Card: 44 cases

README.md

@@ -1,38 +1,29 @@
 ---
-title: TCS Trace Viewer
-emoji: 🔭
-colorFrom: indigo
-colorTo: green
+title: MR-9B Scaffold — 0/7 Trajectories
+emoji: 🔬
+colorFrom: red
+colorTo: indigo
 sdk: static
 app_file: index.html
 pinned: false
 license: mit
-short_description: Browse meta-reasoning trajectories (midtrain + eval)
+short_description: How MR-9B explores SODA-2026 problems 0-step FA can't solve
 ---
 
-# TCS Trace Viewer
+# MR-9B RL scaffold on the FA-0-zero SODA-2026 problems
 
-A static viewer for meta-reasoning trajectories. It renders, with one UI, both
-**Midtrain V3** scout trajectories assembled from an SFT JSONL (e.g.
-[`HerrHruby/MR_midtrain_v3_sft_core`](https://huggingface.co/datasets/HerrHruby/MR_midtrain_v3_sft_core))
-and **eval rollouts** (IMO / SODA) from `inference.mrv3.run_dual`, the latter
-carrying judge scores, a bench label, and the termination reason.
+A static trace viewer for the **44 SODA-2026 `standalone_hard` (proof_writing)**
+problems that the RL policy `meta-learning-llms/MR_9B_rl_215_step45` scored **0/7 on
+all 8 samples** with a **0-step final answer** — after excluding 9 rows with broken
+(`Total: 0`) rubrics that score 0 by construction.
 
-For each trajectory it shows the **problem** and, per layer:
+Each trajectory is the **full 8-layer MR/E/FA scaffold** (MR/FA = the RL policy,
+E = frozen `MR_midtrain_9B_v3`). It shows, in order:
 
-- the **assembled frontier** the planner consumed (the exact `Exploration so far:`
-  context the MR target was generated from),
-- the **directions chosen** at that layer,
-- the **planner CoT**, the **execution traces** (CoT + finding), the per-execution
-  **summaries** (summary / rationale / core result),
-- and the synthesized **final answers** with judge scores and coverage.
+1. the **problem** + its **ground-truth rubric + reference solution**,
+2. the **final answer**, its **judge score** (gpt-oss-120b), and the **judge's
+   reasoning** for that score,
+3. then the **MR/E exploration** — per-layer directions, planner CoT, assembled
+   frontier, and each executed direction's finding + summary.
 
-Use the sidebar to filter and switch between trajectories; the toggles in the top
-bar control which views are shown.
-
-This Space bundles the pre-built trajectory index (`data/`) so it is fully
-self-contained — no external data fetch at runtime. Math renders with KaTeX and
-prose with marked, both vendored locally under `vendor/`.
-
-To regenerate or extend the index, see `build_index.py` (midtrain SFT) and
-`build_eval_index.py` (eval rollouts) in the source repo.
+With one attempt the scaffold rescues **13/44** of these (vs 0/8 for FA-0).