ASTS research desk: an evidence timeline for one stock

1Problem

Following one early-stage, catalyst-driven stock means reading filings, regulator dockets, press and a lot of social noise, and it is easy to let a rumour harden into a "fact" over a few weeks. I wanted a desk where every claim points at its source, speculation is labelled as speculation, and AI helps with the reading without ever becoming a source itself. This one is outside security, and nothing on this page is investment advice.

2What I built

A local-first research app for AST SpaceMobile (ASTS): a Next.js application over a SQLite database with 23 tables. Events link to sources, relationships between events draw a "corkboard" graph, and on top sit the investor views: catalysts with dates and status, constellation deployment, capital and dilution, commercial and government traction, and a bull, base and bear thesis whose pillars cite evidence.

Three source rules are enforced in code, not just in prose: an event cannot be saved without a source; a "confirmed official" label needs an official, SEC or FCC document; and social posts can never confirm anything on their own. AI review is a separate layer, structured into facts, inferences, speculation and open questions, labelled with a model and a date, and never counted as a source.

The copy at the subdomain is a frozen read-only snapshot from October 2026, served from a container with no outbound access. Every write path is disabled four times over: the edge router, the app's request filter, a per-action guard and a read-only database file.

3How AI was used

Claude Code wrote the application and its tests, phase by phase against a written spec with a "done when" block per phase. Scheduled agent runs did the daily reading: a monitor that sweeps primary sources (EDGAR, FCC dockets, procurement notices) before the press covers them, an AI review worker that drains a file-based queue, a research sweep, a hearings watcher and a weekly audit that scores the desk's own calibration.

The plain-English brief is written under a rules layer that caps how confident a sentence may sound: conviction runs from 10 to 95 in steps of 5, the weakest cited record sets the ceiling, and speculation about defense use can never exceed 20.

4Results

293 events, every one with at least one source: 464 event-to-source links over 471 sources. 261 of the 293 sit at the two highest confidence tiers, confirmed by an official document or a reputable outlet. 140 of the sources are primary documents (official statements, SEC and FCC filings and orders, research-service reports) against 16 social posts. 50 catalysts are tracked (30 upcoming, 14 resolved, 6 passed), with 299 AI reviews, 67 patents and 967 cross-links between events.

The calibration layer is where the honest number lives: of 120 scored price-impact estimates, 40 were hits, 31 partial and 49 misses. The desk shows that score on its own audit page rather than hiding it.

5Status and limits

Paused since October 2026, with the scheduled runs switched off; the public copy is a snapshot, not a live feed. The social-media import stopped working in mid-2026 and was never repaired, so the credibility review of X accounts went quiet well before the pause. 11 of the 293 events are seed rows marked as samples. The price-impact scores above are the app's own estimates against a delayed quote feed, not a trading record. Nothing here is investment advice: it is one person's reading of public documents, and the AI layer is analysis, never a source.

6What I learned