experiments: predeclare Pareto and scaling gates

1 files changed, 22 insertions, 0 deletions

diff --git a/experiments/finalize_claims.sh b/experiments/finalize_claims.sh
index 70e53a2..104d87c 100755
--- a/experiments/finalize_claims.sh
+++ b/experiments/finalize_claims.sh
@@ -8,6 +8,7 @@ TRAIN_PY=/home/yurenh2/miniconda3/envs/ep_pascal/bin/python3
 PLOT_PY=/home/yurenh2/miniconda3/bin/python
 OUTDIR=results/figs
 SUMMARY=results/audited_tables.md
+MANIFEST="$OUTDIR/main_figure_manifest.json"
 
 mkdir -p "$OUTDIR"
 
@@ -16,6 +17,27 @@ mkdir -p "$OUTDIR"
 "$PLOT_PY" experiments/plot_main_figures.py \
   --results results --outdir "$OUTDIR"
 
+# These claim boundaries were frozen before the missing d5/d10/d20 cells were
+# generated.  They turn "Pareto endpoint" and "scales" into falsifiable gates
+# rather than captions selected after looking at the completed trajectories.
+# Depth grows 12x from d5 to d60; the 15x wall allowance is a 25% systems-noise
+# margin.  Accuracy retention permits at most one percentage point of loss.
+jq -e '
+  .strict == true and
+  (.missing_cells | length) == 0 and
+  .statistics.sdil_is_high_accuracy_frontier_endpoint == true and
+  .statistics.depth5_to_depth60.SDIL.d60_minus_d5_accuracy_points_mean >= -1.0 and
+  .statistics.depth5_to_depth60.SDIL.d60_over_d5_wall_geomean <= 15.0 and
+  .statistics.cifar_scaling.SDIL_d60.accuracy_percent_mean >
+    .statistics.cifar_scaling.FA_d60.accuracy_percent_mean and
+  .statistics.cifar_scaling.SDIL_d60.accuracy_percent_mean >
+    .statistics.cifar_scaling.DFA_d60.accuracy_percent_mean and
+  .statistics.cifar_scaling.SDIL_d60.early_alignment_mean >
+    .statistics.cifar_scaling.DFA_d60.early_alignment_mean and
+  .statistics.paired_gap_to_bp.BP_minus_SDIL_d60.accuracy_point_gap_mean <= 1.5
+' "$MANIFEST" >/dev/null
+echo "predeclared Pareto/scaling claim gates passed"
+
 "$TRAIN_PY" experiments/baseline_smoke.py
 "$TRAIN_PY" experiments/synthetic_smoke.py
 "$TRAIN_PY" experiments/smoke.py