This repository develops three independent ICML-style research papers:

• sft_paper: supervised fine-tuning methods;
• rl_paper: reinforcement-learning methods; and
• eval_paper: evaluation methods and benchmarks.

The repository is currently at the methods-first drafting stage. Each paper contains a research question, verified related-work bibliography, detailed method, scoped formal analysis, and preregistered experimental protocol. Empirical result tables are intentionally blank until the corresponding experiments are executed; the drafts are not submission-ready results papers.

The templates use the unmodified official ICML 2026 LaTeX style in template/icml2026. You need a TeX distribution with pdflatex (and ideally latexmk). make prepends /Library/TeX/texbin on macOS and falls back to pdflatex + bibtex if latexmk is not installed.

make

If make reports latexmk: command not found or pdflatex not found, install TeX and refresh PATH:

# macOS, full MacTeX (includes latexmk)
brew install --cask mactex-no-gui
eval "$(/usr/libexec/path_helper)"
export PATH="/Library/TeX/texbin:$PATH"

# macOS, if MacTeX is already installed but the terminal cannot see it
export PATH="/Library/TeX/texbin:$PATH"

# Ubuntu/Debian
sudo apt-get install -y texlive-latex-extra texlive-fonts-recommended \
  texlive-science latexmk

Then retry make. Check which engine was found with make check-deps.

The committed main.pdf files must come from pdfTeX. Tectonic/XeTeX builds substitute Latin Modern for the Times fonts selected by icml2026.sty and replace the running head with "Title Suppressed Due to Excessive Size", so the build script refuses Tectonic unless PAPER_ALLOW_TECTONIC=1 is set for a local preview. Before committing PDFs, run:

sudo apt-get install -y poppler-utils   # or: brew install poppler
make pdf-check

make pdf-check fails unless each PDF was produced by pdfTeX on US-letter paper with only Type 1 fonts, a Times body font, and the paper's own running head.

This builds:

sft_paper/main.pdf
rl_paper/main.pdf
eval_paper/main.pdf

Run the CPU-only equation and protocol audits with:

python3 -m pip install -r requirements.txt
make source-audit
make provenance-audit
make content-gates-audit
make external-overlap-audit
make overlap-projection-audit
make design-audits
make protocol-audits

The source audit validates the committed pinned-metadata snapshot but leaves all manual clearances pending. The provenance audit recomputes row identifiers, raw and projected-content hashes, normalized duplicate representatives, and a closed acquired-to-retained/suppressed/quarantined disposition chain. Rehashed orphan, omission, forged-ID, and wrong-representative fixtures fail; the audit also requires a passing content-readiness ledger. The content-gate audit separately exercises the named PII and secret detectors and deterministic field-segment character-ngram comparisons, including unlike field classes. That lexical screen is a blocking heuristic, not a certificate of semantic independence or manual clearance. Artifacts above the exhaustive comparison limit remain inconclusive. The existing external engine contract covers record pairs only; multi-field exports require a reviewed segment-pair successor before hash-bound shards can clear the blocker. The external-overlap audit checks the generated record-pair contract fixture's engine binding, pair-space coverage, shard hashes, and failure handling, then independently rejects omitted qualifying pairs with a disk-backed exact shared-5-gram join. The factorial audit exercises the registered hypervolume, paired difference-in-differences, missing-run, and simultaneous max-$t$ aggregation rules without using empirical endpoints. The overlap-projection audit checks exact source/role coverage and generated multi-field inventory attestations; its eight real source inventories remain explicitly unreviewed blockers. The remaining design audits reproduce the SFT capability/cluster split contract, the RL estimator, cost aggregation, and three-way reference bootstrap from a generated trajectory ledger, and the evaluation paper's 189-condition fractional simulation layout. All use generated identifiers or vectors and set empirical_evidence to false. The ignored protocol artifacts are synthetic software checks, not paper results. See experiments/README.md for the five-stage execution contract and the boundary between these audits and the planned model/human experiments.

Build or clean one paper with:

make sft_paper/main.pdf
make clean

Each methods-first draft has a machine-readable five-stage plan covering data acquisition, processing, model construction, training, and evaluation. Every plan also hashes its section of experiments/pilot_preregistration.json:

python3 scripts/run_experiment_pipeline.py \
  --plan experiments/sft_pipeline.json
python3 scripts/run_experiment_pipeline.py \
  --plan experiments/rl_pipeline.json
python3 scripts/run_experiment_pipeline.py \
  --plan experiments/eval_pipeline.json

These commands validate and display the plans. Site-local execution commands, required artifacts, cost ledgers, and immutable output hashing are documented in experiments/README.md. The repository does not claim that the pending experiments have run.

• Keep review submissions anonymous and leave \usepackage{icml2026} in review mode. Use the accepted option only for camera-ready papers.
• Keep the main body within eight pages. References, the impact statement, and appendices may follow without a page limit; camera-ready papers receive one additional main-body page.
• Put the mandatory Impact Statement before the references.
• Keep each abstract to one self-contained paragraph, ideally four to six sentences.
• Do not modify the ICML style files or compress spacing.
• Never replace -- or pending result cells without an immutable run artifact and aggregation script.
• Add only verified citations and measured results. Never retain invented references, numbers, baselines, or implementation details.

See the official ICML 2026 author instructions and call for papers before submission.