Rename imports to rclc and simplify the CacheBack page

.github/workflows/ci.yml

@@ -36,7 +36,7 @@ jobs:
           .venv/bin/python -m pip install torch --index-url https://download.pytorch.org/whl/cpu
           .venv/bin/python -m pip install ".[dev]" "transformers==${{ matrix.transformers }}"
           .venv/bin/python -m pip check
-          .venv/bin/python -c 'import rcc, sys; from pathlib import Path; assert Path(rcc.__file__).is_relative_to(Path(sys.prefix))'
+          .venv/bin/python -c 'import rclc, sys; from pathlib import Path; assert Path(rclc.__file__).is_relative_to(Path(sys.prefix))'
           scripts/install-hooks.sh
       - name: Harness
         run: .venv/bin/python scripts/check.py

CONTRIBUTING.md

@@ -9,9 +9,9 @@ the same quality.
 
 Write a callable that follows the
 [selector contract](docs/selectors.md#custom-selection) and pass it to
-`rcc.transfer(..., selector=my_selector)`.
+`rclc.transfer(..., selector=my_selector)`.
 
-For a contribution to the package, put the function in `src/rcc/selectors/`.
+For a contribution to the package, put the function in `src/rclc/selectors/`.
 Keep it small and leave sender state unchanged. Extend an existing end-to-end
 scenario to cover selection, delivery and receiver continuation; avoid a test
 per helper. Follow the [development harness](docs/development.md) and run:

README.md

@@ -1,9 +1,9 @@
-# RCLC
+# CacheBack
 
 **Receiver-conditioned latent communication for agents.**
 
 Code for *Receiver-Conditioned Latent Communication gives 94% CacheBack*.
-**[Website](https://maxr0ssi.github.io/rclc/)** · **[Paper](https://arxiv.org/abs/2609.32046)** · [Python API](https://github.com/maxr0ssi/rclc/blob/main/docs/api.md) · [Paper replication](https://github.com/maxr0ssi/rclc/blob/main/docs/replication.md)
+**[Website](https://agentcacheback.github.io/)** · **[Paper](https://arxiv.org/abs/2609.32046)** · [Python API](https://github.com/agentcacheback/cacheback/blob/main/docs/api.md) · [Paper replication](https://github.com/agentcacheback/cacheback/blob/main/docs/replication.md)
 
 Agents distribute large contexts across senders and receivers. Text messages
 require decoding and can omit evidence. Full KV-cache messages accumulate every
@@ -12,17 +12,17 @@ sender's context at the receiver, increasing memory use and context length.
 in a request. **CacheBack** is a training-free selector that uses attention to
 that request to choose which sender positions enter the handoff.
 
-![FanOutQA accuracy and latency across four model families, alongside receiver-conditioned selection of sender state.](https://raw.githubusercontent.com/maxr0ssi/rclc/main/docs/assets/cacheback-overview.png)
+![FanOutQA accuracy and latency across four model families, alongside receiver-conditioned selection of sender state.](https://raw.githubusercontent.com/agentcacheback/cacheback/main/docs/assets/cacheback-overview.png)
 
 *Highest-accuracy CacheBack setting per family versus same-size text on FanOutQA,
 with 50 concurrent tasks on eight H100 GPUs. Operating points differ by family.
 Right: the receiver query guides which sender positions enter the handoff.*
 
 ## Watch the demo
 
-**[▶ Watch the 36-second coding demo](https://maxr0ssi.github.io/rclc/#demo)**
+**[▶ Watch the 36-second coding demo](https://agentcacheback.github.io/#demo)**
 
-[![CacheBack finishes a Django fix while text workers are still writing reports.](demo/coding/poster.jpg)](https://maxr0ssi.github.io/rclc/#demo)
+[![CacheBack finishes a Django fix while text workers are still writing reports.](demo/coding/poster.jpg)](https://agentcacheback.github.io/#demo)
 
 Seven Qwen3-8B workers help a coordinator fix a Django bug. CacheBack completes
 this recorded case in **25.66 s**, versus **113.21 s** for text: **4.41× faster**.
@@ -31,13 +31,13 @@ test grading are excluded from the clocks. The silent video reconstructs the
 interface from separate runs and varies playback speed. This is one case,
 not an aggregate benchmark. [Recorded results](demo/coding/evidence.json).
 
