Universe. Interventional studies with an industry lead sponsor whose primary-completion date falls in calendar 2026–2028, Phase 1 through Phase 3, as registered on ClinicalTrials.gov. The default window is the current month through Dec 2028, the whole registered universe; narrow it in the rail. Registered dates lag real readouts: open-label early-phase programs routinely read out interim data a year or more before the registered date. Dates are sponsor-estimated unless flagged ACTUAL; month-precision dates are placed on the 15th.

Trial health. Enrollment per month is registered enrollment divided by months from study start to primary completion. That span includes follow-up, so it is a schedule-density proxy, not observed recruitment velocity; a low value flags a record to check, not a recruitment problem. Site density is sites per 100 patients. Both are compared with the median of trials in the same area and phase in this universe. A study below half its peer median loses six confidence points unless it is completed or reports an actual date.

rNPV sketch is a deliberately simple model — linear ramp, flat plateau, three-year erosion after exclusivity, single discount rate, one cost-to-complete number — pre-filled from the registry, the literature prior and SEC cover-page share counts. It exists so a number can be stress-tested in the dossier, not to replace a proper model.

Readout confidence is a transparent heuristic, not a model fit: it rewards actual (vs. estimated) completion dates, completed or active-not-recruiting status, actual enrollment counts and recent registry updates, and penalises not-yet-recruiting status, stale records and stopped studies. Use it to triage, not to price.

Flags. Each trial carries the first rule that fired for it, shown with its trigger: stopped → Stopped; cash months < months to primary completion → Cash flag (cash binding; the row says when the test is not applied: large pharma, positive operating cash flow, or no SEC financials); completion date passed, estimated inside 90 days, or slipped → Date passed / Date unconfirmed / Date slipped; ≥2 specialists new last quarter → Specialist adds; an insider buy cluster inside 180 days → Insider buys; enrollment density below half the peer median → Enrollment lag; Phase 3 with an actual registry date and confidence ≥75 → Pivotal, actual date; other Phase 3 inside 6 months → Pivotal inside 6 mo; else No flag. These are descriptive labels for triage; none is a position recommendation, and their false-positive rates are not yet measured (see the roadmap’s evaluation phase). Date reliability is firm (the registry reports primary completion as ACTUAL, i.e. data collection for the primary endpoint has ended; it says nothing about when results are announced), estimated (registry estimate), or slipped (primary completion pushed out by a month or more since the previous weekly snapshot, unless the move was off a June/December placeholder or onto an actual date — that is a firming). A slipped date is a Wait: the registry just said the plan changed. Impact = phase weight × (0.5 + literature prior) × size factor (log public float) × firmness × crowding discount, bucketed High ≥0.62 / Med ≥0.35 / Low. It ranks how much attention a readout deserves in an investable universe: the size factor rewards a liquid, listed sponsor and is not a claim about how far the stock would move — a pivotal readout at a micro-cap can matter far more to that equity. Cash-to-event uses the runway above: cash plus every investment tag the filer uses (current and non-current marketable securities, available-for-sale and held-to-maturity debt), from the company’s own XBRL facts at one balance-sheet date, over operating cash outflow annualised from the latest year-to-date figure in the same filing; the fiscal-year figure and then the annualised net loss are fallbacks, each labelled in the dossier. Debt is shown beside cash and is not netted from it; a sponsor whose reported operating cash flow (or, failing that, net income) is positive is treated as unconstrained, and missing inputs are shown as unknown rather than zero. All of it is rules, none of it is advice.

Snapshot diff. Each weekly build is compared with the previous one, trial by trial: primary-completion date (with the size of the move in months), estimated → actual, status, enrollment and why-stopped. Slipped, Date firmed and New this snapshot are filterable flags in the rail; the dossier lists the raw before → after values, and each build's diff is kept under data/history. Registry dates lag real readouts, so a single slip mostly confirms what the market has priced; the second slip is the signal.

Base-rate priors are phase-transition probabilities from Wong, Siah & Lo (2019, Biostatistics), oncology vs. non-oncology, shown for context only. The prior column is the single next-step transition (e.g. Phase 2 → Phase 3); the rNPV sketch is pre-filled with the cumulative probability of approval from the current phase, the product of the remaining transitions, which is a different and smaller number for early-phase programs. Both are population averages and say nothing about a specific asset.

Therapeutic area and modality are assigned by pattern-matching titles, conditions, keywords and intervention names. A registered intervention type of DRUG with no recognisable name or class stays Drug (unclassified), about half the universe, rather than being called a small molecule. Expect a few percent mis-assignments among the classified ones; the dossier shows the raw fields so you can check.

Sponsor tier. "Large pharma" is a fixed list of ~25 groups with subsidiaries mapped (e.g. Genentech → Roche, Janssen → J&J). "China-domiciled" is inferred from corporate suffixes and city names in the registered sponsor string.

