Skip to content

Round 24: method registry, area inequality indices (SII, RII, concentration) and indirect SMR; v0.22.0 - #22

Merged
cafeTechne merged 15 commits into
defaultBranchfrom
phase24/inequality
Nov 16, 2026
Merged

cafeTechne merged 15 commits into
defaultBranchfrom
phase24/inequality

Conversation

@cafeTechne

Copy link
Copy Markdown
Owner

Centralize manifest method execution, validation, statistic extraction and methods text so new methods no longer require parallel dispatch chains. Add area-weighted SII/log RII/concentration and indirect SMR with strict ACS replicate uncertainty, explicit direction/reference-rate contracts, circularity enforcement, ratio reporting and recorded complete-case exclusions.

Both pre-specified validation scripts and criteria were committed and pushed before full runs:

  • Inequality criterion 60330e3: required conservative 90% coverage 0.876–0.905 (threshold 0.86). Rank CV 30% is informational and under-covers; the validated range remains <=15%.
  • SMR criterion 680707e: combined coverage 0.893–0.909 across all six cells. Fixed-E Garwood intervals are reported separately.

The exploratory MI/NY LBW example ran entirely from caches with HTTP blocked: 144 of 145 counties complete, median rank CV 6.63%; all conservative intervals include their null. The study report is committed; raw outputs remain in a scanned local archive.

Validation:

  • Part A: all 478 original golden snapshots byte-identical; full suite unchanged at 1634 passed / 37 skipped.
  • Final golden: 516 snapshots across 12 manifests; among the 478 legacy snapshots only 48 supported-method-list message fields change.
  • Final full suite: 1703 passed, 37 skipped, 49 subtests passed; 69 new Round 24 tests.
  • Existing MCP tool count remains 63. No existing test files modified.
  • All commits use [skip ci]; no merge, tag or release performed.

SMR exposes only the combined conservative target; reference rates remain fixed. Independent-area inference does not account for spatial dependence.
SMR currently refuses negative ACS band replicates, including when their total expected count is positive, to retain the pre-specified invalid-population domain. Both final-library simulation replays exactly reproduce every committed audit cell; extending that input domain requires separate validation.

Maintainer review (author self-review, not independent)

  • Golden: Part A (ca8604d) byte-identical to d1e4443 (478 snapshots). Final: of the 478 legacy snapshots, exactly 48 leaves differ, all the unsupported method 'no_such_method'; supported [...] message now listing inequality and smr.
  • Statistics checked by hand: weighted mid-rank ridits with tied blocks; SII = WLS slope; HC3 from whitened leverage w(1/W + dr²/Srr); log RII delta method; C = 2·b·Srr/(W·ȳ) = 2cov_w(y,r)/mean_w(y); Garwood limits; replicate variance 4/80.
  • Pre-registration: criterion commits 60330e3 and 680707e precede the result commits; scripts and criteria unchanged afterwards. Replayed both: inequality report and JSON byte-identical; SMR JSON identical (the .md adds a hand-written implementation-domain paragraph).
  • Manifest smr statistic uses the combined interval (the fixed-E Garwood interval under-covers when E is estimated: 0.335 at E=500, CV 20%). Rank CV > 15% gives a note, as the plan specifies.
  • Full suite on this branch: 1703 passed, 37 skipped.
  • Merged defaultBranch (Tests: fix 3 Windows-only WONDER fixture failures (CRLF changed the pinned SHA-256) #21, Windows fixture fix). Study report now points to the release attachment.

@cafeTechne
cafeTechne merged commit 44026b9 into defaultBranch Nov 16, 2026
17 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant