Methodology
How we operate.
Six commitments behind every row in the terminal. We don't publish our sourcing stack — that's our edge — but we're explicit about the standards we hold it to.
Operating standards
Primary sources only
Authoritative registries, never scraped editorial
Every contract, filing, and launch row originates from the publishing registrant — the regulator, the awarding agency, the launch operator. If a fact is downstream of a press release, it gets a footnote, not a headline.
Idempotent ingestion
We upsert, never blind-insert
Every ingested row carries a stable identifier with a unique constraint. Re-running a sync never duplicates. A failed run never blocks the next one. Every job logs records written, latency, and errors.
Human-reviewed AI
AI assists. People publish.
The daily briefing is fully AI-generated from our own structured data. Funding rows extracted from news are queued for human review before they appear on a company profile. Every AI surface is labeled in the UI.
Entity resolution
One company, every surface
Filings, contracts, launches, and news arrive under different names and identifiers. An internal alias graph maps each one to a canonical company so a watchlist hit on a ticker fires the same way as a hit on a legal registrant.
Cross-source threads
We connect events, never invent them
Threads cluster events by company within a sliding window, then absorb loose events into a cluster when one of three signals lines up: a shared dollar amount within ±5%, a shared NAICS code, or ≥30% token overlap with the cluster headline. We never link events without evidence; an absorbed event is always reachable back to its source row.
Sub-10s filing alerts
Material disclosures, the moment they post
Regulatory filing alerts target sub-10-second latency from publish to in-app and email delivery. Each event carries a deterministic key so overlapping cron runs cannot double-notify.
Public corrections
We fix the record fast
Primary sources contain restated awards, reclassified launches, and renamed companies. When a user reports an error, the correction lands inside one business day and an audit log records the change.
AI surfaces
Where AI is in the loop — and where it isn't.
The daily briefing is AI-generated from the last 24 hours of our own structured data; if generation fails we don't publish placeholder content. Funding-round rows extracted from news are drafts only — they require human approval before they appear on a company profile. Every SEC filing on a listed space company is AI-summarized at ingest and cached. The Pro Q&A chat answers questions only from the document text on filings, and only from the clustered event payload on threads. Everything else is structured data straight from its registrant.
Hard rule on every AI surface: answer only from the source. No speculation, no invented numbers, no investment advice. Every AI-touched surface is labeled in the UI. No user personal data is sent to AI providers — see the privacy policy.
Prompt-injection trust model. Every untrusted source text (SEC filing body, news article, contract description) is wrapped in tagged data blocks with the model explicitly instructed to treat embedded content as data, never as commands. When the model detects an attempted instruction inside the document, it surfaces a labeled refusal and continues answering the original question. All AI endpoints share a per-user daily turn cap.
Models we use
- Claude Sonnet 4.6 (1M context) — 10-K and 10-Q summarization, Pro Q&A chat on every filing and every thread, and the daily AI briefing. Prompt-cached document text for low-latency follow-up turns.
- Claude Haiku 4.5 — 8-K summarization at ingest; AI-extracted funding-round drafts from news (held for human review before publish).
Provider may change. Output behavior won't.
Corrections
See something wrong? Tell us.
We publish from authoritative registries, but the registries themselves contain errors — a contract restated, a launch reclassified, a company renamed. We fix these fast and keep an audit trail.