[hmgaudecker]

Scaled continuation expectations can overflow before normalization when values are near the floating-point limit. Small coefficients and small values can also underflow before their finite normalized mean is formed. This change normalizes coefficient mantissas and tracks binary scales so these intermediate operations preserve the represented contributions.

The derivative rule forms normalized finite slopes and transports the centered residual separately. This preserves finite coefficient derivatives when the rounded mean equals a neighboring value, while retaining zero-weight and nonfinite-input conventions. The implementation uses existing JAX binary scaling; it adds no runtime accuracy certificate or exact-arithmetic fallback.

Validation covers float32 and float64 eager, JIT, mapped, forward- and reverse-mode paths with exact-rational references, public continuation/policy examples, and existing CE controls. Tests also check equivalent coefficient/shift representations and the physical scale returned by the shared helper.

The added arithmetic has a measured cost. Relative to the prior revision, the bounded CPU kernel is about 1.5–1.8 times slower for primal evaluation and 1.17–1.20 times slower for coefficient gradients. A matched bounded Eichenbaum A100 solve showed no measurable slowdown: warm solves were 6.16 seconds, within the repeated baseline's 0.44% spread, with unchanged device peak memory and bitwise-identical values and sampled policies. This covers the tested profile; whole-model differentiation and larger calibration grids remain unmeasured. CPU kernel and differentiation cost acceptance remains open.

Process-edge continuation now uses one shared corner program. The loop keeps the certainty equivalent inside each joint corner and retains the original corner order. This removes exponential copies of the continuation graph without changing the economic rule. A regression test limits program construction to two continuation programs.

The additional checks cover 202 native mathematical cases and 1,106 source mutations. On the benchmark runner's V100, the selected ACA simulation program built in 2.13 seconds; both old-code arms exceeded 82 seconds in that phase before the 120-second process cap. The repaired replay matched the captured outputs and layouts exactly, rejected output mutations, and ran three warm calls in 2.412–2.414 seconds. Baseline warm time and a full-model speedup remain unmeasured.

The normal ACA benchmark class completed on the V100 with a fresh JAX cache and configured FP64: cold automatic solve and simulation took 1,803.34 seconds, the following warm call took 250.28 seconds, and the measured CPU-memory peak was 11.14 GiB. Total benchmark-cache setup took 36 minutes 44 seconds. These are one cold and one warm observation on eight CPU cores; a source-only certificate used separate cores during part of the run. There is no controlled full-baseline or A/A comparison. The earlier graph-construction hang is repaired; the remaining full-model cost needs review. This direct invocation used the original benchmark class and measurement subprocess; it does not establish a successful GitHub ASV CLI run.

[read-the-docs-community]

Documentation build overview

📚 pylcm | 🛠️ Build #35652912 | 📁 Comparing 10ecfe0 against latest (f57269f)

  🔍 Preview build  

38 files changed · + 1 added · ± 37 modified

+ Added

• shocks/index.html

± Modified

• age-specialized/index.html
• architecture/index.html
• architecture-transition-ledger/index.html
• benchmarking/index.html
• beta-delta/index.html
• categories-with-characteristics/index.html
• certification/index.html
• choosing-a-solver/index.html
• collective-regimes/index.html
• collective-regimes-1/index.html
• and 27 more...

[codecov]

Codecov Report

❌ Patch coverage is 99.77401% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 94.10%. Comparing base (7b7d298) to head (10ecfe0).

Files with missing lines	Patch %	Lines
tests/test_process_simulation_laws.py	99.44%	2 Missing ⚠️

Additional details and impacted files

@@                         Coverage Diff                         @@
##           codex/on-nodes-transition-rules     #628      +/-   ##
===================================================================
+ Coverage                            94.06%   94.10%   +0.03%     
===================================================================
  Files                                  660      660              
  Lines                               109752   110569     +817     
===================================================================
+ Hits                                103240   104052     +812     
- Misses                                6512     6517       +5     

Flag	Coverage Δ
cpu-python	94.10% <99.77%> (+0.03%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.