# VERDICT WEIGHT

> Confidence scoring framework for autonomous AI systems. USPTO-protected, IEEE-validated, defense-ready.

## Docs

- [Audit logging](https://verdictweight.dev/api/audit-logging.md): Configuring, verifying, and managing the cryptographic audit chain.
- [Hyperparameters](https://verdictweight.dev/api/hyperparameters.md): Configurable surface across the eight streams. Defaults are conservative; tune deliberately.
- [Pipeline](https://verdictweight.dev/api/pipeline.md): Streaming and batched scoring patterns for production deployment.
- [Registry](https://verdictweight.dev/api/registry.md): Configuration registry, integrity hashing, and kill-switch interaction.
- [Scorer](https://verdictweight.dev/api/scorer.md): The primary entry point for single-decision confidence scoring.
- [Streams](https://verdictweight.dev/api/streams.md): Direct access to individual stream evaluators for debugging, ablation, and custom composition.
- [Thresholds](https://verdictweight.dev/api/thresholds.md): How to choose and validate confidence thresholds for gating, escalation, and abstention.
- [Completeness proof](https://verdictweight.dev/architecture/completeness-proof.md): Formal argument that the eight streams cover the failure modes the framework targets.
- [Eight-stream composition](https://verdictweight.dev/architecture/eight-stream-composition.md): The formal composition rule that combines the eight streams into a single confidence score.
- [Architecture overview](https://verdictweight.dev/architecture/overview.md): How the eight streams compose into a single defensible confidence score.
- [Competitive landscape](https://verdictweight.dev/competitive/overview.md): How VERDICT WEIGHT differs from AI security platforms, calibration libraries, and LLM observability tools.
- [vs. AI security platforms](https://verdictweight.dev/competitive/vs-ai-security-platforms.md): How VERDICT WEIGHT differs from HiddenLayer, Robust Intelligence, Lakera, Calypso, ProtectAI.
- [vs. calibration libraries](https://verdictweight.dev/competitive/vs-calibration-libraries.md): How VERDICT WEIGHT differs from Netcal, Uncertainty Toolbox, and scikit-learn calibration utilities.
- [vs. LLM observability](https://verdictweight.dev/competitive/vs-llm-observability.md): How VERDICT WEIGHT differs from Arize, Fiddler, Arthur, WhyLabs, and the broader ML observability category.
- [DoD AI Ethical Principles](https://verdictweight.dev/compliance/dod-ai-ethical-principles.md): How VERDICT WEIGHT supports the Department of Defense's five AI Ethical Principles.
- [EU AI Act](https://verdictweight.dev/compliance/eu-ai-act.md): How VERDICT WEIGHT supports compliance with the EU AI Act's high-risk system requirements (Articles 9-15) and transparency obligations.
- [ISO/IEC 42001](https://verdictweight.dev/compliance/iso-iec-42001.md): How VERDICT WEIGHT supports ISO/IEC 42001 AI management system requirements.
- [NIST AI RMF 1.0](https://verdictweight.dev/compliance/nist-ai-rmf.md): How VERDICT WEIGHT supports the Govern, Map, Measure, and Manage functions of the NIST AI Risk Management Framework.
- [Compliance mappings](https://verdictweight.dev/compliance/overview.md): How VERDICT WEIGHT's eight streams map to the major AI governance frameworks.
- [AFWERX CSO](https://verdictweight.dev/gov/afwerx-cso.md): The active government acquisition pathway for VERDICT WEIGHT pilots.
- [Audit and compliance](https://verdictweight.dev/gov/audit-and-compliance.md): Why VERDICT WEIGHT's audit chain meets the standard required for regulated and defense deployments.
- [The Curveball attack class](https://verdictweight.dev/gov/curveball-attack-class.md): Confidence-flip attacks on autonomous AI systems. What they are, why they matter, what Stream 6 does.
- [Pilot engagement](https://verdictweight.dev/gov/pilot-engagement.md): How to scope, run, and evaluate a VERDICT WEIGHT pilot in a high-stakes deployment.
- [Threat model](https://verdictweight.dev/gov/threat-model.md): What VERDICT WEIGHT defends against, and why that matters in defense and critical-infrastructure deployments.
- [VERDICT WEIGHT](https://verdictweight.dev/index.md): A confidence scoring framework for autonomous AI systems. USPTO-protected, IEEE-validated, defense-ready.
- [Install from source](https://verdictweight.dev/install/from-source.md): Build VERDICT WEIGHT from the GitHub repository for development or audit.
- [Install from PyPI](https://verdictweight.dev/install/pypi.md): The standard installation path for production use.
- [Verifying your install](https://verdictweight.dev/install/verification.md): Confirm the installed package matches the audited release.
