Bounded-context reasoning in Python, built on the Delethink method from McGill-NLP's The Markovian Thinker and integrated into a vendored VERL tree. Generation runs in fixed-size chunks with a configured head and tail carried between them, so the active context stays bounded while the reasoning chain keeps growing.
The base is Delethink, published by McGill-NLP in The Markovian Thinker: the chunked loop, the trimmers, the trainer integration and the reproduction scripts are inherited and credited, and the package still installs as verl.
The mechanism is a budget, not a compression. Generation is cut into chunks of at most C tokens, and between one chunk and the next a fixed head and tail are carried and everything else is dropped. The shipped example sets C to 8,192 with 4,096 carried and the first 100 tokens of the response pinned as the head, so the model re-enters every chunk holding the question and its own opening move. Two configurations ship against a 1.5B distilled reasoning model, at 24K and 96K thinking budgets.
Divyam's own work sits in a separate package outside the inherited namespace, and it is three things. A salience carryover selector that consumes learned per-token utility scores under a hard token budget, a mandatory head and tail, and stable source order — so selection changes what is carried, never how much or in what sequence. A continuity verifier that catches variable and value contradictions across chunk summaries before the next rollout starts. And a carryover benchmark lab that runs competing policies under identical cases, budgets and utility functions and writes raw JSONL receipts, with an offline fixture so it runs without a cluster.
One more piece of engineering saves what bounded context usually loses. Every chunk in a trajectory carries a stable rollout request id, so the server pins the trajectory to one worker and the engine's prefix cache reuses the unchanged question and head instead of re-encoding them per chunk. Each output carries a receipt with its submitted and cache-eligible token counts.