-Also explore the [interactive booking relay](https://maxr0ssi.github.io/rclc/demo/)
+Also explore the [interactive booking relay](https://agentcacheback.github.io/demo/)
 to inspect selected positions and text messages at each handoff. Its
 [demo guide](demo/README.md) covers recording your own.
 
 ## Install
 
-The distribution is named `rclc`; the Python import is `rcc`. Until the PyPI
+Install and import the package as `rclc`. Until the PyPI
 release, install from a checkout:
 
 ```bash
@@ -54,21 +54,21 @@ python -m pip install .
 ```
 
 The engine needs the capture connector; see the
-[vLLM setup](https://github.com/maxr0ssi/rclc/blob/main/docs/api.md#vllm-default) and
-[integration example](https://github.com/maxr0ssi/rclc/blob/main/examples/vllm_transfer.py).
+[vLLM setup](https://github.com/agentcacheback/cacheback/blob/main/docs/api.md#vllm-default) and
+[integration example](https://github.com/agentcacheback/cacheback/blob/main/examples/vllm_transfer.py).
 vLLM 0.26.0 is an explicit opt-in. Both pins passed a Qwen3-8B GPU journey on an
-A100 ([results](https://github.com/maxr0ssi/rclc/blob/main/docs/development.md#gpu-validation)).
+A100 ([results](https://github.com/agentcacheback/cacheback/blob/main/docs/development.md#gpu-validation)).
 
 ## Use it
 
 Bind existing Hugging Face Qwen3 agents by their loaded model and tokenizer:
 
 ```python
-import rcc
+import rclc
 
-agent_x = rcc.bind(model, tokenizer, messages=history_x, backend="hf")
-agent_y = rcc.bind(model, tokenizer, messages=history_y, backend="hf")
-await rcc.transfer(agent_x, agent_y, "Who owns Cedar?")
+agent_x = rclc.bind(model, tokenizer, messages=history_x, backend="hf")
+agent_y = rclc.bind(model, tokenizer, messages=history_y, backend="hf")
+await rclc.transfer(agent_x, agent_y, "Who owns Cedar?")
 
 inputs = agent_y.pop()
 output = model.generate(**inputs, return_dict_in_generate=True)
@@ -82,53 +82,53 @@ supported.
 
 | Supply state with | Use when |
 | --- | --- |
-| `messages=history` | You have chat messages; RCC applies the chat template. |
+| `messages=history` | You have chat messages; RCLC applies the chat template. |
 | `prompt=text_or_ids` | You have rendered text or exact token IDs. |
 | `prompt=ids, past_key_values=cache` | You already have an HF cache. |
 | `generation=output` | You have a native HF generation result. |
 | `prompt=ids, request_id=capture_id` | You already captured the vLLM prompt. |
 | Neither | Receive first, or supply sender state later with `update`. |
 
-- Outside an async function or notebook, use `rcc.transfer_sync(...)` with the
+- Outside an async function or notebook, use `rclc.transfer_sync(...)` with the
   same options. Native generation and `update` block; finish each operation
   before reusing the model
-  ([concurrency](https://github.com/maxr0ssi/rclc/blob/main/docs/api.md#concurrency)).
-- vLLM: `rcc.bind(llm, backend="vllm", messages=history)` with the
-  [engine setup](https://github.com/maxr0ssi/rclc/blob/main/docs/api.md#vllm-default).
+  ([concurrency](https://github.com/agentcacheback/cacheback/blob/main/docs/api.md#concurrency)).
+- vLLM: `rclc.bind(llm, backend="vllm", messages=history)` with the
+  [engine setup](https://github.com/agentcacheback/cacheback/blob/main/docs/api.md#vllm-default).
   Local dense Qwen3 weights are required; hosted chat APIs are not supported.
 - `await agent_y.append(new_text_or_ids)` adds a tool result without losing
   received state; `await agent_y.inputs()` continues locally; `save`/`load`
   snapshot retained state
-  ([continuation and snapshots](https://github.com/maxr0ssi/rclc/blob/main/docs/api.md#continue-append-and-resume)).
+  ([continuation and snapshots](https://github.com/agentcacheback/cacheback/blob/main/docs/api.md#continue-append-and-resume)).
 - `agent_y.inspect()` shows queued handoffs; `record_selection=True` plus
   `agent_y.selection_html()` shows what survived compression
-  ([selection inspection](https://github.com/maxr0ssi/rclc/blob/main/docs/api.md#inspect-check-and-view-selection)).
-- `python -m rcc doctor` (`--backend vllm` for CUDA) or `rcc.check(agent_y)`
+  ([selection inspection](https://github.com/agentcacheback/cacheback/blob/main/docs/api.md#inspect-check-and-view-selection)).
+- `python -m rclc doctor` (`--backend vllm` for CUDA) or `rclc.check(agent_y)`
   checks setup without running a model or downloading weights.
 
-Examples: [quickstart](https://github.com/maxr0ssi/rclc/blob/main/examples/quickstart.py),
-[cached conversation](https://github.com/maxr0ssi/rclc/blob/main/examples/existing_agents.py),
-[Colab GPU notebook](https://github.com/maxr0ssi/rclc/blob/main/examples/colab.ipynb).
+Examples: [quickstart](https://github.com/agentcacheback/cacheback/blob/main/examples/quickstart.py),
+[cached conversation](https://github.com/agentcacheback/cacheback/blob/main/examples/existing_agents.py),
+[Colab GPU notebook](https://github.com/agentcacheback/cacheback/blob/main/examples/colab.ipynb).
 
 ### Options
 
