auto-sync 2026-07-02T22:12:27Z workspace

workspace/docs/cil_format.md

@@ -96,6 +96,15 @@ Learned generator targets:
   reach 23.66% PTR at RMS<=0.20 and 52.69% at RMS<=0.40, while pool64
   negative-margin reranking keeps or lowers strict PTR and does not reduce the
   5.33% strict negative-near rate.
+- `scripts/eval_positive_tangent_chart_synthesis.py` is the next local-chart
+  diagnostic. It keeps train-only positive chart atoms and adds barycentric
+  means over nearby chart neighborhoods, testing whether positive support is
+  better expressed as local chart coordinates than as raw prototype replay.
+  Current best keeps 15 direct local atoms plus one 16-neighbor chart mean,
+  preserving K16 PTR at 23.66% / 52.69% and strict negative-near at 5.33% while
+  improving positives-closer-than-negatives to 65.33%. This remains an offline
+  support proxy; successful settings should motivate the learned object-centric
+  atlas generator rather than become the final deployment method by themselves.
 - `scripts/train_positive_tangent_cvae.py` trains a first raw-action CVAE over
   train-only positive tangents. The companion
   `scripts/summarize_positive_tangent_cvae_sweep.py` ranks temperature/beta

workspace/latex/main.aux

@@ -34,5 +34,5 @@
 \@writefile{lot}{\contentsline {table}{\numberline {4}{\ignorespaces Selector calibration ablations on the current K6 transported residual field. All rows are deployment-clean and use only train-split calibration signals.}}{5}{table.4}\protected@file@percent }
 \newlabel{tab:selector-calibration}{{4}{5}{Selector calibration ablations on the current K6 transported residual field. All rows are deployment-clean and use only train-split calibration signals}{table.4}{}}
 \@writefile{toc}{\contentsline {section}{\numberline {6}Roadmap to the Full Paper}{5}{section.6}\protected@file@percent }
-\@writefile{toc}{\contentsline {section}{\numberline {7}Limitations}{5}{section.7}\protected@file@percent }
-\gdef \@abspage@last{5}
+\@writefile{toc}{\contentsline {section}{\numberline {7}Limitations}{6}{section.7}\protected@file@percent }
+\gdef \@abspage@last{6}

workspace/latex/main.log

@@ -1,4 +1,4 @@
-This is pdfTeX, Version 3.141592653-2.6-1.40.22 (TeX Live 2021 Gentoo Linux) (preloaded format=pdflatex 2023.8.23)  2 JUL 2026 17:00
+This is pdfTeX, Version 3.141592653-2.6-1.40.22 (TeX Live 2021 Gentoo Linux) (preloaded format=pdflatex 2023.8.23)  2 JUL 2026 18:44
 entering extended mode
  restricted \write18 enabled.
  %&-line parsing enabled.
@@ -484,13 +484,13 @@ File: umsb.fd 2013/01/14 v3.01 AMS symbols B
 {/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/var/lib/texmf/fonts/map/pdf
 tex/updmap/pdftex.map}] [2] (./tables/main_results.tex) (./tables/car_decomposi
 tion.tex) (./tables/source_score_sweep.tex)
-(./tables/selector_calibration.tex) [3] [4] [5] (./main.aux)
+(./tables/selector_calibration.tex) [3] [4] [5] [6] (./main.aux)
 Package rerunfilecheck Info: File `main.out' has not changed.
 (rerunfilecheck)             Checksum: B3D461B4FEC67164F7152E09ACAEE494;1245.
  ) 
 Here is how much of TeX's memory you used:
- 9899 strings out of 480884
- 146741 string characters out of 5900692
+ 9900 strings out of 480884
+ 146747 string characters out of 5900692
  460545 words of memory out of 5000000
  26980 multiletter control sequences out of 15000+600000
  414083 words of font info for 68 fonts, out of 8000000 for 9000
@@ -524,10 +524,10 @@ mf-dist/fonts/type1/public/amsfonts/cm/cmsy7.pfb></cvmfs/soft.computecanada.ca/
 gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/symbols/
 msbm10.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-
 dist/fonts/type1/public/cm-super/sfrm1000.pfb>
-Output written on main.pdf (5 pages, 213796 bytes).
+Output written on main.pdf (6 pages, 214742 bytes).
 PDF statistics:
- 154 PDF objects out of 1000 (max. 8388607)
- 127 compressed objects within 2 object streams
- 21 named destinations out of 1000 (max. 500000)
+ 158 PDF objects out of 1000 (max. 8388607)
+ 130 compressed objects within 2 object streams
+ 22 named destinations out of 1000 (max. 500000)
  81 words of extra memory for PDF output out of 10000 (max. 10000000)
 

workspace/latex/main.pdf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:247e89589cfd3fb8279d9c12bd5ea27187f8ba4156af2cc9a86082894801fb1f
+oid sha256:6166521afad4567b2534b40963afe78a2e6822be18cba258bdccfb34d72768fb
 size 213796

workspace/latex/main.tex

@@ -266,6 +266,19 @@ away from measured negatives in raw action space. The next generator should
 learn local positive chart coordinates and use negative contrast as a boundary
 inside that chart, not as a stand-alone selection prior.
 
+The next diagnostic makes this distinction explicit. Chart synthesis keeps a
+small number of local positive atoms, then adds barycentric means over nearby
+chart neighborhoods. The best setting keeps 15 nearest atoms and adds one
+16-neighbor barycentric chart mean. It preserves local-atlas K=16 PTR at
+23.66\% / 52.69\% and the 5.33\% strict negative-near rate, while improving
+positives-closer-than-negatives from 64.00\% to 65.33\% and mean positive
+distance from 0.5330 to 0.5247. Wider barycentric smoothing lowers distance
+slightly further but raises strict negative-near to 6.67\% or lowers loose PTR
+to 50.54\%. Thus chart coordinates help more than raw likelihood models, but
+support is not merely convex averaging in raw action space; the generator must
+learn object-centric chart transport rather than smooth residual chunks
+directly.
+
 A first raw-action CVAE improves the safety side of the proxy---0.00\%
 negative-near at the 0.20 threshold and 65.33\% positives-closer-than-negatives
 at K=16---but covers less positive support (7.53\% PTR at 0.20, 34.41\% at