```mermaid
graph LR
    eval_inference["eval: inference"]
    topogpt3_model["topogpt3: model"]
    topogpt3_inference_hrm["topogpt3: inference_hrm"]
    topogpt3_tasks["topogpt3: tasks"]
    topogpt3_jlens["topogpt3: jlens"]
    eval_data_pipeline["eval: data_pipeline"]
    eval_inference -- "calls into" --> topogpt3_model
    eval_inference -- "calls into" --> topogpt3_inference_hrm
    eval_inference -- "depends on" --> topogpt3_jlens
    eval_inference -- "calls into" --> eval_data_pipeline
    topogpt3_model -- "calls into" --> eval_inference
    topogpt3_model -- "calls into" --> topogpt3_inference_hrm
    topogpt3_inference_hrm -- "calls into" --> topogpt3_model
    topogpt3_inference_hrm -- "calls into" --> topogpt3_tasks
    topogpt3_tasks -- "calls into" --> topogpt3_model
    topogpt3_tasks -- "calls into" --> topogpt3_inference_hrm
    topogpt3_tasks -- "calls into" --> eval_data_pipeline
    topogpt3_jlens -- "depends on" --> eval_inference
    topogpt3_jlens -- "calls into" --> topogpt3_model
    topogpt3_jlens -- "depends on" --> topogpt3_inference_hrm
    click eval_inference href "./eval_inference.md" "Details"
    click topogpt3_model href "./topogpt3_model.md" "Details"
    click topogpt3_inference_hrm href "./topogpt3_inference_hrm.md" "Details"
    click topogpt3_tasks href "./topogpt3_tasks.md" "Details"
```

## Details

Static architecture map of 59 files (py, sh) in 6 components: eval: inference, topogpt3: model, topogpt3: inference_hrm, topogpt3: tasks, topogpt3: jlens, eval: data_pipeline. Structure is extracted statically without LLM calls; component prose may be enriched locally.

### eval: inference (1)

eval: inference groups 12 files (py) around load_humaneval, build_prompt, extract_candidate, run_one_test, run_one_test_sandboxed. Drop-in entry point that demonstrates how to use the topogpt3 package.

**Related Classes/Methods**:

- <a href="../eval/harness.py#L59-L74">eval/harness.py::load_humaneval</a>
- <a href="../eval/harness.py#L75-L99">eval/harness.py::build_prompt</a>
- <a href="../eval/harness.py#L100-L149">eval/harness.py::extract_candidate</a>
- <a href="../eval/harness.py#L150-L171">eval/harness.py::run_one_test</a>
- <a href="../eval/harness.py#L172-L194">eval/harness.py::run_one_test_sandboxed</a>
- <a href="../eval/harness.py#L195-L203">eval/harness.py::make_sampler</a>
- <a href="../eval/harness.py#L204-L216">eval/harness.py::completion_for_problem</a>
- <a href="../eval/harness.py#L217-L219">eval/harness.py::ModelLoader</a>

### topogpt3: model (2)

topogpt3: model groups 12 files (py) around TopoGPT2Config, __post_init__, setup_logger, set_seed, QuaternionOps. Diagnostico estatico de un checkpoint TopoGPT3 congelado.

**Related Classes/Methods**:

- <a href="../topogpt3/model.py#L58-L162">topogpt3/model.py::TopoGPT2Config</a>
- <a href="../topogpt3/model.py#L163-L195">topogpt3/model.py::__post_init__</a>
- <a href="../topogpt3/model.py#L196-L205">topogpt3/model.py::setup_logger</a>
- <a href="../topogpt3/model.py#L206-L217">topogpt3/model.py::set_seed</a>
- <a href="../topogpt3/model.py#L218-L225">topogpt3/model.py::QuaternionOps</a>
- <a href="../topogpt3/model.py#L226-L237">topogpt3/model.py::hamilton_product</a>
- <a href="../topogpt3/model.py#L238-L241">topogpt3/model.py::normalize</a>
- <a href="../topogpt3/model.py#L242-L246">topogpt3/model.py::conjugate</a>

### topogpt3: inference_hrm (3)

topogpt3: inference_hrm groups 8 files (py) around _count_unclosed_brackets, _count_unclosed_fences, is_response_complete, extract_tail_for_continuation, split_at_last_newline. Checkpoint architecture discovery tests.

