Best-effort HTML → Markdown conversion via regex.

Analyze convergence data for multiple methods. Parameters ---------- convergence_data : dict Maps method name to list of {"h": ..

Estimate what fraction of the text is template/placeholder content. Returns 0.0 (fully original) to 1.0 (fully template). Simple heuristic: c

Return headers that signal session completion. Call this when the pipeline run finishes to trigger MetaClaw's post-session processing

Return cached results or None if miss/expired. If *ttl* is not provided, uses source-specific TTL from ``_SOURCE_TTL``, falling back to the g

Check if a domain profile represents an ML/AI domain.

Parse BibTeX text into a list of field dicts. Each dict contains at least ``key`` and ``type``, plus any parsed fields (``title``, ``author``

Increment step counter. Call this once per experiment step.

Generate a BibTeX entry string. If ``_bibtex_override`` was populated (e.g. from CrossRef), return that directly. Otherwise construc

Compute overall novelty score and assessment. Returns (score, assessment) where score is 0-1 (higher = more novel) and assessment is 'high' |

Extract the best available snippet from an Exa result. Cascades through highlights -> summary -> text, returning the first non-empty value fo

Collect content authenticity metrics from stage outputs.

Escape LaTeX special characters in plain text (titles, headings). Does NOT escape inside math delimiters or \\commands.

Jaccard similarity between two keyword lists.

Parse ``name: value`` metric lines from experiment stdout. Handles formats like ``UCB (Stochastic) cumulative_regret: 361.9233`` and simple `

Apply academic styling via rcParams.

Return True if the Docker daemon is reachable.

Compute convergence order via log-log linear regression. Parameters ---------- h_values : list[float] Grid sizes / timesteps (mus

Return top-level module names imported by *code*. Returns an empty set if the code can't be parsed.

Read checkpoint and return the NEXT stage to execute, or None if no checkpoint.

Return the raw system prompt template for *stage*.

Serialise to a plain dict for JSON/JSONL output.

Scan *code* AST for dangerous calls and imports.

Build targeted search queries from topic and hypotheses.

Build the ``docker run`` command list. The container uses ``entrypoint.sh`` which handles: Phase 0: pip install requirements.txt

Send a chat message and return the content string.

Classify an error into a LessonCategory based on keywords.

Combined similarity score between hypotheses keywords and a paper.

Remove duplicates. Priority: DOI > arXiv ID > fuzzy title. When a duplicate is found, the entry with higher citation_count wins (i.e. Semant

Detect NVIDIA GPU via nvidia-smi.

Extract multi-file code blocks from LLM output.

Extract patterns using regex heuristics (no LLM needed).

Generate queries without LLM using keyword extraction.

Return domain IDs whose keywords appear in the topic.

Reset circuit breaker state (for tests).

Replace unverified numerical data in markdown tables with '---'. Loads experiment_summary.json as ground truth, extracts all verified metric

Save intermediate artifact to stage directory.

Skip discussion and proceed directly to S8 for a single agent.

Compute exponential decay weight for a lesson based on age. Uses 30-day half-life: weight = exp(-age_days * ln(2) / 30). Returns 0.0 for less

Return True if *image* exists locally (does NOT pull).

Check if PyTorch is importable; attempt install if not. Returns True if torch is available after this call.

Convert a legacy exp_plan.yaml (baselines/proposed/ablations) to the universal format. This allows existing ML experiment plans to work with

Extra context injected into the blueprint generation prompt. Includes domain-specific file structure guidance, library hints, etc.

Load all YAML profiles from the profiles directory.

Assign pending tasks to idle agents (FIFO pull).

Check if PRM gating is configured for this stage.

Run all validations and return a combined :class:`CodeValidation`. 1. Syntax check (always) 2. Security scan (unless *skip_security*) 3.

Validate that *entry_point* resolves inside *staging*. Returns an error message if invalid, ``None`` if valid. Call this **after** copying fi

Check that all imported modules are available. *available* defaults to ``SAFE_STDLIB | COMMON_SCIENCE`` plus any modules already in ``sys.mod

Search for a paper by title and verify its existence. Uses the unified ``search_papers`` function from our literature module. Returns *None*

Write a single KB entry to the appropriate category directory. Returns the path to the written file.

Write stage results to the knowledge base. Reads the primary output artifact and creates a KB entry in the appropriate category directory.

Convert all non-title/abstract sections to LaTeX body text. When a paper has its title as an H1 heading (``# My Paper Title``), that heading

Build a paper draft with default prose sections.

Build a Gemini prompt for academic figure generation.

Build the base ssh command with common options. *extra_opts* are inserted **before** the hostname so that SSH interprets them as SSH options,

Send a chat message expecting JSON output. Falls back to regex extraction.

Collect generated Python files, requirements.txt, and setup.py. Only collects files that are **new or modified** relative to the code

Lightweight fallback: fetch HTML and strip tags.

Scan Python files for import statements and return pip package names.

Enforce min/max figure counts and required categories.

Execute script inside an isolated Docker container. Security measures: - ``--network none``: No network access (prevents data exfiltr

Filter benchmarks by tier, size, and network policy.

Make a GET request to the GitHub API.

Generate figure decisions without LLM (rule-based fallback).

Run a single aider invocation in the workspace.

Parse a scholarly publication object into ScholarPaper.

Parse the LLM response into a list of skill dicts.

Create a temporary workspace directory with context files.