初始化项目，由ModelHub XC社区提供模型

Model: Italianhype/Blum-Finance-4B
Source: Original Platform

CONTRIBUTING.md

@@ -0,0 +1,56 @@
+# Contributing Learning Evidence
+
+BLUM Finance does not collect prompts, outputs, account data, or usage telemetry.
+Community learning is explicit and evidence-bound.
+
+## Contribution lifecycle
+
+1. Run BLUM Finance locally and retain the point-in-time request and response.
+2. After the stated horizon, attach an observed outcome and verified provenance.
+3. Create a redacted contribution bundle:
+
+   ```bash
+   blum-contribute example.json --output contribution.json --consent
+   ```
+
+4. Inspect the bundle locally. To submit it for review, explicitly run:
+
+   ```bash
+   blum-contribute example.json --output contribution.json --consent --push
+   ```
+
+The upload opens a pull request against `Italianhype/Blum-Finance-Memory`.
+It never writes directly to accepted memory or released model weights.
+
+## Required evidence
+
+A contribution must contain:
+
+- a timestamped request with point-in-time evidence;
+- the model response generated at that timestamp;
+- a mature outcome observed after the decision;
+- verified source provenance and an explicit quality score;
+- explicit consent under the contribution license.
+
+Pending, inconclusive, chronologically invalid, tampered, or unverified examples
+remain quarantined. Secrets, account identifiers, email addresses and Hugging
+Face tokens are removed from generated bundles.
+
+## Local memory
+
+Eligible bundles can improve a local installation without changing weights:
+
+```bash
+blum-memory-add contribution.json
+```
+
+The inference pipeline retrieves only outcomes observable before the new
+request's `as_of` timestamp. Retrieved records are labeled as historical
+analogies and cannot replace current evidence.
+
+## Model updates
+
+Accepted records may enter a future immutable dataset snapshot. A training run
+always creates a challenger. Promotion requires temporal holdout evaluation,
+no-fabrication and schema checks, adequate sample quality, and an explicit
+versioned release. Anonymous inputs never self-modify a published model.