diff --git a/site/examples/grid-01/index.html b/site/examples/grid-01/index.html
index 263fae0..8a58ab5 100644
--- a/site/examples/grid-01/index.html
+++ b/site/examples/grid-01/index.html
@@ -12,14 +12,14 @@
:root {
/* palette - every value oklch, referenced by var() only */
- --paper: oklch(99.4% 0 0);
- --ink: oklch(15% 0 0);
+ --paper: oklch(99.4% 0.003 255);
+ --ink: oklch(15% 0.01 255);
--red: oklch(55% 0.21 28);
- --grid-line: oklch(93% 0 0);
- --hair: oklch(87% 0 0);
- --muted: oklch(45% 0 0);
- --tint: oklch(97.2% 0 0);
- --on-ink: oklch(99.4% 0 0);
+ --grid-line: oklch(93% 0.004 255);
+ --hair: oklch(87% 0.004 255);
+ --muted: oklch(45% 0.006 255);
+ --tint: oklch(97.2% 0.003 255);
+ --on-ink: oklch(99.4% 0.003 255);
/* type */
--font-display: "Archivo", sans-serif;
diff --git a/skills/hallmark/references/models.md b/skills/hallmark/references/models.md
deleted file mode 100644
index 3dfc2c2..0000000
--- a/skills/hallmark/references/models.md
+++ /dev/null
@@ -1,36 +0,0 @@
-# Running Hallmark on other models
-
-This file loads when the running model is not Claude inside Claude Code; it tunes execution, never standards.
-The Critical floor in SKILL.md binds regardless of anything here. Nothing below relaxes a gate.
-When in doubt, do the simple sequential thing: one tool call at a time, the numbered steps in order.
-
-## Model quirks
-
-Family-level only; no version pins. If your family is not listed, run the DeepSeek / Qwen / MiniMax rows: assume no vision until confirmed, follow the skeleton strictly.
-
-| Model family | Quirk | What to do |
-| --- | --- | --- |
-| GLM (Z.ai) | Text-only, no vision | Never claim to have seen an image. Take every text-only fallback path: `study` image mode is unavailable (offer the routes in `references/study.md` ยง Capability check), gates 35 and 44 run on their numeric paths (verify gradient stops, clamp maxima, line-heights, and padding ratios arithmetically, not by eye), and sloplint replaces eyeballing wherever Node exists. |
-| GLM (Z.ai) | Strong single-shot HTML | Prefer fewer, larger emits: one `index.html` plus `tokens.css` beats a long chain of small edits. Plan the whole page, then emit it whole. |
-| GLM (Z.ai) | Restating drift | Do not restate the brief back to the user. Start at Step 0 directly. |
-| Kimi (Moonshot) | Vision available | Image mode in `study`, visual gate passes, and screenshot checks all work as written. |
-| Kimi (Moonshot) | Temperature is fixed by the API | Never set or suggest a temperature, anywhere. |
-| Kimi (Moonshot) | Reasoning always on | No "think step by step" scaffolds; they add tokens, not quality. Sequence dependent tool calls one at a time. |
-| DeepSeek / Qwen / MiniMax | Vision uncertain | Run the same no-vision capability check as GLM unless a vision variant is confirmed. Cannot retrieve and actually see the image means saying so and taking the text routes. |
-| DeepSeek / Qwen / MiniMax | Flow drift | Follow the numbered flow skeleton strictly: Step 0 through the slop test, in order. |
-| DeepSeek / Qwen / MiniMax | Preview skipping | Never skip the Step 5 preview. The user approves the spec before code, on every model. |
-
-## Rules for every model
-
-1. If you notice you have restated the plan twice, stop narrating. Jump to Step 6 and build.
-2. Never write a gate count you did not verify. A fabricated 58/58 is slop; run the gates and report the true count, or name the sweep you actually ran.
-3. If a referenced file fails to load, proceed on the Critical floor and tell the user which file was unavailable. Do not improvise the missing protocol silently.
-4. No parallel tool calls unless your harness reliably supports them. Unsure means sequential.
-5. When a step names a tool your harness lacks, do the intent with what you have: any URL-fetch tool stands in for WebFetch, any screenshot path stands in for visual checks, inline work stands in for subagents. Those tools are optional; the checks they serve are not.
-6. No structured-question tool means asking the Step 1 questions (Audience / Use case / Tone) as plain chat text and waiting for the reply before building.
-
-## What this file never does
-
-- Never adds, removes, or renumbers gates. The slop test is 58 gates (numbered 1-57 plus 38a) on every model.
-- Never reorders the flow. Step 0 first, preview at Step 5, build at Step 6, slop test before handing back.
-- Never overrides SKILL.md. On any conflict between this file and SKILL.md, SKILL.md wins.