 ```python
 from functools import partial
 
-await rcc.transfer(senders, receivers, requests)               # Relative r4 (default)
-await rcc.transfer(senders, receivers, requests, ratio=16)     # Keep 1/16 of new positions
-await rcc.transfer(senders, receivers, requests, budget=1024)  # Bounded message size
-await rcc.transfer(senders, receivers, requests, selector=rcc.selectors.qsnap)
-await rcc.transfer(senders, receivers, requests,
-                   reasoning=partial(rcc.latent_mass, steps=50), budget=128)
+await rclc.transfer(senders, receivers, requests)               # Relative r4 (default)
+await rclc.transfer(senders, receivers, requests, ratio=16)     # Keep 1/16 of new positions
+await rclc.transfer(senders, receivers, requests, budget=1024)  # Bounded message size
+await rclc.transfer(senders, receivers, requests, selector=rclc.selectors.qsnap)
+await rclc.transfer(senders, receivers, requests,
+                   reasoning=partial(rclc.latent_mass, steps=50), budget=128)
 ```
 
 Each argument is one item or a list: two senders, two receivers and two requests
 produce four deliveries of two messages each. See
-[budgets](https://github.com/maxr0ssi/rclc/blob/main/docs/api.md#relative-and-bounded-budgets),
-[selectors](https://github.com/maxr0ssi/rclc/blob/main/docs/selectors.md),
-[reasoning](https://github.com/maxr0ssi/rclc/blob/main/docs/api.md#reasoning) and
-[representations](https://github.com/maxr0ssi/rclc/blob/main/docs/api.md#representations).
+[budgets](https://github.com/agentcacheback/cacheback/blob/main/docs/api.md#relative-and-bounded-budgets),
+[selectors](https://github.com/agentcacheback/cacheback/blob/main/docs/selectors.md),
+[reasoning](https://github.com/agentcacheback/cacheback/blob/main/docs/api.md#reasoning) and
+[representations](https://github.com/agentcacheback/cacheback/blob/main/docs/api.md#representations).
 
 ## Results in the paper
 
@@ -142,10 +142,10 @@ the paper's benchmark results, not performance claims for this package.
 ## Bring your own selector or benchmark
 
 Pass any callable as `selector=` and compare it with CacheBack using the
-[small evaluation script](https://github.com/maxr0ssi/rclc/blob/main/CONTRIBUTING.md#2-run-a-small-comparison).
+[small evaluation script](https://github.com/agentcacheback/cacheback/blob/main/CONTRIBUTING.md#2-run-a-small-comparison).
 The paper's **FanOutQA** and **LongBench v2** panels need separate integration.
 New benchmarks and negative results are welcome; see
-[Contributing](https://github.com/maxr0ssi/rclc/blob/main/CONTRIBUTING.md).
+[Contributing](https://github.com/agentcacheback/cacheback/blob/main/CONTRIBUTING.md).
 
 ## Branches
 
@@ -155,9 +155,9 @@ New benchmarks and negative results are welcome; see
 | `paper` | Four-family runtimes, both benchmarks, twelve experiment configs, pinned serving stacks, data bundles |
 
 The `paper-v1` tag fixes the replication snapshot; see the
-[replication guide](https://github.com/maxr0ssi/rclc/blob/main/docs/replication.md).
+[replication guide](https://github.com/agentcacheback/cacheback/blob/main/docs/replication.md).
 For checks, tests and GPU validation, see the
-[development guide](https://github.com/maxr0ssi/rclc/blob/main/docs/development.md).
+[development guide](https://github.com/agentcacheback/cacheback/blob/main/docs/development.md).
 
 ## License and citation
 
@@ -169,6 +169,8 @@ Apache-2.0. Data licenses and replication requirements are documented on the
   title  = {Receiver-Conditioned Latent Communication gives 94\% CacheBack},
   author = {Rossi, Maximillian and Raghunath, Prajwal and Xuan, Haoqing and Zhang, Yusen and Wu, Eugene},
   year   = {2026},
-  note   = {Preprint}
+  eprint = {2609.32046},
+  archivePrefix = {arXiv},
+  url    = {https://arxiv.org/abs/2609.32046}
 }
 ```

