Auto-sync: 2026-06-30 10:19:50

results/README.md

@@ -154,11 +154,14 @@ scale-gated challenger variants at `0.35` and `0.35/0.40` also tie at 36.00%
 with slightly lower MSE, while `0.03` is slightly lower and adding wrong-gripper
 challengers raises MSE. The clean gain is a tight singleton near-miss calibration
 rather than broad residual mixing or monotone shortest-tangent filtering.
-The next clean support-gap test is pending: `near_miss_policy_bc5_typedprop_p2`
-adds a typed proposal head that generates expert plus primitive non-expert
-proposal families, then evaluates `proposal_lattice` with policy fallback at
-margins `0.00` and `0.05`. The job chain is train `14962264`, eval
-`14962356`/`14962357`, summary `14962363`/`14962364`.
+The typed proposal-head support-gap test completed for the best-rank checkpoint:
+six-family generated proposal lattices reach 31.30% at margin `0.00` and 32.17%
+at margin `0.05`. Anti-goal family masking improves the row but does not beat
+compatible tangent transport: safe4 reaches 33.45%, sparse no-op/wrong-gripper
+34.26%, and no-op-only 34.43%. The result is a useful negative diagnostic:
+direct typed proposal generation adds clean support, but the generated support
+is still not local enough. Best-policy checkpoint ablations are pending as eval
+arrays `14963143` and `14963147`, with summaries `14963150` and `14963149`.
 
 - Code supports `--retrieval-residual-scale`, defaulting to `1.0` so the
   completed 32.12% result is unchanged.
@@ -414,8 +417,13 @@ margins `0.00` and `0.05`. The job chain is train `14962264`, eval
 - Jobs `14936131`/`14936132`/`14936133` completed the compatible-chart
   abstention test at margin `0.10`: it reaches 34.67%, below the 36.06% top row.
   This rules out a naive lower-margin fix for the apparent oracle headroom.
-- Jobs `14962264` -> `14962356`/`14962357` -> `14962363`/`14962364` are pending
-  for the typed proposal lattice support-gap test.
+- Jobs `14962264` -> `14962356`/`14962357` -> `14962363`/`14962364` completed
+  the six-family typed proposal lattice support-gap test at 31.30%/32.17%.
+- Jobs `14963030` -> `14963044`, `14963031` -> `14963042`, and `14963032` ->
+  `14963043` completed typed proposal family masking at 33.45%, 34.26%, and
+  34.43%.
+- Jobs `14963143` -> `14963150` and `14963147` -> `14963149` are pending
+  best-policy checkpoint ablations for sparse/no-op typed proposals.
 
 Paper table automation: