Working notes and internal AI checks

These are summaries of internal AI-assisted checks carried out in the shared research workflow on 17 September 2026. They are not external human peer review. The checked artifacts do not retain the exact historical AI model version or agent execution identifier.

Methodological review

A separate AI reviewer examined the existing historical and global methods and their sources. It distinguished an accounting decomposition from causal identification, a survey filter’s conditional uncertainty from broader model uncertainty, and fixed-member compensation benchmarks from a uniquely identified worldwide risk factor.

The zero-innovation inflation fit remains part of the reference specification because that boundary was allowed by the likelihood under the declared noise assumption. More responsive and unsmoothed paths form the sensitivity exercise. The horizon mapping and index difference remain economic assumptions.

Numerical validation

The AI validation agent had access to the estimator, frozen CSVs, manifests and source artifacts. Its recorded validation report was created at 16:48 UTC on 17 September 2026. It checked 63 five-component cases, 114,675 daily rows, the component identities, source timing and quarter-window comparisons. The checkpoint records nine targeted tests and two byte-identical executions of 21 generated files.

Agreement in these checks establishes consistency with the recorded implementation. It does not independently recreate the model from this public specification or measure economic identification error. The frozen records distinguish directly checked outputs from test results reported by the main workflow.

Publication audit, 18 September 2026

The site review prompted a new descriptive comparison of HMT and MPS observations, explicit checks of long-history gaps, and an interpolation comparison drawn from existing sensitivity outputs. Charts were rerendered from saved values. No expectation parameters or component histories were refitted.

Internal-check provenance (JSON) · Publication audit manifest · Month-by-month gap audit.

Earlier research stages

The earlier US-proxy model, global-source feasibility checks and CPIH integration work preceded this checkpoint. Their useful numerical comparisons remain in Charts 11–12. Source and curve-construction details are consolidated in the Methods and Replication pages. Earlier fiscal, collateral and QE investigations do not supply causal coefficients for the five published components.

Original working documents are retained in the research archive. The public summaries remove agent-to-agent instructions and superseded plans while preserving the scope and limits of the checks.