demo/README.md

@@ -10,7 +10,7 @@ Harbour Hub, whose passport requirement overrides the original desk's ID advice.
 The [booking section](../index.html#booking-demo), below the main coding video, presents the recorded final answers
 and completion times before the replay. Selected positions use teal highlights.
 
-Open <https://maxr0ssi.github.io/rclc/demo/> and press **Run**. To serve it
+Open <https://agentcacheback.github.io/demo/> and press **Run**. To serve it
 locally, run this from the repository root and open <http://127.0.0.1:8765/demo/>:
 
 ```bash
@@ -50,7 +50,7 @@ animation.
 
 ## Record
 
-Install RCC as described in the [repository README](../README.md#install), then
+Install RCLC as described in the [repository README](../README.md#install), then
 run:
 
 ```bash
@@ -70,12 +70,12 @@ GPU with a high-RAM host. Prepare its source bundle from a committed checkout:
 
 ```bash
 mkdir -p output/colab
-git archive --format=zip --output=output/colab/rcc-demo-colab-source.zip HEAD
+git archive --format=zip --output=output/colab/rclc-demo-colab-source.zip HEAD
 ```
 
 Upload that ZIP when prompted, mount Drive and run the cells in order. Each run
 saves its source, configuration, environment, complete logs and traces to a new
-`MyDrive/rcc-demo/<run-id>/` folder. The notebook retains its width comparison
+`MyDrive/rclc-demo/<run-id>/` folder. The notebook retains its width comparison
 for analysis; the viewer shows W=4 only. The last cell downloads a portable
 replay; serve its extracted directory with the command above. Replay needs no
 model.
@@ -88,7 +88,7 @@ handoff alone. Later agents cannot reread discarded evidence.
 
 Text generates new notes; each text column is the actual message generated at
 that step, not a summary written for the viewer. RCLC sends selected input
-embeddings through `rcc.transfer_sync`, with no extra latent steps. Highlights
+embeddings through `rclc.transfer_sync`, with no extra latent steps. Highlights
 follow exact tokenizer offsets, including word fragments. As another document
 arrives, selection can drop evidence retained at the previous handoff.
 

