JWriter20 · JWriter20 · Aug 23, 2026 · Aug 22, 2026 · Aug 23, 2026
diff --git a/README.md b/README.md
@@ -23,7 +23,7 @@
 </p>
 
 > ⭐ **Enjoying CaptchaKraken?** [Star & watch the repo](https://github.com/JWriter20/CaptchaKraken)
-> for new puzzle types, smaller models, and video support. One repo, two
+> for new puzzle types and new models. One repo, two
 > published ports — the TypeScript browser driver (**npm:** `captchakraken`) and
 > the Python engine (**PyPI:** `captchakraken`).
 
@@ -178,7 +178,7 @@ CaptchaKraken detects the captcha, solves it, clicks, and verifies — end to en
 | ✅ **NetEase Yidun** | Jigsaw, picture-click, icon-click |
 | ✅ **Lemin, Prosopo, Tencent** | Cropped-image, grid, and slide flows |
 | ✅ **Distorted text** | BotDetect, MTCaptcha, Yandex — read and typed, not clicked |
-| 🟡 Video challenges | Driver support ships; the adapter `setup.sh` installs skips them ([roadmap](./docs/roadmap.md)) |
+| ✅ **Animated / video challenges** | Recorded, cut into keyframes, solved as one multi-image prompt, then clicked once the widget returns to the chosen frame |
 
 The non-Google/hCaptcha vendors are driven end-to-end in CI against generated
 fixtures in **both** ports. Per-vendor accuracy varies more than the headline
@@ -341,26 +341,23 @@ cell indices:
 
 ### End-to-end, in a real browser
 
-Both runs below drove **live reCAPTCHA challenges** at
+The run below drove **live reCAPTCHA challenges** at
 `google.com/recaptcha/api2/demo` on 2026-07-27, headless, against the hosted
 API. A "solve" means reCAPTCHA accepted and the widget cleared — not that the
 model's first answer was right.
 
 | Browser | Solved | Rounds billed | Median wall clock |
 |---|---:|---:|---:|
 | [Camoufox](https://camoufox.com) 0.4.11 | **3 / 3** | 21 | 96 s |
-| Holo 152.0.3 | **1 / 3** | 21 | 85 s |
 
 A solve took 5–8 model rounds, because reCAPTCHA replaces tiles after each
 click and every replacement is a fresh puzzle. That is the whole reason the
 hosted API meters per round rather than per solve.
 
-**Solve rate in a browser is not model accuracy, and the gap is mostly your
-IP.** reCAPTCHA rejects correct answers from addresses it distrusts, and both
-runs above came from the same datacenter address in the same hour — which is
-the least favourable condition there is, and the likeliest explanation for the
-difference between the two rows. Neither figure is a benchmark; they are proof
-the path works. See
+**Solve rate in a browser is not model accuracy, and much of it is your IP.**
+reCAPTCHA rejects correct answers from addresses it distrusts, and the run above
+came from a datacenter address — which is the least favourable condition there
+is. The figure is not a benchmark; it is proof the path works. See
 [Rate limiting & IP reputation](./docs/performance.md#rate-limiting--ip-reputation).
 
 Reproduce it yourself — the demos in this repo drive a real browser end to end:
@@ -392,7 +389,7 @@ Most of the detail lives in the docs hub — start at **[docs/](./docs/README.md
 | 🚀 [Usage](./docs/usage.md) | Install, the 4 browser frameworks, the Python CLI, migrating from v1 |
 | ⚙️ [How it works](./docs/how-it-works.md) | The solve pipeline, `find_grid`, the freshness guard, dedup |
 | 📊 [Performance](./docs/performance.md) | Accuracy, speed-by-device tables, IP-reputation & rate limits |
-| 🗺️ [Roadmap](./docs/roadmap.md) | Video support, more captcha types, and what shipped |
+| 🗺️ [Roadmap](./docs/roadmap.md) | What shipped, what's in progress, and what's planned |
 | 📜 [Licensing](./docs/licensing.md) | Plain-English: what you can and can't build |
 
 ---

diff --git a/docs/README.md b/docs/README.md
@@ -6,7 +6,7 @@ guides. Start here.
 > One repo, two published ports — the TypeScript browser driver
 > (npm: `captchakraken`) and the Python engine (PyPI: `captchakraken`). ⭐ **Star
 > & watch** [the repo](https://github.com/JWriter20/CaptchaKraken) for smaller
-> models, video support, and new puzzle types.
+> models and new puzzle types.
 
