---
license: mit
tags: [agent-trace, game-ai, behavior-trace]
pretty_name: Parry BrainTraces
---

# Parry — BrainTraces (shared agent traces)

Decision-by-decision traces of the **Parry** duelist — a Qwen2.5-1.5B running
**in the player's browser via WebGPU** as a real-time fighting-game opponent
(HF x Gradio Build Small Hackathon, Track 2).

Each row of `rows` is ONE decision of the two-speed brain:
`{tick, stateDigest, planString (the Analyst's live read of the player),
planSource (analyst | override), prediction {habit, confidence}, intent
(the grammar-constrained action), opMs (decision latency), events}`.

Header carries the full provenance: model id, backend, engine/prompt/grammar
versions, match seed, and final score — every trace is replayable against the
deterministic engine at that version.

Export your own: play https://huggingface.co/spaces/build-small-hackathon/parry,
press **B** (judge panel) → *Export trace*. The Space's `trace_digest` Gradio
API validates and summarizes any trace via `gradio_client`.

---

### 🧭 All Parry artifacts

**Play:** [the Space](https://huggingface.co/spaces/build-small-hackathon/parry) ·
**Models:** [merged](https://huggingface.co/Jainamshahhh/parry-tactician-1.5b-merged) / [LoRA](https://huggingface.co/Jainamshahhh/parry-tactician-1.5b-lora) / [GGUF · llama.cpp](https://huggingface.co/Jainamshahhh/parry-tactician-1.5b-gguf) ·
**Data:** [agent traces](https://huggingface.co/datasets/Jainamshahhh/parry-traces) / [field notes](https://huggingface.co/datasets/Jainamshahhh/parry-field-notes)