demo/colab.ipynb

@@ -5,7 +5,7 @@
    "id": "cell-0",
    "metadata": {},
    "source": [
-    "# Record the RCC relay with Qwen3-8B\n",
+    "# Record the RCLC relay with Qwen3-8B\n",
     "\n",
     "This notebook runs the existing demo recorder at **W=16 and W=4**, with the same\n",
     "r4 budget, questions and source documents. It preserves unedited answers,\n",
@@ -18,16 +18,16 @@
     "   **Runtime > Change runtime type**. An L4 or A100 is the intended target.\n",
     "   This is an unquantized BF16/FP16 run. The notebook checks GPU memory before\n",
     "   downloading weights; a 16 GB T4 is too small for this configuration.\n",
-    "2. Upload the separate `rcc-demo-colab-source.zip` when prompted, then allow\n",
+    "2. Upload the separate `rclc-demo-colab-source.zip` when prompted, then allow\n",
     "   Drive mounting. This does not require a public repository or a GitHub token.\n",
     "3. Run the remaining cells in order. Results go to a unique directory under\n",
-    "   `MyDrive/rcc-demo/`. The final cell downloads a portable replay ZIP.\n",
+    "   `MyDrive/rclc-demo/`. The final cell downloads a portable replay ZIP.\n",
     "\n",
-    "Build the source archive from the RCC repository after committing your changes:\n",
+    "Build the source archive from the RCLC repository after committing your changes:\n",
     "\n",
     "```bash\n",
     "mkdir -p output/colab\n",
-    "git archive --format=zip --output=output/colab/rcc-demo-colab-source.zip HEAD\n",
+    "git archive --format=zip --output=output/colab/rclc-demo-colab-source.zip HEAD\n",
     "```\n",
     "\n",
     "Colab hardware availability and session lifetime vary; see the\n",
@@ -86,22 +86,22 @@
     "    )\n",
     "    uploaded = files.upload()\n",
     "    if len(uploaded) != 1:\n",
-    "        raise ValueError(\"Upload exactly one RCC source ZIP.\")\n",
+    "        raise ValueError(\"Upload exactly one RCLC source ZIP.\")\n",
     "    SOURCE_ZIP = Path(next(iter(uploaded)))\n",
     "    drive.mount(\"/content/drive\")\n",
-    "    OUTPUT_ROOT = Path(\"/content/drive/MyDrive/rcc-demo\")\n",
+    "    OUTPUT_ROOT = Path(\"/content/drive/MyDrive/rclc-demo\")\n",
     "    DEVICE = \"cuda\"\n",
     "\n",
     "run_id = datetime.now(timezone.utc).strftime(\"%Y%m%dT%H%M%SZ\") + \"-\" + uuid.uuid4().hex[:8]\n",
     "RUN = OUTPUT_ROOT / run_id\n",
     "RUN.mkdir(parents=True)\n",
-    "SOURCE = Path(tempfile.mkdtemp(prefix=\"rcc-demo-source-\"))\n",
+    "SOURCE = Path(tempfile.mkdtemp(prefix=\"rclc-demo-source-\"))\n",
     "with zipfile.ZipFile(SOURCE_ZIP) as archive:\n",
     "    if any(Path(name).is_absolute() or \"..\" in Path(name).parts for name in archive.namelist()):\n",
     "        raise ValueError(\"Source ZIP contains an unsafe path.\")\n",
     "    archive.extractall(SOURCE)\n",
     "if not (SOURCE / \"demo/record.py\").is_file() or not (SOURCE / \"pyproject.toml\").is_file():\n",
-    "    raise ValueError(\"Expected a git archive of the RCC repository root.\")\n",
+    "    raise ValueError(\"Expected a git archive of the RCLC repository root.\")\n",
     "shutil.copyfile(SOURCE_ZIP, RUN / \"source.zip\")\n",
     "(RUN / \"config.json\").write_text(\n",
     "    json.dumps(\n",

demo/index.html

@@ -93,7 +93,7 @@ <h1 id="question">Where should Maya collect her pass on Friday, and what ID shou
     <p class="sr-only" id="status" role="status" aria-live="polite"></p>
     <footer>
       <p><span id="run-info">Recorded replay</span> · <span id="timing-note">Estimated streaming, measured completion times.</span></p>
-      <a href="https://github.com/maxr0ssi/rclc/tree/main/demo">Code &amp; recording</a>
+      <a href="https://github.com/agentcacheback/cacheback/tree/main/demo">Code &amp; recording</a>
     </footer>
   </main>
   <script>

demo/record.py

@@ -17,8 +17,8 @@
 import transformers
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-from rcc import Delivery, SenderState, transfer_sync
-from rcc.selectors import cacheback
+from rclc import Delivery, SenderState, transfer_sync
+from rclc.selectors import cacheback
 
 DOCUMENTS = [
     {