Auto-sync: 2026-06-30 03:27:59 (part 3)

scripts/build_paper_analysis.py

@@ -393,6 +393,14 @@ METHODS = [
             "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_oraclek8trace_summary.json"
         ),
     ),
+    MethodSpec(
+        key="residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002",
+        label="K4 compatible tangents, no-op bonus 0.03, near-miss challenger gate 0.02",
+        summary_path=(
+            "h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_"
+            "k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p02_summary.json"
+        ),
+    ),
     MethodSpec(
         key="residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus001",
         label="K4 composed compatible tangents, no-op bonus 0.03, singleton near-miss bonus 0.01",

scripts/build_paper_table_status.py

@@ -745,6 +745,16 @@ SPECS = [
         story_role="diagnostic branch trace for deployable selector calibration over the local tangent chart",
         pending_job="14953960/14953961",
     ),
+    ResultSpec(
+        key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmchallenger002",
+        label="K4 compatible residual retrieval, near-miss challenger gate 0.02",
+        path="h16_policy_ckpt_near_miss_policy_bc5_bestpt_retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_safe_margin0p20_noopbonus0p03_nmchallenger0p02_summary.json",
+        clean_deployment="yes",
+        same_state_proposals="no",
+        expert_proposal="no",
+        story_role="trace-motivated two-stage selector calibration over the compatible local tangent chart",
+        pending_job="14954280/14954281",
+    ),
     ResultSpec(
         key="retrieval_residual_k4_composemasked_dropnmnoop_grid035040045_noopbonus003_nmbonus001",
         label="K4 composed compatible residual retrieval, no-op bonus 0.03, singleton near-miss bonus 0.01",

scripts/eval_maniskill_policy_rollout.py

@@ -234,6 +234,19 @@ def main(argv: list[str] | None = None) -> int:
         "'compose_mean_by_type' also adds pairwise sums of type-consensus tangents; "
         "'field_softmax' forms a field-weighted tangent barycenter before final scoring.",
     )
+    parser.add_argument(
+        "--retrieval-residual-challenger-types",
+        default="",
+        help="Optional comma-separated residual candidate types that may override the "
+        "primary retrieval-residual selection in a second-stage challenger gate.",
+    )
+    parser.add_argument(
+        "--retrieval-residual-challenger-margin",
+        type=float,
+        default=0.0,
+        help="Required field-potential margin for a challenger candidate to override the "
+        "primary retrieval-residual selection.",
+    )
     parser.add_argument(
         "--lattice-exclude-types",
         default="",
@@ -272,6 +285,11 @@ def main(argv: list[str] | None = None) -> int:
     lattice_exclude_types = tuple(
         item.strip() for item in args.lattice_exclude_types.split(",") if item.strip()
     )
+    retrieval_residual_challenger_types = tuple(
+        item.strip()
+        for item in args.retrieval_residual_challenger_types.split(",")
+        if item.strip()
+    )
     candidate_type_bonuses: dict[str, float] = {}
     try:
         for item in args.candidate_type_bonuses.split(","):
@@ -339,6 +357,8 @@ def main(argv: list[str] | None = None) -> int:
         retrieval_residual_anchor=args.retrieval_residual_anchor,
         retrieval_residual_direction=args.retrieval_residual_direction,
         retrieval_residual_reduce=args.retrieval_residual_reduce,
+        retrieval_residual_challenger_types=retrieval_residual_challenger_types,
+        retrieval_residual_challenger_margin=args.retrieval_residual_challenger_margin,
         lattice_exclude_types=lattice_exclude_types,
         candidate_type_bonuses=candidate_type_bonuses,
         candidate_type_bonus_components=args.candidate_type_bonus_components,