FDA actions come from the openFDA Drugs@FDA endpoint. It covers approval actions on NDAs/BLAs/ANDAs, not target dates, which the FDA does not publish.

Regulatory calendar. PDUFA target dates, advisory-committee meetings and planned resubmissions are read from the companies' own SEC filings: EDGAR full-text search over 8-K, 6-K, 10-Q and 10-K documents and their press-release exhibits for the trailing 24 months, refreshed weekly on the trailing three weeks. Each date is classified from its own sentence, carries day, month or quarter precision (≈ marks a period-end placeholder), and cites the filing it came from; when a later filing states a different date the earlier statement is kept as history and the row is marked moved. Confidence is firm for an exact date in a filing under six months old and guided otherwise. Filers are resolved to tickers by CIK. Coverage trails a paid calendar for companies that only say the date on a call or a slide, and a date inside an image exhibit is invisible to text search. Market cap is the last close times cover-page shares outstanding where a quote is available (Yahoo chart endpoint, Nasdaq quote API as fallback; unofficial, dated in the dossier); otherwise the SEC-reported public float, marked “f”.

Private companies. Sponsors with no SEC equity match — about two thirds of the universe: private companies, unlisted subsidiaries, ex-US listings and names the join cannot resolve, which the dashboard does not tell apart — are resolved by name (current and previous legal names) to SEC Form D issuers in the Biotechnology and Pharmaceuticals industry groups, which gives Regulation D placements since 2019: amount sold per notice, offering size, investor count, and the officers, directors and promoters named on the filing. NIH RePORTER supplies non-dilutive awards by organisation name for the venture-stage tail and for the longevity companies, most of which are pre-IND and absent from ClinicalTrials.gov. Neither source contains valuations, terms, cap tables, burn or board seats; those are stated as unknown rather than estimated.

Programs & peers. The sponsor, not the trial, is the unit of analysis: each program is a vector over area × phase × modality, and peers are the sponsors with the highest cosine similarity on that vector. Trials are coordinates, not oracles — the peer set is how a readout should be read.

Specialist ownership. Latest two 13F-HR information tables for a fixed list of biotech-specialist managers (Baker Bros, Perceptive, RA, OrbiMed, Deerfield, Redmile, Avoro, EcoR1, Logos, Cormorant, Boxer, Vivo, Frazier, Rock Springs, Braidwell, Commodore, Darwin, Paradigm, Bain LS, Vestal Point, Samsara, TCG Crossover, Ally Bridge, Sofinnova, Sectoral, Tang, BVF); multi-strategy managers such as Point72, Viking, Woodline and Polar are deliberately excluded because their books dilute the signal, aggregated by issuer and matched to sponsors by name. This is the "specialist consensus" quality signal; count of holders matters more than dollars, and a 45-day lag applies.

Insider activity. Open-market purchases and sales (Form 4 transaction codes P and S) from the SEC's quarterly insider-transactions data sets, aggregated by issuer over the last two published quarters and joined by issuer CIK or ticker. The current quarter is not yet published, so the window ends at the last quarter-end. FINRA consolidated short interest is shown where the public feed responds.

Longevity companies. Names compiled from public reports (July 2026) are matched to lead-sponsor strings; aging-adjacent trials are flagged by mechanism and indication keywords across all sponsors. Capital raised is as publicly reported and is not re-verified here.

Company-guided milestones. Sentences in 8-K and 6-K filings (trailing 120 days, EDGAR full-text search) in which a company gives a period for topline data, an interim analysis, enrollment completion or a study start: "topline data in the second half of 2026". Each row keeps the sentence, the period precision (day, month, quarter, half, year) and the filing. They are the company’s own dates, shown beside the registry date and never merged with it; a period is placed at its end. Boilerplate (forward-looking statements) and PDUFA sentences are excluded, the latter because the regulatory calendar already carries them.

Registry versions. For every trial whose record changed since the weekly baseline, the dashboard fetches the ClinicalTrials.gov version list and shows the versions saved since then with the modules the sponsor edited. The snapshot diff also compares the primary-outcome text, so a changed endpoint shows its before and after; an edit to the outcome module alone is reported as an edit, not as an endpoint change.

Daily delta and the changes feed. A daily job re-pulls the registry, the regulatory calendar, filings and prices against the Monday baseline; the weekly job rebuilds everything. What moved is published as changes.json and changes.xml (RSS) beside the app: slipped, firmed and new trials, status changes and moved PDUFA dates, each linked to its record. The in-app change log reads the same list.

Live refresh queries the ClinicalTrials.gov v2 API for records updated since the snapshot and merges them in place; the status bar then shows the merged count separately from the weekly snapshot diff. It works from saros.gcarosso.bio and from the local file; a host that sandboxes outbound requests (such as an embedded artifact viewer) reports the block instead.