Agentic GraphRAG over S&P 10-Ks with Laya as a shadow decision engine beside Jev

# Agentic GraphRAG over S&P 10-Ks (Aiven talk demo)

A portfolio manager holds sixteen S&P 500 companies. The system holds their 2024 10-K filings as a
**vertical knowledge graph** (what the agent knows: typed facts `Company -relationship-> Target`,
each pointing at the passage it came from, plus the passages themselves with embeddings) and runs a
**horizontal workflow graph** (how the agent acts: a DAG of decision, reasoning, execution and
generation nodes). A **harness** advances the DAG, and the screen shows both graphs live, plus the
**execution graph**: the append-only record of every node firing with engine, model version,
confidence, latency and cost.

Typed decisions (Route, Enough?, Layer 0 per claim, Promote?) are yes/no and pick-one questions with declared
answer spaces, never generated prose. The acting engine is **TypeSafe Jev** when a key is present, else a labelled
Claude Haiku fallback. **Laya** (open, Apache-2.0, 421M ModernBERT, the three-day reimplementation of Jev's
interface) runs locally as a **shadow engine** on the same questions and its answers are shown beside the acting
engine's: on the TSMC run it routed the headline as a plain question and rejected every true claim, which is the
book's point that calibration belongs to the answer type and did not clone with the interface. Set
`DEMO_DECISION_ENGINE=laya` to let it act anyway. Structure-producing reasoning (Plan, Pick tool, Layer 1) runs on
Haiku 4.5; the one prose step (Synthesize) on Sonnet 5 or Haiku, as Route decided.

**Walkthrough mode** (button next to Run) replays the cached run one node at a time: what happened, the data that
moved, the code logic, the pillar, and where the piece sits on AWS (`docs/aws-positioning.md`, served as
`/api/positioning`). `docs/graph-harness-landscape.md` m