auto-sync 2026-07-02T23:07:00Z workspace

workspace/cil/__init__.py

@@ -7,26 +7,38 @@ stable place to import benchmark metrics from.
 
 from cil.metrics import (
     bootstrap_ci,
+    branch_car,
     branch_causal_action_regret,
     car_decomposition,
     macro_micro_summary,
+    measured_support_gap,
+    MetricInputError,
     negative_near_at_threshold,
+    outcome_ptr_at_k,
     pairwise_causal_dominance_ece,
     positive_tangent_recall_at_k,
     positives_closer_than_negatives,
+    proxy_positive_tangent_coverage_at_k,
+    proxy_support_distance,
     selector_regret_at_k,
     support_gap,
 )
 
 __all__ = [
     "bootstrap_ci",
+    "branch_car",
     "branch_causal_action_regret",
     "car_decomposition",
     "macro_micro_summary",
+    "measured_support_gap",
+    "MetricInputError",
     "negative_near_at_threshold",
+    "outcome_ptr_at_k",
     "pairwise_causal_dominance_ece",
     "positive_tangent_recall_at_k",
     "positives_closer_than_negatives",
+    "proxy_positive_tangent_coverage_at_k",
+    "proxy_support_distance",
     "selector_regret_at_k",
     "support_gap",
 ]

workspace/cil/metrics.py

@@ -12,6 +12,10 @@ Vector = Sequence[Number]
 Matrix = Sequence[Vector]
 
 
+class MetricInputError(ValueError):
+    """Raised when a measured metric is requested from proxy-only inputs."""
+
+
 def branch_causal_action_regret(best_measured_branch: Number, selected_branch: Number) -> float:
     """Regret to the best measured branch in a same-state action chart.
 
@@ -22,28 +26,93 @@ def branch_causal_action_regret(best_measured_branch: Number, selected_branch: N
     return max(0.0, float(best_measured_branch) - float(selected_branch))
 
 
-def positive_tangent_recall_at_k(
+def branch_car(best_measured_branch: Number, selected_branch: Number) -> float:
+    """Alias matching the paper metric name BranchCAR."""
+
+    return branch_causal_action_regret(best_measured_branch, selected_branch)
+
+
+def outcome_ptr_at_k(
     generated_utilities: Sequence[Number],
     base_utility: Number,
     *,
     epsilon: float = 0.0,
     k: int | None = None,
+    candidates_evaluated: bool = True,
 ) -> float:
-    """Return 1.0 if any generated proposal improves over the base action."""
+    """OutcomePTR@K for generated candidates with measured outcomes.
+
+    This is a measured metric. It must not be computed from nearest-neighbor
+    distances to hidden positives. Use ``proxy_positive_tangent_coverage_at_k``
+    for distance-only support diagnostics.
+    """
 
+    if not candidates_evaluated:
+        raise MetricInputError(
+            "OutcomePTR@K requires rolled-out/evaluated generated candidates; "
+            "use PPTC@K for distance-only support proxies"
+        )
     prefix = _prefix(generated_utilities, k)
     threshold = float(base_utility) + float(epsilon)
     return 1.0 if any(float(value) > threshold for value in prefix) else 0.0
 
 
+def positive_tangent_recall_at_k(
+    generated_utilities: Sequence[Number],
+    base_utility: Number,
+    *,
+    epsilon: float = 0.0,
+    k: int | None = None,
+) -> float:
+    """Legacy alias for measured OutcomePTR@K.
+
+    New code should call ``outcome_ptr_at_k``. Distance-based support diagnostics
+    must call ``proxy_positive_tangent_coverage_at_k`` and be reported as PPTC.
+    """
+
+    return outcome_ptr_at_k(
+        generated_utilities,
+        base_utility,
+        epsilon=epsilon,
+        k=k,
+        candidates_evaluated=True,
+    )
+
+
+def proxy_positive_tangent_coverage_at_k(
+    generated_tangents: Matrix,
+    positive_tangents: Matrix,
+    *,
+    threshold: float,
+    k: int | None = None,
+) -> float:
+    """PPTC@K: distance-only proxy coverage of measured positive tangents."""
+
+    if threshold < 0:
+        raise ValueError("threshold must be non-negative")
+    generated = _prefix(generated_tangents, k)
+    if not generated or not positive_tangents:
+        return 0.0
+    for tangent in generated:
+        nearest_positive = min(_rms_l2(tangent, positive) for positive in positive_tangents)
+        if nearest_positive <= threshold:
+            return 1.0
+    return 0.0
+
+
 def selector_regret_at_k(
     generated_utilities: Sequence[Number],
     *,
     selected_index: int = 0,
     k: int | None = None,
+    candidates_evaluated: bool = True,
 ) -> float:
     """Selector regret inside the generated proposal prefix."""
 
