Events & premise invalidation

The EventAgent ingests policy and macro events from authoritative feeds, LLM-tags them against a fixed vocabulary, and cross-references them against every open thesis. When an event’s tags intersect a thesis’s premise_tags, a PREMISE_INVALIDATION alert fires.

This is the substrate that makes regime-aware investing mechanical rather than a vibe.

The vocabulary

EVENT_TAGS (in cents.models.event) is a fixed set of ~30 themes covering the regime variables a US-equities thesis might depend on:

# Trade / tariffs
"tariffs.universal", "tariffs.china", "tariffs.eu", "tariffs.mexico_canada",
"tariffs.sectoral", "export_controls", "sanctions",

# Fiscal / tax
"tax_policy", "fiscal_spending", "debt_ceiling", "shutdown",

# Monetary / FX
"fed_policy", "rates", "dollar",

# Sectoral policy
"energy_policy", "energy_permitting", "clean_energy_credits",
"semis_policy", "ai_policy", "healthcare_policy", "drug_pricing",
"antitrust", "financial_regulation", "crypto_policy", "defense_spending",
"labor_policy", "immigration_policy",

# Legal / regulatory
"scotus_ruling", "executive_order",

# Macro themes
"ai_capex", "reshoring", "deglobalization", "geopolitical_conflict",

Both sides of the system — the EventAgent (tagging fetched events) and the premise classifier (tagging theses with their dependencies) — draw from this single list. That’s what makes the intersection meaningful.

Adding tags is safe; renaming them is not.

The data source

v1 source is the Federal Register API (/api/v1/documents.json), which publishes every executive order, agency rule, and regulation. The EventAgent pulls types PRESDOCU (presidential documents incl. executive orders), RULE (final rules), and PRORULE (proposed rules). No API key required.

Subsequent sources — SEC EDGAR (8-Ks for sectoral exposure), the SCOTUS calendar, USTR/Treasury/OFAC press feeds, FOMC schedule — are follow-ups tracked in GitHub issues. The EventAgent pipeline (tagging, persistence, premise matching) is source-agnostic; adding a source is a new _fetch_* method that yields raw event dicts.

How tagging works

For each ingested event, the EventAgent calls Claude Haiku with the event’s title + abstract + type, and the EVENT_TAGS vocabulary. The prompt asks: “which regime variables does this US federal action relate to?” — explicitly framed so a tag belongs only if a thesis depending on that variable would be materially affected. Tags outside the vocabulary are dropped.

The same regime-variable framing is used by the premise classifier on the thesis side. Same vocab, same lens, both sides — so the intersection is mechanical and meaningful, not coincidental.

Premise invalidation, end-to-end

1. You (or the factory) create a thesis with premise_tags = ["tariffs.china", "ai_capex"].
2. EventAgent.refresh() pulls new Federal Register documents, LLM-tags them, persists to the events table.
3. For each newly-ingested event, the agent cross-references against every open thesis. Any thesis whose premise_tags intersect the event’s tags gets a PREMISE_INVALIDATION alert with data.thesis_id, data.matched_tags, and the event’s polarity.
4. cents factory run sees the alert in its close phase, closes the matching thesis as INVALIDATED, and records the closure reason on the outcome.
5. The cooldown rule prevents the same symbol from being reopened in the same run (until the event ages out).

CLI

cents event refresh                  # pull new Federal Register events
cents event list --tag tariffs.china # filter by tag
cents event list --since-days 7      # window
cents event show <event-id>          # full detail incl. agencies, polarity

The refresh hooks into cents scan automatically — every scan cycle pulls events once and fires any cross-reference alerts before the per-symbol scan runs.

What it deliberately doesn’t do

• No real-time push — refresh is pull-based, on demand or via cron
• No source other than Federal Register in v1
• No NLP beyond LLM tagging — no entity extraction, no sentiment beyond the bullish/bearish/neutral coarse polarity
• No backtest mode — as_of parameter exists on the agent surface but the factory doesn’t use it yet