[jadenfix]

Summary

• compute posterior updates in log-odds space so extreme finite Bayes factors do not overflow through inf / inf
• treat positive infinity as the limiting posterior 1.0 and reject NaN Bayes factors
• add focused regressions for the extreme finite, infinite, NaN, and robust-interval boundaries
• include a reproducible, source-pinned GIF/MP4 comparison with a machine-readable receipt

Reproduced boundary

On base 4e88e771e33bddca3ce3e718da7ed4b531c987d2:

posterior_from_bayes_factor(0.9, 1e308) -> nan
posterior_from_bayes_factor(0.5, inf) -> nan
robust_posterior_interval((0.1, 0.9), (0.5, inf)) -> (..., nan)

At fix commit ec2ad8caeeabbba28252b799cf5ca8101881b641, the same calls return 1.0, 1.0, and (..., 1.0); a NaN Bayes factor raises ValueError.

Verification

• PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src python -m pytest -p no:cacheprovider: 426 passed in 1337.38s
• focused inference suite: 7 passed
• python -m compileall -q src scripts: 109 files compiled
• python -m simtheory.experiments: exit 0, 151 output lines
• additional monotonicity/range grid: 88 boundary combinations
• MP4: H.264, 960×540, yuv420p, 30 fps, 13.833333 s

Source-pinned media

The receipt compares base 4e88e771e33bddca3ce3e718da7ed4b531c987d2 with fix ec2ad8caeeabbba28252b799cf5ca8101881b641 by loading inference.py directly from both Git revisions. It also records the focused test command and verified paths.

• docs/media/posterior-boundary-demo.gif — SHA-256 4caa0bacdec47f69132ff9c36c1325cceebc47ce07f33e2007f031f58115bfc9
• docs/media/posterior-boundary-demo.mp4 — SHA-256 1fa27eaf90e27551f220519aa56f40be249b9caff99eb13a435d3a40905f4224
• docs/media/posterior-boundary-demo-receipt.json — SHA-256 5349d9d86e9b6b71e263fb6904e469cc046bebc5fe403b7fe6b34eddda98ee51

Scientific boundary

This is a numerical-stability and input-validation fix. It does not validate any supplied likelihood model, change identifiability assumptions, or provide evidence for any simulation hypothesis.

[jadenfix]

Independent exact-head review — CLEAR

Reviewed exact head bef838efa89e346d940ed87e7f0c3385e1bdebde.

• Stable log-odds implementation avoids finite-BF overflow.
• Boundary contract verified: BF 0 → posterior 0; positive infinity → 1; NaN/negative/non-finite-negative inputs reject.
• Independent 117-case high-precision comparison plus monotonicity/range probes passed.
• Focused inference suite passed 7/7; full suite passed 426 tests on Python 3.11, 3.12, and 3.13.
• PR-triggered rerun is fully green: https://github.com/jadenfix/simulation-theory/actions/runs/38411588888
• Current PR state recalculated as clean and mergeable.
• All six changed files have zero path overlap with the repository's 20 pre-existing draft PRs.
• Fresh renderer execution reproduced GIF, MP4, and receipt byte-for-byte.
• Full GIF and MP4 decoding plus representative-frame inspection passed.
• Working tree remained clean.

This is narrowly a numerical-stability/input-validation repair. It does not validate likelihood models or provide evidence for any simulation hypothesis.

No blocker found. Ready for designated sole-merger review; not merged here.