The Broadcast Ceiling v0.1.0

The Broadcast Ceiling v0.1.0

First public research release for the Limes Labs mechanism/theory paper:

The Broadcast Ceiling: Information and Reliability Limits of Advantage Estimators in Long-Horizon RL

Release commit: 1d2956502cba967939551856e60d2e68d707d610

What This Is

A focused mechanism/theory artifact on long-horizon post-training credit assignment. The release studies when trajectory-level reward coefficients are enough, and when token/state-varying value-style credit is needed.

Main Result Package

• Broadcast-ceiling theory and information/reliability framing.
• CPU-only finite-MDP audits for trajectory-broadcast, structural, sampled-value, learned-value, oracle-value, and policy-implied estimators.
• Exact-gradient audit with 200 replications and true null-action diagnostics.
• Tabular closed-loop training and short-budget autoregressive MLP sequence-policy training.

Attached Assets

• the-broadcast-ceiling-v0.1.0.pdf - canonical rendered 32-page author PDF.
• paper_manifest.json - SHA-256 manifest and page-count metadata for the generated PDF. The manifest records the repository's internal build path for the PDF.
• sequence_policy_training_seedset.json - canonical 10-paired-seed sequence-policy result artifact.
• the-broadcast-ceiling-v0.1.0-latex-source.zip - compact LaTeX/source bundle for rebuilding the paper PDF. Includes manuscript source, bibliography, generated LaTeX tables, public figures, result JSON inputs, and paper build scripts.
• the-broadcast-ceiling-v0.1.0-full-source.zip - complete repository source archive at the release tag, including experiments, tests, docs, results, paper sources, and public artifacts. GitHub also provides automatic source archives; this named archive is included for stable citation and checksum verification.
• the-broadcast-ceiling-v0.1.0-artifact-readme.md - concise guide to the release assets and reproducibility commands.
• the-broadcast-ceiling-v0.1.0-submission-metadata.md - copy-ready title, abstract, keywords, URLs, and upload guidance for Zenodo, OSF, Figshare, and OpenReview.
• SHA256SUMS-v0.1.0.txt - checksums for manually uploaded release assets.

Sequence-Policy Audit Note

Table 15 reports 10 paired seeds: neural value TD reaches final return 0.839 vs group broadcast 0.811, with paired Delta R = 0.028 and 95% paired CI [0.007, 0.050]. Environment-interaction budgets are matched; total compute is not matched because the TD arm additionally trains a critic.

Rebuild and Verify

From the full source archive or repository checkout:

./scripts/build_latex_paper.sh
python3 -m unittest discover -s tests
./scripts/run_smoke.sh

Release verification before publication:

• python3 -m unittest discover -s tests passed: 51 tests.
• ./scripts/run_smoke.sh passed.
• ./scripts/build_latex_paper.sh passed: 32-page PDF.
• git diff --check clean.

Non-Claims

This is not a transformer-scale PPO-vs-GRPO benchmark and not an LLM leaderboard result. It is a focused, reproducible mechanism/theory release with synthetic audits and explicit external-validity limitations.