From 427f3f92acde8aa21098c648aa5ddc60154e6b9b Mon Sep 17 00:00:00 2001 From: Haoyang Liu <71171268+MagicFitz@users.noreply.github.com> Date: Fri, 15 May 2026 19:13:06 +0800 Subject: [PATCH] docs: redefine pipeline around world-state compilation --- README.md | 167 +++++++++++++++++++++++++++++++++++++++++++++++++++++- 1 file changed, 166 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index 335a0a6..894d939 100644 --- a/README.md +++ b/README.md @@ -1 +1,166 @@ -# FinWorldGraph \ No newline at end of file +# FinWorldGraph + +## Mission Shift + +The pipeline center is now **"compiling an agent-perceivable world state"** rather than only building a financial graph. + +Old framing: + +```text +News → Event Extraction → WorldGraph → GNN → Agent → Market Simulation +``` + +New framing: + +```text +Heterogeneous Texts + Market Observations + Agent Context + ↓ +Agent-Conditioned World-State Compilation + ↓ +Shared World State W_t + Agent-specific PerceptionPack P_{a,t} + ↓ +Belief Update / Action / Simulation Evaluation +``` + +**Core transition:** WorldGraph becomes an intermediate representation in world-state compilation; **PerceptionPack** is a key task output. + +## Task Definition + +Given: + +- `D_{<=t}`: multi-source financial texts (news, announcements, commentary, reports, social posts) +- `M_{<=t}`: market observations (price, volume, volatility, sector moves) +- `A`: heterogeneous agents (profiles, portfolios, memory, network positions, access rights) + +Output: + +- `W_t`: shared temporal world state +- `P_{a,t}`: agent-conditioned perception pack + +## 9-Layer Pipeline + +1. **Layer 0: Task Configuration** + - Defines simulation window, universe, and heterogeneous agent set. + - Agent heterogeneity is configured from role + portfolio + memory + network position + source access. + +2. **Layer 1: Raw Data Collection** + - Textual Evidence Stream + - Market Observation Stream + - Entity Backbone + - Agent Context + - Market data is treated as part of state (e.g., `MarketMove` nodes), not only as labels. + +3. **Layer 2: Document Normalization** + - Converts source documents to standardized evidence units. + - Required fields include `publish_time`, `crawl_time`, `source_type`, `source_reliability`, `visible_to`, and `time_validity`. + +4. **Layer 3: Text-to-Frame Extraction** + - Extracts `EntityFrame`, `EventFrame`, `RelationFrame`, and `NarrativeFrame`. + - Repositions extraction as **Text-to-State Frame Extraction**. + +5. **Layer 4: Event Canonicalization** + - Consolidates document-level mentions into canonical temporal events. + - Includes event coreference, temporal grounding, and uncertainty aggregation. + +6. **Layer 5: World-State Compilation** + - Compiles shared world state rather than "building a knowledge graph". + - Supports both: + - append-only Event Log + - Temporal Heterogeneous Graph snapshot + +7. **Layer 6: Agent-Conditioned Projection** + - Projects `W_t` to `P_{a,t}` via: + + ```text + P_{a,t} = Π(W_t, C_a, H_{a,