AIVEX
Back to Blog

AIVEX in Production: System Status Q1 2026

4 min readby AIVEX Team

What Is Running Right Now

All eight modules are live under the canonical aivex_runtime.py supervisor:

  • news — financial news ingestion and NLP sentiment pipeline (30 min cycle)
  • new_media — social media signal pipeline (30 min cycle)
  • chart — OHLCV ingest and technical indicator computation (60 min cycle)
  • companies — fundamental data and balance sheet updates (24 h cycle)
  • metrics — macro indicators (VIX, Fear & Greed, etc.) (60 min cycle)
  • signal_engine — atomic signal composition and final signal persistence (30 min cycle)
  • governor — policy evaluation and gating control (60 min cycle)
  • brain — pattern discovery and candidate generation (60 min cycle)

Each module runs as a separate OS subprocess. If one crashes, it does not affect the others. The watchdog checks every 60 seconds, restarts crashed modules with exponential backoff, and detects stuck processes via log growth and heartbeat files.

What We Shipped This Quarter

End-to-end signal traceability. Every pipeline run now generates a trace_id (tr_<16hex>) that is propagated through all atomic signals, composed signals, and the final final_signals database row. This makes it possible to query exactly which pipeline run produced any given signal.

Deep health diagnostics in Eye. The /eye/health endpoint now returns data_age_seconds, freshness_status (fresh / stale / very_stale), latency_ms, and a trust_verdict per module. This gives operators a clear, auditable view of system health at any point in time.

Watchdog alerting. The runtime now emits structured alerts through AlertManager on crash, stuck detection, permanent failure, kill-switch activation, and module gating. Alerts route to a rotating log at errors/alerts.log and optionally to a webhook.

Governor gating integrity. The gating.json write path is atomic (write-to-tmp → fsync → replace), making partial writes impossible. Fail-closed semantics: a corrupt file retains the last known gating state rather than clearing restrictions.

Current API Surface

The Eye REST API (/eye/*) exposes:

| Endpoint | Description | |---|---| | GET /eye/health | Per-module health with freshness and trust verdicts | | GET /eye/signals | Paginated final signals with filtering | | GET /eye/signals/{id} | Single signal by ID or signal_id | | GET /eye/prices | Latest prices per tracked symbol | | GET /eye/company/{symbol} | Company fundamentals | | GET /eye/snapshots | Pipeline run snapshots | | POST /eye/replay | Replay a past snapshot for comparison |

Authentication is role-based: EYE_VIEWER_TOKEN (read-only), EYE_OPERATOR_TOKEN (read + replay/resolve), EYE_ADMIN_TOKEN (full). If no tokens are configured, the API is permissive (development mode).

Signal Schema

Every signal persisted by the system carries these fields:

{
  "signal_id":    "sig_<12 hex chars>",
  "symbol":       "AAPL",
  "direction":    "BUY | SELL | HOLD",
  "confidence":   0.74,
  "strength":     "weak | moderate | strong | very_strong",
  "sources":      ["news", "chart"],
  "reasons":      ["..."],
  "atomic_count": 2,
  "has_conflict": false,
  "trace_id":     "tr_<16 hex chars>",
  "created_at":   "ISO8601"
}

This schema is stable. The trace_id field was added this quarter via a non-destructive migration (nullable, default NULL for historical rows).

What Is Not Yet Done

We deliberately document gaps here.

  • new_media excluded from composition — the NewMediaAtomicSource runs and persists data, but its output is not yet consumed at the composition step because it has no symbol-level mapping. This is a known gap.
  • outcomes_db not connected — Governor's drawdown-based kill-switch logic exists but has no live PnL data to operate on. Governor evaluates signal pattern statistics only.
  • Brain → Signal Engine link is planned, not built — Brain pattern candidates (human-approved) do not yet influence Signal Engine outputs.

We are documenting these openly because the architecture is already built to support them — the contracts, schemas, and module boundaries are in place.

For research and informational purposes only. Not financial advice.

For research and informational purposes only. Not financial advice.