[imerad]

Description

• Adds generate_batched_binary_dataset_with_oracle_sampling, which generates a stream of binary datasets with oracle uncertainty scores, one per batch, with per-batch n_samples, true_mean, proxy_mean and correlation (e.g. to inject a drift in actively sampled data). It is exported from glide.simulators and listed in the API reference.
• Closes Add generate_batched_binary_dataset_with_oracle_sampling simulator #508
• Any noteworthy implementation decisions or trade-offs?

Checklist

Quality gates that must be satisfied before requesting a review:

• I have read CONTRIBUTING.md
• make lint passes
• make type-check passes
• make tests passes
• make coverage reports 100% coverage
• make test-notebooks passes
• New public API has numpy-style docstrings
• New public API is inserted in the API reference section of the documentation
• Docs build without warnings (make doc)
• CHANGELOG.md updated if the change is user-facing

LLM usage

Disclose if an LLM was used in writing this PR:

• No LLM used
• I used an LLM and I went through and validated all the code myself

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[imerad]

Code review

Six findings, none blocking.

1. Duplicated batching scaffolding (glide/simulators/batched_oracle_binary.py:77)

The per-batch loop, SeedSequence spawning, per-batch call, np.full_like batch ids and np.hstack assembly are copied from batched_stratified_binary and batched_clustered_binary, with only the inner generator swapped. This is a third copy, so a future change to batch-id dtype or seed derivation has to be made in three places and can silently diverge. It goes against the "No Redundancy" rule. A shared private helper taking the inner generator would avoid it (could also be a follow-up refactor).

2. No test for an infeasible batch (tests/unit/simulators/test_batched_oracle_binary.py)

The docstring's Raises section says a ValueError is propagated when any batch is infeasible, but no unit test asserts it (e.g. a correlation outside the feasible range for batch 1 only). The sibling batched tests cover this path.

3. SeedSequence input has no direct test (glide/simulators/oracle_binary.py:14)

random_seed was widened to accept np.random.SeedSequence, but only the batched wrapper exercises it, indirectly. A regression would only surface through the batched test's hard-coded values, not in test_oracle_binary.py.

4. Doctest is weak (glide/simulators/batched_oracle_binary.py:67)

The example only checks lengths and the number of unique batches. It never inspects y_proxy, the uncertainty output or the batch ids, so a bug applying batch 0's parameters to every batch would still pass, and it does not show the batch-wise drift the function exists for.

5. Type annotation gap (glide/simulators/batched_oracle_binary.py:15) (minor)

The batched wrapper still annotates random_seed as Optional[int], while generate_binary_dataset_with_oracle_sampling now takes Optional[Union[int, np.random.SeedSequence]]. The batched stratified and clustered wrappers have the same gap, so this is a consistency issue, not a crash.

6. API index row placement (docs/api/index.md:15) (minor)

The new row sits between the batched clustered and batched stratified entries, not next to the non-batched oracle simulator (line 11), and is not in alphabetical or logical order.