- [Quickstart](https://verdictweight.dev/introduction/quickstart.md): Score your first decision with VERDICT WEIGHT in under five minutes.
- [What is VERDICT WEIGHT?](https://verdictweight.dev/introduction/what-is-verdict-weight.md): An eight-stream composition that turns model outputs into auditable, calibrated confidence scores.
- [Why it matters](https://verdictweight.dev/introduction/why-it-matters.md): Confidence is the missing layer between deployable AI and auditable AI.
- [Stream 1: Evidence aggregation](https://verdictweight.dev/streams/01-evidence-aggregation.md): Uncertainty-aware fusion of model outputs, retrieval signals, and structured priors.
- [Stream 2: Uncertainty quantification](https://verdictweight.dev/streams/02-uncertainty-quantification.md): Decomposing total uncertainty into aleatoric and epistemic components.
- [Stream 3: Temporal stability](https://verdictweight.dev/streams/03-temporal-stability.md): Penalizes confidence that fluctuates across semantically equivalent inputs.
- [Stream 4: Cross-source coherence](https://verdictweight.dev/streams/04-cross-source-coherence.md): Rewards corroboration across independent evidence sources; surfaces contradiction.
- [Stream 5: Calibration](https://verdictweight.dev/streams/05-calibration.md): Post-hoc reliability correction so reported confidence matches empirical correctness.
- [Stream 6: SIS / Curveball detection](https://verdictweight.dev/streams/06-sis-curveball-detection.md): Detects adversarial inputs designed to manipulate confidence without flipping the prediction.
- [Stream 7: CPS / hash-chain integrity](https://verdictweight.dev/streams/07-cps-hash-chain-integrity.md): Cryptographic provenance for every scoring event. Tamper-evident audit by construction.
- [Stream 8: RIS / registry kill switch](https://verdictweight.dev/streams/08-ris-registry-kill-switch.md): Binary, registry-level abort condition. Last line of defense against scoring-layer compromise.
- [Coverage overview](https://verdictweight.dev/testing/coverage-overview.md): 673 tests across 27 suites covering fuzz, mutation, differential, regression, concurrency, and formal verification.
- [Formal verification](https://verdictweight.dev/testing/formal-verification.md): Property-based and constraint-solver checks of the composition rule and audit-chain invariants.
- [Fuzz and mutation testing](https://verdictweight.dev/testing/fuzz-and-mutation.md): Two complementary approaches to finding bugs the unit tests miss.
- [AI security operations](https://verdictweight.dev/use-cases/ai-security-soc.md): Vulnerability triage and threat detection where confidence determines action priority. Direct alignment with the published CVE/KEV validation.
- [Defense autonomy](https://verdictweight.dev/use-cases/defense-autonomy.md): Confidence-gated autonomous decisioning in adversarial environments. Curveball-class threats. Tamper-evident audit.
- [Use cases](https://verdictweight.dev/use-cases/overview.md): Scoped scenarios where VERDICT WEIGHT's eight-stream composition produces concrete operational value.
- [Regulated industry](https://verdictweight.dev/use-cases/regulated-industry.md): AI decisioning in healthcare, finance, and legal contexts where audit defensibility is mandatory.
- [Ablation studies](https://verdictweight.dev/validation/ablation-studies.md): Empirical demonstration that every stream is necessary. Removing any one degrades coverage.
- [Calibration curves](https://verdictweight.dev/validation/calibration-curves.md): Empirical reliability of VERDICT WEIGHT's confidence values. ~9.6× better than averaging.
- [Complexity analysis](https://verdictweight.dev/validation/complexity-analysis.md): VERDICT WEIGHT runs in O(1) per decision in the core path; O(K) in the pipeline path.
- [CVE dataset](https://verdictweight.dev/validation/cve-dataset.md): 120 real CVEs from NIST NVD and CISA KEV used as a real-world proxy for benchmarking.
- [Head-to-head comparison](https://verdictweight.dev/validation/head-to-head-comparison.md): VERDICT WEIGHT benchmarked against Dempster-Shafer, Naive Bayes, Simple Averaging, and Max Voting.
- [Known limitations](https://verdictweight.dev/validation/known-limitations.md): Honest enumeration of what the validation does and does not establish.
- [NVD / KEV methodology](https://verdictweight.dev/validation/nvd-kev-methodology.md): How CVE records and KEV entries are mapped to evidence vectors and ground-truth labels.

## OpenAPI Specs

- [openapi](https://verdictweight.dev/api-reference/openapi.json)

## Optional

- [GitHub](https://github.com/Odingard/verdict-weight)
- [PyPI](https://pypi.org/project/verdict-weight/)
- [Zenodo (DOI)](https://doi.org/10.5281/zenodo.19447547)