Content moderation is inherently policy-dependent: moderation boundaries are often subtle, subjective, and frequently updated, while conventional rule based or fine-tuning based systems require costly rule rewriting, model retraining, or system redeployment. We propose CaseLaw, a policy-decoupled and training-free framework for adaptive content moderation. CaseLaw reformulates moderation as precedent-based retrieval and voting, where multimodal content is transformed into evidence-preserved structured semantics, embedded into a stable semantic space, and classified by thresholded voting over retrieved precedent labels. We train NoteCap on a large-scale corpus of structurally annotated notes to ensure consistent semantic extraction and freeze it during subsequent policy adaptation, which requires no additional training. Policy adaptation is achieved by updating the semantic case library, including case labels and mappings, without retraining models or modifying prompts. To evaluate CaseLaw, we construct MFO-RedNotes, a benchmark for the male-female opposition risk domain. We additionally provide a five category CoPro comparison. Experiments show that CaseLaw consistently outperforms competing baselines in both precision and recall across policy-dependent risk levels and broader safety categories, with category-wise precision/recall reaching up to 86.8%/90.1% on MFO-RedNotes across both policy settings and 94.5%/92.3% on CoPro. Our work highlights the potential of precedent-based reasoning for policy adaptive content moderation and offers a practical direction toward more stable, transparent, and training-free moderation systems.
Figures & tables
Figure 1: Comparison of policy adaptation in conventional moderation pipelines and CaseLaw. Policy changes in conventional pipelines can require rule rewriting, model retraining, and redeployment of updated moderation components. CaseLaw adapts by updating the external case library while keeping its deployed models fixed.
Figure 2: Overview of CaseLaw. NoteCap converts text, images, video, and audio into evidence-preserving structured fields. A frozen encoder embeds the serialized fields, enabling retrieval of precedent cases from an editable library. Each case whose similarity meets or exceeds its stored threshold casts a single vote based on its current risk label, without invoking an LLM classifier after retrieval. Policy changes are implemented by updating precedent labels and metadata while keeping the deployed representation models fixed.
Method
Safe
Low-risk
Mid-risk
High-risk
Prec (%)
Rec (%)
Prec (%)
Rec (%)
Prec (%)
Rec (%)
Prec (%)
Rec (%)
Original policy
Zero-shot
56.6
54.5
39.7
41.0
52.0
53.0
50.7
50.7
Fixed few-shot
71.5
72.2
55.4
56.0
64.6
63.4
68.1
67.8
Fine-tuning
78.1
85.0
73.3
71.0
76.3
74.0
81.2
76.2
Class-RAG
81.0
85.3
75.8
74.5
78.5
76.3
82.5
78.3
Table 1: Comparison on MFO-RedNotes before and after policy change. Precision and recall are expressed as percentages for four policy-dependent risk levels. Yellow highlights the best score, and green highlights the strongest baseline in each column. Superscripts report absolute gains in percentage points over the strongest baseline within the same policy setting.
Method
Safe
Sexual
Hate
Violence
Shocking
Prec (%)
Rec (%)
Prec (%)
Rec (%)
Prec (%)
Rec (%)
Prec (%)
Rec (%)
Prec (%)
Rec (%)
Zero-shot
48.4
51.2
51.8
49.1
50.9
55.8
53.8
60.4
56.6
43.9
Fixed few-shot
65.6
68.2
71.7
71.3
65.4
64.7
70.5
76.1
66.2
59.3
Fine-tuning
80.0
82.0
84.7
83.2
80.6
84.8
83.1
85.9
81.7
74.1
Class-RAG
83.0
84.3
87.8
87.0
82.8
87.2
86.2
88.0
84.2
77.4
Llama Guard
82.2
83.1
88.1
87.0
81.6
84.3
87.0
82.3
82.6
84.5
Table 2: Comparison on the five-category CoPro benchmark. Precision and recall are expressed as percentages. Yellow highlights the best score, and green highlights the strongest baseline in each column. Superscripts report absolute gains in percentage points over the strongest baseline. Evaluation uses 5,000 validation instances, with 1,000 instances per category.
Figure 3: Impact of structured semantic extraction on embedding distributions. We visualize 1,000 safe notes and 1,000 risk notes from MFO-RedNotes, where risk includes Mid-risk and High-risk instances. Compared with direct embedding of raw notes and summary-only representations, NoteCap produces more compact intra-class clusters and clearer inter-class separation. The concentrated risk regions further indicate that structured semantic extraction can form risk-oriented semantic anchors for more accurate downstream case retrieval.
Method
Safe
Unsafe
Prec
Rec
Var(P)
Var(R)
Prec
Rec
Var(P)
Var(R)
MFO-RedNotes
Direct splicing
66.0
68.8
0.00
0.00
67.4
64.6
0.00
0.00
Summary only
73.0
74.5
0.09
0.09
74.0
72.5
0.09
0.09
NoteCap
85.6
88.0
0.00
0.00
87.7
85.2
0.00
0.00
CoPro
Table 3: Representation ablation on MFO-RedNotes and CoPro. The MFO-RedNotes subset contains 1,000 Safe and 1,000 Unsafe notes. Prec and Rec are means over five runs (%); Var(P) and Var(R) are sample variances in squared percentage points (pp 2 ).
