PhD-Brown · Nov 11, 2026 · Nov 11, 2026 · Nov 11, 2026 · Nov 13, 2026 · Nov 13, 2026
diff --git a/audit/M28_DIMRED_VISUALIZER.md b/audit/M28_DIMRED_VISUALIZER.md
diff --git a/audit/M28_PROOFS.md b/audit/M28_PROOFS.md
@@ -0,0 +1,84 @@
+# M2.8 — immutable red evidence manifest
+
+> Revision 2026-11-13: this manifest and its counts describe the **first pass**.
+> The original archive is preserved unchanged. See [revised proof manifest](M28_REVIEW_PROOFS.md)
+> for corrected oracles, new archive and review campaigns.
+
+Refs #43. Base39dad874610745ccd456d26972e90f6fb3a11405;
+production visualizer blob624b857bca5833c3401204b0a592b9ae6ffc2902 unchanged.
+`tests/proofs_dimred_visualizer.py` is explicitly executed, never discovered
+by the normal test naming pattern. It is committed in one final archive,
+pushed/read back, then removed normally from the Draft PR head. No xfail.
+The designated archive is **ed1d91ec0efb619aeb77a5bf7e4a0ab43fa84cac**; its
+ancestry preserves exact conftest, green tests, source, scripts and results.
+
+Final baseline: **22 cases,21FAIL/1PASS/0ERROR/SKIP**. This is **15 grouped
+findings**, not21bugs. The passing `n_show=0` case already refuses an impossible
+subplot count. It is a control/observation, not a requirement to keep one
+particular exception text. Green valid-data cases accompany invalid proofs.
+
+| ID | Red test selection / failed cases | Independent invariant and admitted outcomes |
+|---|---|---|
+| D01 | `kde_labeled_regions` /2 | Real SciPy and analytic Gaussian covariance diag(4,9). Independent expanded-grid quadrature and analytic `max(0,1-12*pi*t)`; density-height percentiles do not specify enclosed probability. Integrated thresholds accepted. M1 already noted this risk; no separate corrective issue existed. |
+| D02 | `hr_region_vertices` /1 | All gravity coordinates must lie in gravity-domain bounds. Current mixed K/dex margin yields−93.3175dex for supplied ordinary data. Nonempty informative regions required; corrected gravity margin admitted. |
+| D03 | `hr_population_title` /2 | Displayed count must be supplied population count6/9 or omitted. Hardcoded43019 is neither. Omission admitted; stellar-region taxonomy is not validated. |
+| D04 | `reconstruction_cardinality` /1 | Deleted middle observation: actual F mean1140/13, plotted1060/12. Incompatible lengths must not reassign labels by truncation. Pertinent ValueError/RuntimeError refusal or true correspondence admitted; valid aligned F10/G12 still must work. No external equal-size IDs inferred. |
+| D05 | `cluster_colorbar` /2 | Nonconsecutive2/7 labels and actual analyzer palette. Displayed colorbar mapping at each CID must match the point RGB; an informative categorical annotation can replace a colorbar. No fixed aesthetic palette imposed. |
+| D06 | `actual_unit_confidence` /1 | Values all1 are compatible with actual probabilities. Without an availability flag the renderer cannot assert predict_proba absent. Missing probability still tested as None in green; no calibrated probability claim. |
+| D07 | `no_physical_columns` /2 | Class-only AE/zoom inputs have a defined nonempty class plot or relevant explicit refusal; scalar axes TypeError is not a decision. Correct scalar-to-array handling admitted. |
+| D08 | `cluster_profiles_unavailable` /2,1PASS | n_show1 must address an actual axis; n_show4/two clusters must hide unused data slots. n_show0 explicit refusal admitted. Top-five signed numeric data remain protected in green. |
+| D09 | `more_reconstructions` /1 | Requested4/available2 must not present blank measurement panels. Hide unused or allocate actual count; valid1/3/6 rows still exact. |
+| D10 | `truncated_variance_annotation` /1 | Reported3PCs cannot be paired with85% when3 actually explain100%. Correct numeric association or removal/clarification of that claim accepted; no PCA sign convention. |
+| D11 | `log_histogram_keeps` /1 | Strictly positive1e-8..1 MSE values cannot disappear below an arbitrary1e-4 bin floor without indication. Positive-support bin choice admitted; truezero/negative policy remains open. |
+| D12 | `fifth_class` /1 | All five supplied classes represented or relevant explicit refusal; no silent loss due palette length. Nominal <=4class density arithmetic separately protected. |
+| D13 | `default_output_directory` /2 | Configured output_dir must actually create a default file on successful training/comparison plot; direct calls currently bypass_save. Explicit exports do work and are green. |
+| D14 | `trustworthiness_axis_contains` /1 | Calculated.5,.55882353,.50396825 cannot all be below a fixed.75 display floor without indication. Full0..1 range admitted; independent numerical computation stays green. |
+| D15 | `default_eigenspectrum` /1 | Actual default Balmer mathtext must draw with Agg. Doubled backslashes cause a parser error. Supplied custom marker labels are already green; unescaping default labels admitted. Notebook01 supplies RAIES, so no historical failure asserted. |
+
+## Reproduction and admissibility
+
+On the archived checkout:
+
+```powershell
+python -m pytest -q tests/proofs_dimred_visualizer.py --tb=short
+```
+
+From the green checkout with the fixed archive SHA:
+
+```powershell
+python audit/check_m28_evidence.py --mode baseline --proofs-ref ed1d91ec0efb619aeb77a5bf7e4a0ab43fa84cac
+python audit/check_m28_evidence.py --mode proofs --proofs-ref ed1d91ec0efb619aeb77a5bf7e4a0ab43fa84cac
+```
+
+The runner reads exact proof/green/fixture bytes from Git into its own synthetic
+TemporaryDirectory; production source is unchanged. No real datasets are copied.
+Exit1 for explicit proof pytest is expected. The runner itself reports evidence
+without treating expected counterexamples as an infrastructure failure.
+
+Mutations M01–M16 are separately scoped to green tests, all reached and KILLED
+by assertions with matching collection. They are not15production fixes.
+Fourteen admissible C01–C14 candidates exercise percentages, export, aesthetic
+colors, explicit alignment refusal, gravity margin, count omission, small MSE
+support, scalar axes, profile addressing, full trust range, hidden unused
+reconstructions, integrated KDE and valid mathtext. These are experimental
+copies, not patches proposed for production. Selections and test counts are in
+the ledgers; each candidate was not run against the entire repository.
+
+F01 always-refuse is rejected by the valid aligned mean oracle, F02 empty HR
+by explicit informative-region assertion. F03 standard-normal thresholds for
+every KDE initially survived an almost isotropic sample. Covariance diag(4,9)
+now discriminates it: F03 KILLED, while integrated C13 remains PASS. The final
+KDE ledger supersedes only this narrow earlier evidence. Old results remain
+unchanged. Analytic mass is clamped to0 above peak density; this avoids an
+incorrect negative mass in the audit oracle itself.
+
+## Limits
+
+Policy adapters are bounded: explicit diagnostic keywords, current French count
+and variance annotation grammar, Matplotlib colorbar/annotation interfaces and
+probability contour order. They do not parse arbitrary scientific language or
+all future plotting implementations. A future API/representation change must
+update adapters while preserving numerical contracts. In particular diagnostic
+wording is not universal proof of correct refusal, class-based availability is
+not object identity and a renderable figure is not astrophysical validation.
+No green test demands retaining the demonstrated red behavior.