Research reliability and the next PureMacro milestones
The agreed order is research reliability and replications, then economic models and policy applications, then teaching and usability.
Published corrections and current release
Version 6.4.0 (published to PyPI on 2026-11-12) succeeds published versions 6.3.0 (2026-11-10), 6.2.0 and 6.1.0 (published to PyPI on 2026-11-03 after 6.0.1 on 2026-10-29), incorporating the corrections below and those recorded in the changelog. The 5.8.0, 5.9.0 and 6.0.0 source tags were not published. See the current roadmap and API policy.
Native model adapters now preserve the supplied equations and loadings.
First-order DSGE scores use Kalman sensitivities; DAG scores centrally difference
the native Whittle likelihood while holding the supplied steady state fixed.
Unsupported native operations raise errors. Explicit model=None demonstrations
warn about their scope. See adapter details.
The default nonlinear HANK DAG uses prescribed household responses. Its
aggregate-risk example uses prescribed moments and matrices. They are explicitly
demonstrations; a supplied model cannot be silently ignored. The legacy
asymmetry_ratio reports nonlinear/linear peak amplification, not two solved
opposite-shock paths. See HANK scope.
Sovereign corrections include deterministic default, stationary-mean default
costs, zero-debt reentry and unadjusted simulation statistics. A constant spread
has an undefined correlation, reported as NaN with a status field.
Tests to run first
From the repository root, in a Python 3.11+ environment with .[dev,notebooks]:
python -m pytest tests/test_protocol_model_fidelity.py tests/test_research_reliability_references.py tests/test_sovereign/test_reference_reliability.py -q
python tools/release_check.py --no-tests
For an interactive tour, run notebook 72 in English or Spanish. Notebook 73 explores sovereign debt; 74 covers synthetic distributional LPs, Gaussian inference and reduced transitions. Their numerical scorecards have the specific scope described in each notebook, not blanket economic validation.
Independent evidence so far
| Comparison | Result on the tested grids | What remains open |
|---|---|---|
| Arellano short-term deterministic equations | Matched-grid equations pass; archived Fortran and MATLAB moments reproduce; public simulation timing and zero-debt reentry are corrected; public-calibration grids converge at 6,401 × 641, and the public solver reaches that grid with a certified monotone search | Published-moment discrepancies and extension-specific economic convergence |
| One-asset household block and published HANK equilibrium | Household policies and consumption match the SSJ authors' code within 5e-11; return Jacobians within 4e-4. The complete published equilibrium, both nonlinear shock signs and an extended asset/income/horizon ladder are independently checked; see the HANK evidence | Household incidence with observed data and general higher-order aggregate risk |
The original equation/block references are frozen with hashes and provenance in
tests/fixtures/research_reliability/; the linked study guides locate the
additional equilibrium and grid records. Regeneration tools live in
tools/reference_validation/; the external SSJ package is not a runtime
dependency. Matched-grid agreement does not prove discretization accuracy.
Sovereign study: baseline completed, public grids converged
The original-code and published-moment comparison now includes 16 independent simulations per case, nested debt grids and 21 to 321 income states. At the archived calibration, refinement moves the moments far more than simulation uncertainty, and spread volatility halves. For the public calibration, a pre-registered paired-seed criterion establishes grid convergence at 6,401 debt nodes and 641 income states; the same criterion establishes the archived calibration's convergence at 6,369 × 641, where spread volatility is about 2.9 percentage points against the published 6.36. A small Bellman residual alone would not have justified that conclusion.
- Public reentry timing, zero-debt initialization and measurement are now documented and tested. Keep the archive and public-calibration results distinct; they use different economic conventions.
- Done for the public calibration: grids converge at 6,401 × 641 under the
registered criterion, with sampling uncertainty kept separate from
discretization error and every raw moment retained.
solve_sovereign_default(policy_search="monotone")solves that grid, certified by an exhaustive check, and matches the reference arrays. - Study how income volatility, persistence and market-reentry probability change sustainable debt and default exposure in the validated short-term model. Report welfare and numerical uncertainty before adding maturity or restructuring mechanisms that require additional validation.
This creates a research result that can be reviewed from equations through policy tables. It also supplies the model and output contracts for future work.
The sovereign extension references now check the stated long-term smoothed operator, haircut/maturity states and collateral-planner policies independently. Long-term simulation draws from the same debt-choice distribution used in pricing. Collateral diagnostics are measured from simulated paths, with explicit undefined amplification and no prescribed fallback. The survival-spread residual is a compounding adjustment. These are small-grid equation checks, not published extension-moment replications.
Then: household incidence and teaching
The published one-asset HANK equilibrium now has matched fiscal rules, return timing and market clearing, independent nonlinear paths for both shock signs, and a passing extended grid/horizon study. Next, link household incidence to distributional LPs on observed data. A general higher-order aggregate-risk solver remains a separate later task.
Use the validated study to build a short bilingual path: install a verified wheel, reproduce one baseline, change one policy assumption, and inspect a diagnostic table. Keep demonstrations visibly separate from empirical results.