# Pending Guideline Fixes (September 2025)

## Overview
The following gaps violate our "no stubs, no synthetic data" policy and must
be resolved before the dashboard can operate in production. Inline TODOs with
matching wording have been added in the codebase.

## Items
1. **Prediction aggregation** – `TradingOrchestrator._get_all_predictions` still
   raises until the real ModelManager integration is written. The decision loop
   intentionally skips synthetic fallback signals.
2. **Device handling for CNN checkpoints** – the orchestrator references
   `self.device` while loading weights; define and manage the device before the
   load occurs.
3. **Trading balance access** – `TradingExecutor.get_balance` is currently
   `NotImplementedError`. Provide a real balance snapshot (simulation and live).
4. **Fallback pricing** – `_get_current_price` now raises when no market price
   is available. Implement a real degraded-mode data path instead of hardcoded
   ETH/BTC prices.
5. **Pivot context prerequisites** – ensure pivot bounds exist (or are freshly
   calculated) before requesting normalized pivot features.
6. **Decision-fusion training features** – the dashboard still relies on random
   vectors for decision fusion. Replace them with real feature tensors derived
   from market data.

## Next Steps
- Prioritise restoring real prediction outputs so the orchestrator can resume
  trading decisions without synthetic stand-ins.
- Sequence the remaining work so that downstream components (dashboard panels,
  executor feedback) receive genuine data once more.