Rewrite arena_inputs/README.txt; numbered-emoji steps; contact

README.txt claimed only third_view ships prompt_*.txt — all four exist
with 25 rows each, and captions carry a per-view prefix. Rewritten to
match the filesystem, with the intrinsics format, non-uniform image
sizes and FOV range that adapters need. The skill's three
"README is stale" warnings are dropped accordingly; the general
trust-the-filesystem rule stays.

Root README: 1/2 emoji for the two steps, contact address.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: xjxu21

.claude/skills/world-model-adapter/SKILL.md

@@ -64,9 +64,10 @@ Converse in the user's language. Artifacts follow fixed rules regardless:
 
 ### 1. Read the inputs first
 
-`arena_inputs/action_protocol.txt` is authoritative for the action semantics. For the file
-layout, **list the directories yourself** — `README.txt` has been stale before (it still
-understates which prompt files ship).
+`arena_inputs/action_protocol.txt` is authoritative for the action semantics and
+`arena_inputs/README.txt` describes the layout. Still **list the directories yourself** to
+confirm — the inputs can be revised, and a doc that has drifted is invisible until it costs you
+a full run.
 
 **[references/output_spec.md](references/output_spec.md) §1** records the layout and the
 verified facts about the data — including three things that catch people out:

.claude/skills/world-model-adapter/references/known_traps.md

@@ -164,9 +164,6 @@ Failure modes seen or easily reachable:
 Use `{view}/prompt_{domain}.txt` row `i` for image `i`, verbatim, and record
 `prompt.source` as `file:row` in the manifest so it is auditable.
 
-Note the zip's own `README.txt` is stale here (it claims only `third_view` has prompts) — see
-trap 12.
-
 ## 15. Inputs are not uniform in size or aspect ratio → E5, H3
 
 Measured: `real` images span 1024×768 … 1920×1200 (AR 1.33–1.79); `style` is uniformly

.claude/skills/world-model-adapter/references/output_spec.md

@@ -23,9 +23,8 @@ arena_inputs/
     style/ style_intrinsics/ style_action.txt  prompt_style.txt
 ```
 
-> `README.txt` in the zip is **stale** on one point: it says only `third_view` ships
-> `prompt_*.txt`. All four `{view}/prompt_{domain}.txt` now exist, 25 rows each. Trust the
-> filesystem over the README (trap 12), and re-check both when the zip is updated.
+> `README.txt` matches this layout as shipped. Still trust the filesystem over any doc
+> (trap 12), and re-check both if the inputs are revised.
 
 Dimensions: **view** ∈ {`first_view`, `third_view`} × **domain** ∈ {`real`, `style`}.
 Per (model × view × domain): 25 images × 5 actions = **125 videos**.

README.md

@@ -18,17 +18,17 @@ memory, and visual quality.
 Using WorldMark is two steps.
 
 ```
-arena_inputs/  ──①generate──▶  {domain}/{view}/{MODEL}/*.mp4  ──②evaluate──▶  scores
+arena_inputs/  ──1️⃣ generate──▶  {domain}/{view}/{MODEL}/*.mp4  ──2️⃣ evaluate──▶  scores
 ```
 
 | | | |
 |---|---|---|
-| **①** | **[Generate videos](#-generate-videos)** | drive your model with our images + action sequences |
-| **②** | **[Evaluate](#-evaluate)** | score the videos on the 9 metrics |
+| 1️⃣ | **[Generate videos](#1️⃣-generate-videos)** | drive your model with our images + action sequences |
+| 2️⃣ | **[Evaluate](#2️⃣-evaluate)** | score the videos on the 9 metrics |
 
 ---
 
-## ① Generate videos
+## 1️⃣ Generate videos
 
 Inputs live in **[`arena_inputs/`](arena_inputs/)** (read-only): 25 starting images per
 `view × domain`, the action assignment for each image, per-image intrinsics, and captions.
@@ -80,7 +80,7 @@ will produce meaningless numbers.
 
 ---
 
-## ② Evaluate
+## 2️⃣ Evaluate
 
 ```bash
 cd evaluation
@@ -118,6 +118,8 @@ figures: **[`evaluation/README.md`](evaluation/README.md)**.
 }
 ```
 
+Questions, or a model you would like added: **xjxu21@gmail.com**
+
 ## License
 
 See [`LICENSE`](LICENSE).

arena_inputs/README.txt

@@ -1,6 +1,29 @@
-WorldMark benchmark inputs: first-/third-person starting images + action sequences.
-  first_view/  real|style  : 25 first-person starting images (000..024.jpg) + *_action.txt + *_intrinsics
-  third_view/  real|style  : 25 third-person starting images + *_action.txt + *_intrinsics + prompt_*.txt
-  action_protocol.txt      : action_id -> key sequence
-The per-image action assignment (*_action.txt) is identical for first and third view.
-Each (image, action) pair is one generated video, named {image}_{action:03d}.mp4.
+WorldMark benchmark inputs  —  read-only, never edit.
+
+Layout
+  action_protocol.txt              action_id (1..15) -> key sequence
+  {view}/                          view = first_view | third_view
+    {domain}/NNN.jpg               domain = real | style; 25 images, 000..024
+    {domain}_intrinsics/NNN_intrinsics.npy
+    {domain}_action.txt            row i = image i's 5 assigned action_ids
+    prompt_{domain}.txt            row i = image i's caption (25 rows)
+
+All four {view}/prompt_{domain}.txt exist. Captions differ by view: every first_view caption
+starts with "First-person view." and every third_view caption with "Third-person view." — the
+prompt already carries the view signal, so never reuse a caption across views or domains.
+
+Facts
+  *_action.txt is identical between first_view and third_view for a given domain.
+  Intrinsics are (4,) float32 [fx, fy, cx, cy], per-image and all distinct, in ORIGINAL-image
+    pixel units (cx ~ (W-1)/2, cy ~ (H-1)/2) — transform them for your model's resolution,
+    do not pass them through.
+  Image sizes are not uniform: real spans 1024x768 .. 1920x1200 (AR 1.33-1.79);
+    style is uniformly 1024x1024 (AR 1.00).
+  Horizontal FOV spans 22-90 degrees (median ~66 for real, ~45 for style).
+
+Scope
+  Each (image, action) pair is one generated video, named {image:03d}_{action:03d}.mp4.
+  One (view, domain) pair = 25 images x 5 actions = 125 videos; all four = 500.
+
+See ../generation/README.md to produce videos, and
+../.claude/skills/world-model-adapter/references/output_spec.md for the full I/O contract.

generation/README.md

@@ -29,7 +29,7 @@ actually been made — worth a skim before you conclude anything about a repo.
 
 The skill in `../.claude/skills/world-model-adapter/` does the adapter and the batch render.
 It needs the **model repo** and **which view + domain**; see the repo
-[README](../README.md#-generate-videos).
+[README](../README.md#1%EF%B8%8F%E2%83%A3-generate-videos).
 
 ## `check_delivery.py`