Extractor
Safe
Unsafe
Top5-diff
Prec
Rec
Var(P)
Var(R)
Top5-diff
Prec
Rec
Var(P)
Var(R)
Gemini-3.1-Pro
16.8
77.5
84.4
0.20
0.22
12.0
94.6
91.8
0.02
0.05
GPT-6.1-sol
11.6
80.0
86.2
0.12
0.13
7.6
95.3
92.8
0.01
0.03
Qwen-3.8-Max
22.6
75.5
82.9
0.31
0.53
17.5
94.1
91.0
0.05
0.08
NoteCap
0.0
85.3
89.0
0.00
0.00
0.0
96.3
94.9
0.00
0.00
Table 4: Five-run extraction stability and moderation performance on the MFO-RedNotes validation split. Evaluation uses the original policy and voting depth k=16 . Unsafe combines Low-risk, Mid-risk, and High-risk. Top5-diff is the percentage of notes whose top-five retrieved case sets change across five runs. Precision and recall are five-run means (%); their sample variances are measured in squared percentage points (pp 2 ).
Figure 4: Voting-depth ablation under the original-policy setting of Table 1 , with all other experimental conditions fixed. Precision (top) and recall (bottom) are expressed as percentages. Blue dashed squares denote Safe and pink solid circles denote Unsafe.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Risk Level
Typical Manifestation
Safe
No obvious risk of gender-based antagonism. The content is typically neutral, factual, or rights-oriented, and respects diverse viewpoints without inflammatory framing.
Low Risk
The content may involve sensitive gender-related topics but remains restrained in tone. It usually presents controversial opinions in a moderate manner without explicit hostility, group attribution, or provocative escalation.
Medium Risk
The content shows a clear antagonistic tendency but does not reach an extreme or explicit violation. It often uses oppositional framing, implicit group attribution, emotionally charged claims, or partially matches policy-defined risky patterns.
High Risk
The content exhibits strong antagonism or directly violates the moderation policy. Typical manifestations include original-sin rhetoric, extreme demands, explicit abuse, hate speech, deliberate provocation, or intensified gender-based opposition.
Appendix
Table 5: Risk-level taxonomy for the male-female opposition moderation domain. Each level reflects increasing severity of gender-based antagonism and is defined according to expert-designed content-safety policies used for MFO-RedNotes.
Figure 5: Impact of different structured semantic extraction backbones on embedding distributions. We visualize the representations produced by (a) InternVL3-14B-Instruct, (b) Llama-4-Scout-17B, and (c) NoteCap on MFO-RedNotes. NoteCap produces more compact intra-class clusters and clearer safe–risk separation, indicating better formation of risk-oriented semantic anchors for downstream retrieval.
Method
braincels
theredpill
chapotraphouse
Prec (%)
Rec (%)
Prec (%)
Rec (%)
Prec (%)
Rec (%)
Few-shot
68.3
69.8
64.9
69.2
70.6
67.1
Class-RAG
83.7
84.6
81.9
83.2
78.2
79.3
Llama Guard
83.6
85.8
80.6
87.4
85.5
80.3
CaseLaw (ours)
87.9 ↑4.2
90.3 ↑4.5
86.7 ↑4.8
92.1 ↑4.7
89.8 ↑4.3
90.5 ↑10.2
Appendix
Table 6: Illustrative moderation performance on The Big Ban Theory benchmark. Precision and recall are expressed as percentages for three subreddit subsets. Superscripts indicate gains in percentage points over the strongest baseline in each column.
Cases
Macro-P
Macro-R
Macro-F1
1,000
70.3
60.7
64.0
5,000
75.4
70.0
72.3
20,000
79.9
77.3
78.4
40,000
84.3
83.6
83.8
Appendix
Table 7: Effect of case-library size on moderation performance. Label proportions, evaluation queries, and retrieval settings are fixed. All metrics are percentages. Macro-F1 averages class-wise F1 scores.
Largest:smallest
Macro-F1
Worst-label recall
1:1
83.8
76.6
2:1
81.8
71.9
4:1
76.5
63.7
8:1
69.3
51.4
Appendix
Table 8: Effect of label imbalance in the case library. Library size, evaluation queries, and retrieval settings are fixed. The balanced condition bases on the 40,000-case reference configuration in Table 7 . All metrics are percentages.
Near-duplicate share
Clusters
Macro-F1
0%
40,000
83.8
25%
33,848
81.8
50%
25,616
76.3
75%
17,240
69.5
Appendix
Table 9: Effect of semantic redundancy in the case library. The 40,000-case reference configuration matches Table 7 . Near-duplicate share is the fraction of cases belonging to clusters containing at least two cases; the cluster count includes singleton clusters. Macro-F1 is expressed as a percentage.
Figure 6: Evidence-preserving extraction for an allegation and warning. Identifying and sensitive details are masked.