This guide is based on committed source and experiment records. It does not rerun GPU benchmarks or claim a fresh November 14 training result.

Source of each claim

Preservation is not complete replay

The project keeps compact reports, selected reproduction scripts and the graph in Git. Bulk logs, generated binaries, large tensor captures and several obsolete checkpoints were removed to reclaim laptop space. Remaining artifacts and a small installed runtime were archived privately on HF. An evidence path in an older ledger entry may therefore describe an artifact that is no longer local or publicly downloadable. A summary with a hash is useful provenance, but it is not a replacement for the full original run directory. I do not claim that every one of the 240 experiments can be replayed from the public checkout alone. The kernel fork and submodule revision also matter. A plain clone with uninitialized submodules does not contain the full CUDA dependency source, and the pinned Torch/vLLM/CUDA stack is not installed by cloning the repository.

What is still open

  • Show sustained task learning on an environment with useful reward contrast. Wordle and Reverse Text exact success remain zero in these records.
  • Establish GSM8K baselines and a matched post-training evaluation before treating it as a learning success.
  • Resolve the work and trainer-trajectory divergence in the ten-update startup comparison.
  • Make a kernel change produce a repeatable whole-RL gain under comparable rollout work.
  • Measure capacity for longer contexts and additional model sizes. A 1B full-training or coding-agent result is not part of the evidence here.

Documentary provenance

The guide was assembled against source snapshot a0d23a0e58daeebde74dbed0484295116724351a on 14 November 2026. manifest.json records chapter order and evidence paths. evidence.json carries the selected numeric records with source-file hashes. The portfolio copy is generated from these same pages. Return to the overview.