Owner's verdict on the v1 dashboard: a complete shit show — can't keep track of or find anything. Fair. v1 renders all 94 trials as one stack of dense tables with no triage, no per-name view, no search, and the journal buried at the bottom.

• Full redesign spec written at stockradar/design/ui-spec.md: five views (Today / Names / Calendar / Candidates / More), global search, group filter chips, a rolling 14-day activity feed, routine-trial noise hiding, ticker cards with a per-name detail panel, bottom tab bar on phone, strict color discipline. Vanilla JS now allowed; still one static file, no framework.
• The spec went to Claude Design for a visual draft; implementation in render.py follows once the design is settled.
• The main Signal Radar dashboard sidebar now links to Catalysts (signal_radar/preview.py, _NAV_MAIN).

Goal set: find and be positioned for genomics innovation over the next 3–5 years. The radar exists so that nothing discussed in a Claude session gets lost: findings land here in the journal, candidates land in the candidates section with a first-seen date, and the calendar tracks every known catalyst.

What went live today

• The Phase 3 screener was validated live against ClinicalTrials.gov (its query syntax works as written) and deployed unchanged, wrapped by radar_run.py. Nightly run at 05:30 ET weekdays; discovery sweep plus candidate mapping on Sundays.
• The wrapper adds what the raw screener lacked: a Phase 2 pass for the therapeutic names (the biggest platform jumps come from early data, not Phase 3), estimated-versus-actual completion-date tracking (a flip to "actual" means last patient visit done, topline typically inside one to four months), sponsor-and-collaborator ticker mapping through SEC EDGAR so discovery surfaces tradeable names instead of raw sponsor strings, and MarketData-based LEAPS availability per name for long-horizon positioning.
• Data providers confirmed: marketdata.app is primary (token already lives on this box and the account is IP-locked here, which is correct) and Databento is the backup when MarketData credits run out. The IV term-structure scan around catalysts is built but off by default to protect the MarketData credit ceiling — flip iv_scan: true in universe.yaml to enable.

First findings (from the live runs)

• The 12-month Phase 3 window holds 17 trials across only 9 of 30 watchlist names. Nearest: REGENXBIO's Duchenne gene therapy (September 2026), the AbbVie-run RGX-314 eye trials (October and December 2026), Sarepta's high-dose eteplirsen (end of October 2026).
• The discovery sweep surfaced two apparently-new names; the tradeability check then killed one of them, which is the system working: **Cartesian Therapeutics (RNAC)** — mRNA CAR-T for myasthenia gravis, pivotal window November 2026, options listed but only out to February 2027 — was added to the watchlist. Adverum looked like a find, but SEC EDGAR and MarketData both came back empty, and it turned out Eli Lilly acquired Adverum ($3.56/share plus CVR, delisted 2025-12-09) — before the pivotal readout, which is the M&A-before-data pattern to expect more of in gene therapy.
• Same check exposed a stale name in the original watchlist: Abbott completed its $21B takeover of Exact Sciences on 2026-03-23, so EXAS is delisted too. Both now live in the delisted list in universe.yaml.
• Regulatory picture (from today's search, sources in events.yaml): Intellia's HAELO pivotal for hereditary angioedema succeeded in April 2026; BLA guided for H2 2026, launch planned H1 2027 — would be the first in-vivo CRISPR approval. Note the FDA also placed a hold on two Intellia nex-z (ATTR) trials earlier in 2026 — the registry may lag on this; treat NTLA's 2027 ATTR timeline as soft. uniQure got FDA alignment in June 2026 for an accelerated-approval BLA on AMT-130 (Huntington's), submission guided Q3 2026. Alnylam/Regeneron's cemdisiran has a November 2026 FDA action date.

Post-mortem, same evening: would this have flagged today's MRNA double?

Honest answer: no, not as built this morning — and the reason reshaped the radar. Moderna roughly doubled today on INTerpath-001: the Merck-partnered individualized mRNA cancer vaccine plus Keytruda met its endpoints in resected melanoma — the first positive Phase 3 for any mRNA cancer therapy (FDA also cleared the mFLUSIVA flu shot the same day). The trial WAS reachable by our sponsor query (ModernaTX is listed as collaborator under Merck's lead on NCT05933577), but its registry primary-completion date is October 2029 — today's result came from an event-driven interim analysis, three years ahead of the registry date, so a forward completion-date window could never contain it. Meanwhile the readout was guided by the companies for 2026, meaning the guidance layer (events.yaml) was the right net and simply hadn't been fed.

Three changes went in tonight:

• Beyond-window pass: every watchlist name's Phase 3 trials out to five years now render in their own section, so standing heavyweights like INTerpath-001 are visible even when the registry date is far and unreliable.
• Lookback window: trials whose primary completion passed in the last ~4 months (including newly COMPLETED ones) now stay on the calendar marked "data due" — completion-to-topline is the hottest zone and the old filter structurally dropped it.
• Standing chore, not yet automated: once a quarter (earnings season), sweep each watchlist name's guidance for "data expected in ..." statements into events.yaml. The registry times the ordinary readouts; only guidance times the big event-driven ones.

Also from tonight: a daily email brief now goes out after each run — changes, data-due list, everything inside 30 days, new candidates — via the same Resend account the Signal Radar digest uses.

Decisions

• Frontend is a static page at radar.gpt.cafe/catalysts/ — no framework, no build chain, so it will still render in five years.
• Diagnostics names stay on the radar but their real catalysts are coverage and guideline decisions, not trials; don't read their empty trial rows as "no catalysts".
• MarketData enrichment kept deliberately cheap (expirations only, cached a day). The screener file itself stays untouched per the deployment brief.