diff --git a/docs/adr/0123-a-linear-observable-parameter-is-profilable-only-in-the-space-its-own-noise-family-scores-so-the-log-family-half-is-refused-on-the-math-and-the-linear-gaussian-half-stays-open-on-a-measurement-the-corpus-cannot-make.md b/docs/adr/0123-a-linear-observable-parameter-is-profilable-only-in-the-space-its-own-noise-family-scores-so-the-log-family-half-is-refused-on-the-math-and-the-linear-gaussian-half-stays-open-on-a-measurement-the-corpus-cannot-make.md index 822aca2e..0d08e7cf 100644 --- a/docs/adr/0123-a-linear-observable-parameter-is-profilable-only-in-the-space-its-own-noise-family-scores-so-the-log-family-half-is-refused-on-the-math-and-the-linear-gaussian-half-stays-open-on-a-measurement-the-corpus-cannot-make.md +++ b/docs/adr/0123-a-linear-observable-parameter-is-profilable-only-in-the-space-its-own-noise-family-scores-so-the-log-family-half-is-refused-on-the-math-and-the-linear-gaussian-half-stays-open-on-a-measurement-the-corpus-cannot-make.md @@ -1,14 +1,14 @@ # A linear observable parameter is profilable only in the space its own noise family scores, so the log-family half is refused on the math and the linear-Gaussian half stays open on a measurement the corpus cannot make (issue #572) -**Status: Superseded in part by ADR-0129 (2026-08-25).** The two measurements this ADR named as -#572's gate have been taken, and they close it: see ADR-0129 for the decision, the numbers, and the +**Status: Superseded in part by ADR-0130 (2026-08-25).** The two measurements this ADR named as +#572's gate have been taken, and they close it: see ADR-0130 for the decision, the numbers, and the scope. **What still stands:** finding 1's log-versus-linear split, finding 2's double-binding -refusal and its two instances, and finding 5's census. **What ADR-0129 corrects:** finding 1's table +refusal and its two instances, and finding 5's census. **What ADR-0130 corrects:** finding 1's table sorts families by the space their residual lives in, which is the wrong axis -- `laplace` scores a linear residual but sums absolute values, so least squares is not its conditional optimum and it has to be refused for the coupled form; and finding 3's reading of `Laske`'s `Int_nuc_off` ("profiles to approximately 57.13 at essentially every box draw"), which is 9 of the 15 draws that integrate, with -five going to effectively zero -- the compression it reports does reproduce. **What ADR-0129 +five going to effectively zero -- the compression it reports does reproduce. **What ADR-0130 answers:** finding 4's redundancy question, and the "what would decide it" measurements at the end. *(Originally -- Narrowed, not decided, 2026-08-21.)* ADR-0066/0099 profile a declared column's @@ -255,9 +255,8 @@ If (1) shows the coupled pair behaves like `Int_nuc_off` did and (2) shows a ben gate. * Two reusable tools land in the corpus with their gotchas written down: `Grein-2026-benchmark-subset-I/tools/linear_scope.py` and `.../linear_profile.py`. - **They were committed and not merged**, to `wshlavacek/BNGL-Models` branch - `tools/572-linear-observable-profiling`, so this ADR has been pointing at files that are not on - `main`. ADR-0129's branch merges them. + They landed as that repository's PR #47, four days after this ADR was written; ADR-0130 adds a + third tool and two more confs beside them. ## Prior art diff --git a/docs/adr/0129-a-linear-coefficient-profile-is-gated-on-a-quadratic-loss-not-on-a-linear-residual-space-and-on-a-gaussian-family-it-sharpens-the-ordering-rather-than-collapsing-it-and-reaches-the-optimum-sooner.md b/docs/adr/0130-a-linear-coefficient-profile-is-gated-on-a-quadratic-loss-not-on-a-linear-residual-space-and-on-a-gaussian-family-it-sharpens-the-ordering-rather-than-collapsing-it-and-reaches-the-optimum-sooner.md similarity index 97% rename from docs/adr/0129-a-linear-coefficient-profile-is-gated-on-a-quadratic-loss-not-on-a-linear-residual-space-and-on-a-gaussian-family-it-sharpens-the-ordering-rather-than-collapsing-it-and-reaches-the-optimum-sooner.md rename to docs/adr/0130-a-linear-coefficient-profile-is-gated-on-a-quadratic-loss-not-on-a-linear-residual-space-and-on-a-gaussian-family-it-sharpens-the-ordering-rather-than-collapsing-it-and-reaches-the-optimum-sooner.md index 8a9b340e..7e8e69c5 100644 --- a/docs/adr/0129-a-linear-coefficient-profile-is-gated-on-a-quadratic-loss-not-on-a-linear-residual-space-and-on-a-gaussian-family-it-sharpens-the-ordering-rather-than-collapsing-it-and-reaches-the-optimum-sooner.md +++ b/docs/adr/0130-a-linear-coefficient-profile-is-gated-on-a-quadratic-loss-not-on-a-linear-residual-space-and-on-a-gaussian-family-it-sharpens-the-ordering-rather-than-collapsing-it-and-reaches-the-optimum-sooner.md @@ -30,12 +30,10 @@ The short version: ## Where the measurements and the tools live -In `wshlavacek/BNGL-Models`, branch `572-linear-profiling-race`, which also carries the two tools -ADR-0123 described. **ADR-0123 recorded those as having landed in the corpus and they had not** -- -they were committed to `tools/572-linear-observable-profiling` and never merged, so the ADR has been -pointing at files that are not on `main`. This is the branch that merges them, together with the -fixture, the additions below, and the failure modes each one hides, written up in -`Grein-2026-benchmark-subset-I/tools/README.md`. +In `wshlavacek/BNGL-Models`. The two tools ADR-0123 describes, and the fixture, landed there as +that repository's PR #47; the additions below are PR #48. Every failure mode these tools hide is +written up in `Grein-2026-benchmark-subset-I/tools/README.md`, which is the file to read before +believing any number taken with them. What was already there: `linear_scope.py` (which free parameters enter an observable linearly, and in what space), `linear_profile.py` (simulate a box point once, then minimize **PyBNF's own