**Related Classes/Methods**:

- <a href="../topogpt3/continuation.py#L25-L35">topogpt3/continuation.py::_count_unclosed_brackets</a>
- <a href="../topogpt3/continuation.py#L36-L44">topogpt3/continuation.py::_count_unclosed_fences</a>
- <a href="../topogpt3/continuation.py#L45-L74">topogpt3/continuation.py::is_response_complete</a>
- <a href="../topogpt3/continuation.py#L75-L104">topogpt3/continuation.py::extract_tail_for_continuation</a>
- <a href="../topogpt3/continuation.py#L105-L114">topogpt3/continuation.py::split_at_last_newline</a>
- <a href="../topogpt3/checkpoint_meta.py#L24-L34">topogpt3/checkpoint_meta.py::read_state</a>
- <a href="../topogpt3/checkpoint_meta.py#L35-L80">topogpt3/checkpoint_meta.py::probe_checkpoint</a>
- <a href="../topogpt3/checkpoint_meta.py#L81-L119">topogpt3/checkpoint_meta.py::config_from_checkpoint</a>

### topogpt3: tasks (4)

topogpt3: tasks groups 7 files (py) around pre_processing_chat, post_processing_chat, _fmt_tool_defs, _tool_calls_text, apply_chat_template. Generic task eval: JSONL cases + safe verifiers.

**Related Classes/Methods**:

- <a href="../topogpt3/chat.py#L44-L57">topogpt3/chat.py::pre_processing_chat</a>
- <a href="../topogpt3/chat.py#L58-L63">topogpt3/chat.py::post_processing_chat</a>
- <a href="../topogpt3/chat.py#L64-L88">topogpt3/chat.py::_fmt_tool_defs</a>
- <a href="../topogpt3/chat.py#L89-L115">topogpt3/chat.py::_tool_calls_text</a>
- <a href="../topogpt3/chat.py#L116-L167">topogpt3/chat.py::apply_chat_template</a>
- <a href="../topogpt3/chat.py#L168-L177">topogpt3/chat.py::parse_tool_calls</a>
- <a href="../topogpt3/chat.py#L178-L184">topogpt3/chat.py::parse_thinking</a>
- <a href="../topogpt3/chat.py#L185-L196">topogpt3/chat.py::split_reasoning_content</a>

### topogpt3: jlens (5)

topogpt3: jlens groups 5 files (py) around LensModel, encode, forward, unembed, TopoGPT3LensConfig. TestValidPositionMask: Feature: valid_position_mask excludes attention-sink and final...

**Related Classes/Methods**:

- <a href="../topogpt3/lens_model.py#L23-L39">topogpt3/lens_model.py::LensModel</a>
- <a href="../topogpt3/lens_model.py#L40-L44">topogpt3/lens_model.py::encode</a>
- <a href="../topogpt3/lens_model.py#L45-L51">topogpt3/lens_model.py::forward</a>
- <a href="../topogpt3/lens_model.py#L52-L58">topogpt3/lens_model.py::unembed</a>
- <a href="../topogpt3/lens_model.py#L59-L83">topogpt3/lens_model.py::TopoGPT3LensConfig</a>
- <a href="../topogpt3/lens_model.py#L84-L103">topogpt3/lens_model.py::from_topogpt2_config</a>
- <a href="../topogpt3/lens_model.py#L104-L141">topogpt3/lens_model.py::probe_checkpoint</a>
- <a href="../topogpt3/lens_model.py#L142-L149">topogpt3/lens_model.py::_TopoGPT3ResidualForward</a>

### eval: data_pipeline (6)

eval: data_pipeline groups 4 files (py) around SandboxConfig, _names_imported, _blocked_dunder_access, _max_depth, d. Sandbox for executing model-generated code during HumanEval evaluation.

**Related Classes/Methods**:

- <a href="../eval/sandbox.py#L53-L99">eval/sandbox.py::SandboxConfig</a>
- <a href="../eval/sandbox.py#L100-L113">eval/sandbox.py::_names_imported</a>
- <a href="../eval/sandbox.py#L114-L122">eval/sandbox.py::_blocked_dunder_access</a>
- <a href="../eval/sandbox.py#L123-L124">eval/sandbox.py::_max_depth</a>
- <a href="../eval/sandbox.py#L125-L132">eval/sandbox.py::d</a>
- <a href="../eval/sandbox.py#L133-L253">eval/sandbox.py::check_safety</a>
- <a href="../eval/sandbox.py#L254-L269">eval/sandbox.py::_build_worker_src</a>
- <a href="../eval/sandbox.py#L270-L372">eval/sandbox.py::safe_exec</a>