scripts/slurm/eval_maniskill_policy_rollout.sbatch

@@ -70,6 +70,11 @@ fi
 RETRIEVAL_RESIDUAL_ANCHOR="${RETRIEVAL_RESIDUAL_ANCHOR:-expert}"
 RETRIEVAL_RESIDUAL_DIRECTION="${RETRIEVAL_RESIDUAL_DIRECTION:-candidate_minus_anchor}"
 RETRIEVAL_RESIDUAL_REDUCE="${RETRIEVAL_RESIDUAL_REDUCE:-none}"
+RETRIEVAL_RESIDUAL_CHALLENGER_TYPES="${RETRIEVAL_RESIDUAL_CHALLENGER_TYPES:-}"
+if [[ -n "${RETRIEVAL_RESIDUAL_CHALLENGER_TYPES_COLON:-}" ]]; then
+  RETRIEVAL_RESIDUAL_CHALLENGER_TYPES="${RETRIEVAL_RESIDUAL_CHALLENGER_TYPES_COLON//:/,}"
+fi
+RETRIEVAL_RESIDUAL_CHALLENGER_MARGIN="${RETRIEVAL_RESIDUAL_CHALLENGER_MARGIN:-0.0}"
 LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES:-}"
 if [[ -n "${LATTICE_EXCLUDE_TYPES_COLON:-}" ]]; then
   LATTICE_EXCLUDE_TYPES="${LATTICE_EXCLUDE_TYPES_COLON//:/,}"
@@ -148,6 +153,8 @@ apptainer exec --nv \
     --retrieval-residual-anchor "$RETRIEVAL_RESIDUAL_ANCHOR" \
     --retrieval-residual-direction "$RETRIEVAL_RESIDUAL_DIRECTION" \
     --retrieval-residual-reduce "$RETRIEVAL_RESIDUAL_REDUCE" \
+    --retrieval-residual-challenger-types "$RETRIEVAL_RESIDUAL_CHALLENGER_TYPES" \
+    --retrieval-residual-challenger-margin "$RETRIEVAL_RESIDUAL_CHALLENGER_MARGIN" \
     --lattice-exclude-types "$LATTICE_EXCLUDE_TYPES" \
     --candidate-type-bonuses "$CANDIDATE_TYPE_BONUSES" \
     --candidate-oracle-rollouts "$CANDIDATE_ORACLE_ROLLOUTS" \

scripts/slurm/summarize_h16_policy_ckpt.sbatch

@@ -101,6 +101,12 @@ for result_path in sorted(base_dir.glob(f"seed_*/{out_name}")):
                 "retrieval_residual_direction", "candidate_minus_anchor"
             ),
             "retrieval_residual_reduce": data.get("retrieval_residual_reduce", "none"),
+            "retrieval_residual_challenger_types": data.get(
+                "retrieval_residual_challenger_types", []
+            ),
+            "retrieval_residual_challenger_margin": data.get(
+                "retrieval_residual_challenger_margin", 0.0
+            ),
             "lattice_exclude_types": data.get("lattice_exclude_types", []),
             "candidate_type_bonuses": data.get("candidate_type_bonuses", {}),
             "candidate_type_bonus_components": data.get(
@@ -382,7 +388,7 @@ for row in rows:
     scales = row.get("retrieval_residual_scales") or []
     scale_grid = ",".join(f"{float(scale):.2f}" for scale in scales) if scales else "none"
     lines.append(
-        "| {seed} | {mode} | {k} | {policy_cand} | {retrieval} | {metric} | {anchor} | {direction} | {reduce} | {min_success:.2f} | {type_success_bonus:.3f} | {consensus_penalty:.3f} | {min_source_progress:.2f} | {source_progress_bonus:.3f} | {source_score_bonus:.3f} | {scale:.2f} | {scale_grid} | {margin:.3f} | {sigma:.2f} | {steps} | {trust:.2f} | "
+        "| {seed} | {mode} | {k} | {policy_cand} | {retrieval} | {metric} | {anchor} | {direction} | {reduce}{challenger} | {min_success:.2f} | {type_success_bonus:.3f} | {consensus_penalty:.3f} | {min_source_progress:.2f} | {source_progress_bonus:.3f} | {source_score_bonus:.3f} | {scale:.2f} | {scale_grid} | {margin:.3f} | {sigma:.2f} | {steps} | {trust:.2f} | "
         "{success:.2%} | {progress:.2%} | {oracle:.2%} | {candidate_oracle} | {oracle_gain} | {mse:.3f} |".format(
             seed=row["seed"],
             mode=row.get("selection_mode") or "policy",
@@ -394,6 +400,13 @@ for row in rows:
             direction=row.get("retrieval_residual_direction")
             or "candidate_minus_anchor",
             reduce=row.get("retrieval_residual_reduce") or "none",
+            challenger=(
+                " + challenger "
+                + ",".join(row.get("retrieval_residual_challenger_types") or [])
+                + f"@{row.get('retrieval_residual_challenger_margin') or 0.0:.2f}"
+                if row.get("retrieval_residual_challenger_types")
+                else ""
+            ),
             min_success=row.get("retrieval_type_min_success") or 0.0,
             type_success_bonus=row.get("retrieval_type_success_bonus_scale") or 0.0,
             consensus_penalty=row.get("retrieval_residual_consensus_penalty_scale")