Intermittent 'rules-engine formulas must be deterministic' errors from /api/series

Twice on 2026-07-03 (Europe/Brussels morning), in-app series calculations against the deployed API failed with policyengine-core's randomness guard, e.g.:

Calculation failed: The formula for 'slcsp_age_0' called numpy.random.seed(), but rules-engine formulas must be deterministic functions of their inputs. Remove the random call. If you need a stochastic input, compute it once when building the dataset (with a seeded generator) and store it as an input variable instead.

and the same wording naming age_head during a cliff-refinement request. Both appeared in the frontend as "Calculation failed: …" (the api/* error prefix), for an NC household (adult 30, children 4 and 8, ZIP 27601, all programs) swept 0–100k.

The guard landed in policyengine-core on 2026-06-04 (randomness_guard.py, "Forbid randomness inside variable formulas"); this repo floats policyengine-us>=1.715.2, so deployments resolve whatever us/core pair exists at build time.

What I could not do is reproduce it on demand: direct POSTs to /api/series on the current production deployment succeed for the same household — full 0–100k sweep at step 1000, refinement-shaped requests with min_earned_income, and ACA-region ranges where slcsp_age_0 computes (aca_ptc > 0). That leaves an intermittent trigger — plausibly a conditional randomness path (e.g., a take-up or imputation branch) that only some request shapes or instance states reach.

Suggestions:

• Pin policyengine-us (and transitively policyengine-core) to an exact known-good pair in requirements.txt/pyproject.toml so deploys stop drifting, and bump deliberately.
• Grep the pinned policyengine-us for np.random reachable from formulas (slcsp_age_0, age_head dependency chains) and either fix upstream or pin below the offending release.
• If it recurs, capture the full traceback server-side — the guard's error names the innermost formula but not the caller that seeded.

Start at the /api/series entry point and inspect the resolved policyengine-us/policyengine-core versions in requirements.txt or pyproject.toml. Read randomness_guard.py, then grep the pinned policyengine-us dependency chains for np.random near slcsp_age_0 and age_head; if the issue recurs, capture the server-side traceback. Done means the offending dependency pair or upstream formula is identified and the deployment no longer reaches nondeterministic formulas.