stella-ops.org/git.stella-ops.org
1
0
Fork 0
Code Issues Pull Requests Actions 2 Releases Wiki Activity

save checkpoint: save features

Browse Source
This commit is contained in:
master
2026-02-12 10:27:23 +02:00
parent dca86e1248
commit 5bca406787
8837 changed files with 1796879 additions and 5294 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
docs/features/unimplemented/binaryindex/ensemble-decision-engine-for-multi-tier-matching.md Normal file
View File

@@ -0,0 +1,37 @@
# Ensemble decision engine for multi-tier matching
## Module
BinaryIndex
## Status
PARTIALLY_IMPLEMENTED
## Description
Ensemble decision engine combines multiple matching tiers (range match, Build-ID, fingerprint) with configurable weight tuning for vulnerability classification.
## Implementation Details
- **Modules**: `src/BinaryIndex/__Libraries/StellaOps.BinaryIndex.Ensemble/`
- **Key Classes**:
- `EnsembleDecisionEngine` (`src/BinaryIndex/__Libraries/StellaOps.BinaryIndex.Ensemble/EnsembleDecisionEngine.cs`) - combines multiple matching signals with configurable weights into a final vulnerability classification decision
- `FunctionAnalysisBuilder` (`src/BinaryIndex/__Libraries/StellaOps.BinaryIndex.Ensemble/FunctionAnalysisBuilder.cs`) - builds function analysis inputs including optional ML embeddings
- `WeightTuningService` (`src/BinaryIndex/__Libraries/StellaOps.BinaryIndex.Ensemble/WeightTuningService.cs`) - tunes ensemble weights based on golden set validation results
- `EnsembleOptions` (`src/BinaryIndex/__Libraries/StellaOps.BinaryIndex.Ensemble/Models.cs`) - configurable weights and thresholds for matching tiers
- `MlEmbeddingMatcherAdapter` (`src/BinaryIndex/__Libraries/StellaOps.BinaryIndex.ML/Training/MlEmbeddingMatcherAdapter.cs`) - adapts ML function embeddings for ensemble use
- **Interfaces**: `IEnsembleDecisionEngine` (`src/BinaryIndex/__Libraries/StellaOps.BinaryIndex.Ensemble/IEnsembleDecisionEngine.cs`)
- **Registration**: `EnsembleServiceCollectionExtensions.AddBinarySimilarityServices()` for full pipeline setup
- **Benchmarks**: `EnsembleAccuracyBenchmarks`, `EnsembleLatencyBenchmarks` (`src/BinaryIndex/__Tests/StellaOps.BinaryIndex.Benchmarks/`)
## E2E Test Plan
- [ ] Submit a binary with known vulnerability and verify ensemble produces correct classification
- [ ] Verify weight tuning: adjust instruction weight to 0.6 and verify it changes classification outcomes
- [ ] Verify multi-tier integration: Build-ID match, fingerprint match, and ML embedding all contribute to score
- [ ] Verify `FunctionAnalysisBuilder` correctly assembles all matching dimensions
- [ ] Verify `WeightTuningService` optimizes weights based on golden set validation accuracy
- [ ] Run accuracy benchmark and verify F1 score meets minimum threshold
## Verification
- Run: `run-001` (2026-02-11 UTC).
- Tier 1/2 builds and tests passed (`37/37`), but parity review found contract mismatch and missing coverage for key claims.
- Ensemble signal model currently exposes syntactic, semantic, embedding, and exact-hash signals, but the feature contract claims range-match, Build-ID, and fingerprint tiers (`src/BinaryIndex/__Libraries/StellaOps.BinaryIndex.Ensemble/Models.cs:232`).
- `FunctionAnalysisBuilder` explicitly retains a simplified semantic-graph path when binary data is unavailable (`src/BinaryIndex/__Libraries/StellaOps.BinaryIndex.Ensemble/FunctionAnalysisBuilder.cs:87`).
- No direct tests were found for `FunctionAnalysisBuilder` or `MlEmbeddingMatcherAdapter` in `src/BinaryIndex/__Tests/StellaOps.BinaryIndex.Ensemble.Tests`.