[SUPR1M3]

Problem

With random_seed set (the default is 42), every LLM request carries the same seed. On providers that honour it (Ollama, vLLM, and OpenAI on a best-effort basis), identical prompts return identical responses. Identical prompts are common early in a run, when several iterations share the same parent, so those iterations produce duplicate children and waste their evaluations.

In a 10-iteration pilot on Ollama, up to four iterations of one run returned byte-identical programs. (How deterministic a backend is varies: under concurrent GPU load, Ollama can diverge even with the same seed. Where it is deterministic, a fixed seed causes duplicates; where it isn't, a per-iteration seed costs nothing.)

This is a follow-up to #554/#559, which fixed the same problem for model selection (reseed(iteration)). The request seed itself was still fixed.

Fix

After reseed(iteration), LLMEnsemble sends a seed derived from the sampled model's random_seed and the iteration (a SHA-256 hash, truncated to a non-negative 31-bit int).

• Reproducible: the same config and iteration always send the same seed, whichever worker runs it, so the reproducibility from Support seed for reproducibility #97 is kept.
• An explicit seed kwarg still wins.
• No seed configured, no seed sent: requests carry no seed, as before.
• Ensembles that are never reseeded are unchanged, such as the evaluator's LLM-feedback ensemble.

_sample_model() now delegates to a new _sample_index(), so generate and generate_with_context know which model's seed to use.

Tests

Adds tests to tests/test_ensemble_worker_seed.py, which runs fresh worker processes and checks that:

• identical prompts get a different seed on each iteration
• the request seed depends only on the iteration, not on which worker runs it
• unseeded configs send no seed, and their model selection stays random

python -m unittest discover tests: 1016 tests, OK.

Related

Found while running the full_rewrite_probability experiment for #116 (#580). Independent of that PR, the two merge cleanly in either order.