 ## 📚 Guides
 
@@ -17,7 +17,7 @@ guides. Start here.
 | [🧑‍💻 Usage](./usage.md) | Install both ports, the Python one-liner, all four browser frameworks (Playwright / Patchright / camoufox-js / Puppeteer), and migrating from v1. |
 | [⚙️ How it works](./how-it-works.md) | The detect → grid → click → verify pipeline, the OpenCV grid tracer, the stale-frame freshness guard, and solution dedup. |
 | [📊 Performance](./performance.md) | Model accuracy, the memory-bandwidth speed model + per-device throughput, and IP-reputation guidance. |
-| [🗺️ Roadmap](./roadmap.md) | What shipped recently, what's in progress, and what's planned (video support, more captcha types). |
+| [🗺️ Roadmap](./roadmap.md) | What shipped recently, what's in progress, and what's planned (more captcha types). |
 | [⚖️ Licensing](./licensing.md) | Plain-English explainer of the source-available license — what you may and may not build. |
 
 ## ✅ What it solves today
@@ -32,7 +32,7 @@ CaptchaKraken detects the captcha, solves it, clicks, and verifies.
 | ✅ **hCaptcha 3×3 image grid** | Works end-to-end |
 | ✅ **hCaptcha click / drag puzzles** | Full-puzzle model → pixel click/drag actions |
 | ✅ Cloudflare Turnstile | Works via the checkbox flow |
-| 🟡 **Animated / video challenges** | Driver support ships: recorded, cut into keyframes, solved as a multi-image prompt, then clicked once the widget returns to the chosen frame. Accuracy depends on the adapter; the one `setup.sh` installs skips them ([roadmap](./roadmap.md)) |
+| ✅ **Animated / video challenges** | Recorded (4 s @ 10 fps), cut into keyframes, solved as one multi-image prompt, then clicked once the widget returns to the chosen frame. The model half shipped with **v1.2**, which `setup.sh` installs and the hosted API serves |
 
 Non-grid still-image puzzles — **click** ("click each …"), **drag** ("drag the
 piece into place"), path/connect, "choose the card" — route to the full-puzzle

diff --git a/docs/performance.md b/docs/performance.md
@@ -87,12 +87,12 @@ any harness that trusts the file naively:
 2. **`rows`/`cols` disagree with `puzzle_type` on those same records.** Trust
    `puzzle_type`; it is the only field that agrees with the answer.
 
-**Browser solve rates are a different measurement.** In live runs against
-`google.com/recaptcha/api2/demo` on 2026-07-27, Camoufox cleared 3/3 challenges
-and Holo 1/3, each taking 5–8 model rounds — reCAPTCHA replaces tiles after
-every click, and each replacement is a fresh puzzle. Both runs came from one
-datacenter IP within one hour, which is the least favourable condition
-available; see [below](#rate-limiting--ip-reputation).
+**Browser solve rates are a different measurement.** In a live run against
+`google.com/recaptcha/api2/demo` on 2026-07-27, Camoufox cleared 3/3 challenges,
+taking 5–8 model rounds — reCAPTCHA replaces tiles after every click, and each
+replacement is a fresh puzzle. The run came from a datacenter IP, which is the
+least favourable condition available; see
+[below](#rate-limiting--ip-reputation).
 
 ## A note on speed
 

diff --git a/docs/roadmap.md b/docs/roadmap.md
@@ -26,7 +26,7 @@ Where CaptchaKraken is headed. Legend: 🟢 shipped · 🟡 in progress · ⚪ p
 | ⬛ **Abyss** | 🟡 in progress | The next hosted-only model, trained against the open weights' measured failures. **Not serving yet** — the endpoint answers with Twilight v1.2 until it lands. |
 | 🪶 **Sunlight / Twilight merges** | 🟢 shipped | The adapter merged into the base at 4-bit (~9 GB) and 8-bit (~14 GB), so self-hosting is one download instead of two. Published for both v1.1 and v1.2, all public on [HuggingFace](https://huggingface.co/CaptchaKraken). |
 | 📈 **More real labeled data** | 🟡 in progress | Broader coverage for under-represented prompts. |
-| 🎥 **Video challenge support** | 🟢 shipped | **The driver half has shipped.** A challenge that never settles is now recorded (4 s @ 10 fps), cut into keyframes, and sent to the model as one multi-image prompt; the answer names which keyframe it acted on, and the driver waits for the widget to return to that frame before clicking. The model half shipped with **v1.2**, which is trained on the keyframe format. |
+| 🎥 **Video challenge support** | 🟢 shipped | **Both halves are out.** A challenge that never settles is recorded (4 s @ 10 fps), cut into keyframes, and sent to the model as one multi-image prompt; the answer names which keyframe it acted on, and the driver waits for the widget to return to that frame before clicking. The model half shipped with **v1.2** — trained on the keyframe format, installed by `setup.sh`, and what the hosted API answers with. |
 
 ## ⚪ Planned