TODO.md

Trenches — TODO

Reward System

• Event-prediction RL rewards — when a real-world event occurs and an agent's prior prediction/action aligns with it, grant a positive reward signal. This closes the loop between live data ingestion and agent learning.

  • Track agent predictions per turn (e.g., "Iran will retaliate within 2 turns")
  • Compare predictions against actual events that fire from RSS/OSINT feeds
  • Reward = f(prediction accuracy, lead time, specificity)
  • Only real events (from live feeds or env-generated stochastic events) impact the reward signal

• Chat-injected fake events — allow manual event injection via the chat panel that influences agent behavior but does not affect reward calculations.

  • Tag chat-injected events with source: "manual" vs real events with source: "live" or source: "env"
  • Agents still react to fake events (observe and act), but the reward function filters them out
  • Useful for demos, testing edge cases, and probing agent behavior without polluting the training signal

UI / Frontend

• Event timeline with time control — scrubber bar (like a video editor) for navigating, rewinding, and branching the simulation

  • Scrubber bar at the bottom: drag to jump to any turn/timestamp, play/pause, rewind, fast-forward
  • Two event types on the timeline: predictions (agent forecasts) and actuals (confirmed real events)
  • Predictions that matched actual outcomes are visually linked; incorrect ones shown faded
  • Branching: when a fake scenario is injected via chat, the timeline forks — you can scrub back to before the injection and see the "what if" branch vs the real timeline
  • Playback controls: step-by-step (turn by turn), continuous playback at adjustable speed
  • Markers on the scrubber for key events (escalations, interventions, injected scenarios)
  • Filterable by agent, event type, and time range
  • Feeds into the reward system — correct predictions on the timeline = positive RL signal

• Merge tension/stats pills into top bar

• Disable text selection on floating panels

• Remove Mapbox logo

• Clean up README

Infrastructure

• Push to HF Space (AlazarM/trenches)
• Add NEXT_PUBLIC_MAPBOX_TOKEN as HF Space secret

Post-Training

• 6 synthetic seed replay datasets (in synthetic_historical_replays/)
• Training CLI with GRPO, hyperparameter args, checkpointing
• Local smoke test (tiny-gpt2, US + Israel)
• HF GPU smoke test on T4 (trenches-training-smoke)
• All 6 entity models → Qwen/Qwen3-8B (no quantization)
• Historical data collection pipeline (GDELT → replay JSON)
• Run historical collector for all 6 entities
• Curator-review collected replay data
• Spin up 6 HF A100 Spaces for production training
• Evaluation/baseline reporting