amendment-proxy methodology follows AMEND++ (arXiv 2601.06300) in using eligibility-criteria version changes as an amendment signal, computed here directly from the official registry history endpoint.