The Governor Module: Risk Controls for AI Research Systems
Why a Separate Governance Layer?
Signal composition produces research outputs. But producing an output and emitting it are different decisions.
Between composition and the API, AIVEX inserts the Governor module — a dedicated governance layer that can block, modify, or delay any signal regardless of its confidence score. The Governor is the system's conscience.
What the Governor Checks
Minimum Confidence Threshold
The first check is simple: does the composite signal meet the minimum confidence threshold? The default is 0.40. Signals below this threshold are logged and discarded, not emitted.
This threshold is configurable per deployment. Research teams that want lower-confidence exploratory signals can lower it; teams that want high-conviction outputs only can raise it.
Symbol-Level Block List
The Governor maintains a block list of symbols that should never appear in outputs, regardless of signal confidence. This list can be updated at runtime via environment variable or database, without restarting the system.
Common uses: regulatory exclusions, pending litigation, actively managed positions that the research system should not comment on.
Kill Switch
A global kill switch disables all signal emission instantly. The kill switch can be toggled via environment variable (GOVERNOR_KILL_SWITCH=1) or via the admin API endpoint.
When the kill switch is active, the system continues running, logging, and analyzing data — it simply does not emit signals. This allows the team to investigate anomalies without shutting down data ingestion.
Exposure Concentration Limits
The Governor tracks sector concentration across recent signal emissions. If more than a configurable proportion of recent signals (default: 40%) point to the same sector, it throttles new sector-concentrated signals.
This prevents the system from over-indexing on a single sector narrative driven by correlated news events.
Audit Logging
Every Governor decision — allow or deny — generates a log entry with:
- Symbol and timestamp
- Composite signal details
- Which check caused the block (if denied)
- The Governor configuration at decision time
This audit trail is critical for post-hoc analysis and regulatory review.
Human Override
The Governor supports explicit human override in two directions:
Override allow: A symbol that would normally be blocked can be explicitly allowed for a configurable time window. Requires manual intervention (not API-accessible from the standard surface).
Override block: A symbol that passes all checks can be manually blocked. This is the "human-in-the-loop" mechanism for situations where the system cannot know what a human knows.
Both override types are logged with a required reason string.
What the Governor Is Not
The Governor is not a trading system. It does not make position-sizing decisions, calculate expected returns, or issue buy/sell orders.
It is a research output governance layer. Its only job is to decide whether a research signal should be emitted — and to log that decision so researchers can evaluate its quality over time.
For research and informational purposes only. Not financial advice.
For research and informational purposes only. Not financial advice.