+    if not candidates_evaluated:
+        raise MetricInputError(
+            "SelectorRegret@K requires measured generated-candidate outcomes"
+        )
     prefix = _prefix(generated_utilities, k)
     if not prefix:
         return 0.0
@@ -58,6 +127,23 @@ def support_gap(best_hidden_chart: Number, best_generated_set: Number) -> float:
     return max(0.0, float(best_hidden_chart) - float(best_generated_set))
 
 
+def measured_support_gap(
+    best_hidden_chart: Number,
+    best_measured_generated_candidate: Number,
+    *,
+    candidates_evaluated: bool = True,
+    generated_subset_of_hidden: bool = False,
+) -> float:
+    """SupportGap when generated candidates have measured outcomes."""
+
+    if not candidates_evaluated and not generated_subset_of_hidden:
+        raise MetricInputError(
+            "SupportGap requires evaluated generated candidates or a generated set "
+            "that is a subset of measured hidden chart branches"
+        )
+    return support_gap(best_hidden_chart, best_measured_generated_candidate)
+
+
 def car_decomposition(
     *,
     best_hidden_chart: Number,
@@ -102,6 +188,24 @@ def car_decomposition(
     return payload
 
 
+def proxy_support_distance(
+    generated_tangents: Matrix,
+    positive_tangents: Matrix,
+    *,
+    k: int | None = None,
+) -> float | None:
+    """Minimum RMS-L2 distance from generated tangents to target positives."""
+
+    generated = _prefix(generated_tangents, k)
+    if not generated or not positive_tangents:
+        return None
+    return min(
+        _rms_l2(tangent, positive)
+        for tangent in generated
+        for positive in positive_tangents
+    )
+
+
 def negative_near_at_threshold(
     generated_tangents: Matrix,
     negative_tangents: Matrix,

workspace/data/cil_charts/train/charts_00000.npz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2561285532fe5f7445b34c726fc592f53863ba6e3ab05ffabd693c8b40e82528
+oid sha256:e310a85a4b772d3013c51c1283d49f6a1365d8309e335d4f2e2c2beb6d97269b
 size 39156136

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 18:44
+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 19:02
 entering extended mode
  restricted \write18 enabled.
  %&-line parsing enabled.
@@ -489,11 +489,11 @@ Package rerunfilecheck Info: File `main.out' has not changed.
 (rerunfilecheck)             Checksum: B3D461B4FEC67164F7152E09ACAEE494;1245.
  ) 
 Here is how much of TeX's memory you used:
- 9900 strings out of 480884
- 146747 string characters out of 5900692
+ 9904 strings out of 480884
+ 146795 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
+ 26983 multiletter control sequences out of 15000+600000
+ 414251 words of font info for 69 fonts, out of 8000000 for 9000
  36 hyphenation exceptions out of 8191
  71i,8n,74p,371b,393s stack positions out of 5000i,500n,10000p,200000b,80000s
 {/cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/font
@@ -521,13 +521,15 @@ share/texmf-dist/fonts/type1/public/amsfonts/cm/cmr9.pfb></cvmfs/soft.computeca
 nada.ca/gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/
 cm/cmsy10.pfb></cvmfs/soft.computecanada.ca/gentoo/2023/x86-64-v3/usr/share/tex
 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 (6 pages, 214742 bytes).
+gentoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/amsfonts/cm/cmtt1
+0.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/ge
+ntoo/2023/x86-64-v3/usr/share/texmf-dist/fonts/type1/public/cm-super/sfrm1000.p
+fb>
+Output written on main.pdf (6 pages, 226804 bytes).
 PDF statistics:
- 158 PDF objects out of 1000 (max. 8388607)
- 130 compressed objects within 2 object streams
+ 162 PDF objects out of 1000 (max. 8388607)
+ 133 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)