r"""Find the character position at the END of a \section{heading}'s content. Returns the position just before the next \section{} or bibliogr

Fix known SLURM script issues that LLMs commonly produce.

Fix invalid JSON escape sequences produced by LaTeX content. LLM outputs often contain raw LaTeX like \\cite{}, \\textbf{}, \\frac{} inside J

Auto-fix ``\\begin{X}...\\end{Y}`` mismatches in LaTeX source. Common LLM errors: ``\\begin{equation}...\\end{parameter}``, ``\\begin

When quick-eval times out, apply deterministic speed-ups to main.py. Instead of sending a vague "timeout" to the LLM for batch-fix, we apply

Format code files into a single context string.

Keep generated captions/prose from exposing raw float precision.

Format downloaded resource paths for inclusion in prompts.

Fallback for failed figure generation. Data charts are marked failed rather than filled with made-up values. Conceptual figures get a

Generate a single source file.

LLM generates the next iteration hypothesis from feedback.

Return result data from the best round, or the last round as fallback.

Query SLURM for job status.

Insert a LaTeX preamble snippet before ``\\begin{document}``.

Escape LaTeX-special chars in table cells (headers and method names).

Least-squares slope (no numpy needed).

Load iteration checkpoint if available. Returns (state, start_round) where start_round is the round to resume from (1 if no checkpoin

Load previously saved output for a completed stage.

Run local checks before submitting to SLURM. Tests: 1. Python syntax check (py_compile) on all .py files 2. Import check -- t

Poll SLURM until job completes. Returns final status.

Parse LLM response into edit objects.

Extract structured sections from RAM output text.

Deterministic conda env name for the current session. Uses ``nanoresearch_{sanitized_label}`` so that resume reuses the same env (ide

r"""Run deterministic layout hygiene before LaTeX compilation. The goal is to make the first compiled PDF reasonable: compact floats,

Pretty-print the result of an edit operation.

Print health check results as a table.

Pretty-print paper edit result.

r"""Move non-architecture figures out of the Introduction section. Ablation, results, and training-curve figures don't belong in Intr

r"""Move any \begin{figure}...\end{figure} blocks that appear after \bibliography back to their proper positions in the document body.

Attempt to repair JSON that was truncated by output token limit. Strategy: close any open strings, arrays, and objects from the end.

Attempt to repair JSON that was truncated mid-output. Handles common truncation patterns: unterminated strings, missing closing brack

Resolve a user-supplied environment spec to a Python executable. Accepts three formats: 1. Direct path to python executable (file ex

Review a single section of the paper. Returns a SectionReview with detailed feedback including strengths (to be preserved during revi

Revise a single section based on reviewer feedback. Uses the WRITING-stage LLM (typically a strong generation model) rather than the

When patching fails, ask LLM to rewrite the entire file.

Parse JSON from LLM response, stripping markdown fences.

Fix common Unicode issues in BibTeX entries and deduplicate.

Prepare a Python environment for experiment execution. Reuses a configured or already-compatible local environment first, and only cr

Compact metric labels for paper tables.

Show paper overview: sections, figures, review score.

Split a list of requirement specifiers into torch-family and non-torch. Returns (torch_specs, other_specs). torch_specs are raw specifier string

Bound tool output before it re-enters the LLM context.

Verify generated Python files have valid syntax via compile().

Verify that torch imports and CUDA is available post-install.

Run the ReAct agent loop to fulfill the user's instruction.

Bootstrap confidence interval (no scipy needed).

Execute a registered tool by name. Validates required parameters before calling the handler, and strips unexpected parameters to prev

Check all files for import mismatches. Returns list of mismatch dicts. Each dict contains: importer: filename that has the bad im