Our methodology

Every generated file is validated against a strict schema. If anything produced in a run fails, the run fails and no pull request opens.

Automated rules catch known model failure modes before publication — most importantly, a PDUFA or approval event is rejected unless the asset is actually at a filed or approved stage, because models invent FDA decision dates for phase 1–2 programs.

Generated content never publishes automatically. It opens a pull request labelled automated-content with a review checklist, and a human merges it or does not.

Before a calendar publishes, every trial-readout entry is searched against ClinicalTrials.gov. A match is only accepted when the drug, sponsor, phase, indication and timing line up and no rival study is close behind — otherwise the entry publishes as unverified rather than being guessed at.

Every guard on this page is executed on each push and pull request, alongside the type checker, the linter, the build, the crawlability invariants, and a ratchet that fails if the number of content files breaking their schema goes up. Before September 2026 none of them ran automatically, which made them advisory — a guard nobody runs is a comment. Separately, a scheduled job refreshes the site from ClinicalTrials.gov, SEC EDGAR and openFDA; it checks it can reach those hosts before it starts, so a run that changed nothing means nothing needed changing rather than that nothing was checked. No step that reads a primary source is allowed to continue past a failure, and a test enforces that.

Company profiles used to carry financial figures written as prose by a language model ("Healthy balance sheet backed by…", a runway picked from a menu of round numbers). Those fields have been removed and the schema now rejects them. Each company page opens with cash and investments, quarterly operating burn, runway and shares outstanding read from the company’s XBRL filings, each with its date and a link to the filings. Where no filing has been parsed, or a parsed figure failed a consistency check, the page shows no number and says why.

A trial readout can be checked against ClinicalTrials.gov. An FDA action date cannot be checked against anything except the 8-K or press release in which the sponsor announced it, because the FDA publishes no feed of pending decisions. The search for that filing is scoped to the sponsor and to the two years before the date, and a filing naming a decision that already happened is treated as a different submission rather than as evidence the date moved. Every decision date we could not trace to that filing has been taken off the calendars — not deleted, but held in the open in src/data/pdufa-leads.json, and returned the moment the filing is found. The resolver treats a filing that states a different date as news in its own right, because a PDUFA date that moved is exactly what a reader needs to know.

The PDUFA calendar no longer starts from dates someone expects and then looks for evidence. It starts from the evidence: SEC full-text search for 8-Ks that state a PDUFA or target action date. The date is read from the filing text with the sentence that states it, and that sentence is quoted on the decision’s page with a link to the filing. A date earlier than its filing describes a decision already made and is ignored; a filing that names no drug near the date is held for a person rather than guessed; a later filing with a new date replaces the old one and keeps it on record. After the date, the outcome is taken from Drugs@FDA or the sponsor’s own 8-K, and until one is found the decision says so.

Every pending FDA decision on the calendar is checked against Drugs@FDA and the approved label before it publishes, and an entry that presents a first approval the FDA has already granted is dropped. The check is deliberately narrow: a label expansion, a new indication, a move to an earlier line of therapy and a confirmatory study are all normal entries for an approved drug, and the obvious version of this rule flagged ten of them. If openFDA is unreachable, nothing is screened and the run says so rather than reporting a clean pass.

Each calendar’s search description is checked against the entries in the file. Five of the six promised FDA advisory committee meetings and carried none, and four claimed to be updated daily; both claims are gone and a test now compares the promise against the contents. Advisory committee dates have no machine-readable source, so any that ship must carry a disclosure URL a person read.

Where we state a probability of success, it starts from a published phase-transition rate with the source linked, and the factors that push against it are named with their direction. We do not multiply invented weights into a single percentage. The anchor is the BIO, Informa and QLS study of 12,728 phase transitions over 2011–2020: 52.0% of Phase I programmes advance, 28.9% of Phase II, 57.8% of Phase III. Those figures were read off the report itself on 19 September 2026, and the file records the size and SHA-256 of the exact PDF so you can re-download it and check that we read the same document. Any rate we have not opened the source for still shows as a gap rather than a recalled number.

Where we hold one, a company page shows the sponsor’s full Phase 2 and 3 registry record — how many studies completed, how many stopped early, how many are running — with each stopped study linked. It is reported as counts with the denominator visible, never as a sponsor grade: trials stop for many reasons and the registry often does not say which.

Cash runway is calculated from the disclosed cash position and quarterly burn and displayed with that arithmetic beside it. Where either input is prose rather than a figure, no runway number is shown. Where a stored figure disagrees with its own inputs, the discrepancy is shown. Commercial-stage companies get no runway at all: a company funding itself from product revenue is not counting down to a raise.

A company profile is maintained only when we can say something specific and current about it — an upcoming tracked catalyst or a derivable cash runway. Holding the sponsor’s registry record used to count too, until a scheduled job began building one for nearly every sponsor: a claim about a build artifact is not a claim that anyone keeps the page current. Everything else is archived: the page stays live and readable, labelled as no longer reviewed, and excluded from our published profile count. The tiering comes from criteria in code, not a hand-picked list, and can be recomputed.

Each content type has a review window set by what it depends on — 45 days for a catalyst calendar because guided dates move, 120 for a company profile because companies report quarterly, a year for an educational guide. Under every company and trial title, one line gives the date of the data (the filing period, for company figures) and when the text was last updated, with its age in days once it is past its window. Content with no recorded review date says so rather than looking fresh.

Every dated catalyst call we publish is listed at /accuracy with its outcome — hit, slipped, wrong, or not yet checked. An outcome is recorded only against a source you can open; we do not score our own calls from memory. Calls that have come due without being checked are published as a visible backlog rather than quietly ageing out of view.

FDA must announce every advisory committee meeting in the Federal Register. The meetings listed on the PDUFA calendar are read from those notices through the Federal Register API: the meeting days from the notice’s dates section (not the comment deadline beside them), the committee from its title, and the drugs and application numbers from its agenda. Each links to its notice, and a meeting about a drug the decision tracker follows is shown on that decision’s page. A committee recommends; the FDA decides.

Where a runway is derivable, we compare it against the catalyst date and say whether the cash reaches the readout. Runway is counted from the date the cash was reported, not from today, and a coarse catalyst window is read from its earliest point so a wide window cannot manufacture a warning. It states the arithmetic between two dates and does not predict what a company will do about it.

Each catalyst is graded company-confirmed, registry-confirmed, or unverified, and links out to the record you would use to check it.