Skip to main content

Documentation Index

Fetch the complete documentation index at: https://verdictweight.dev/llms.txt

Use this file to discover all available pages before exploring further.

The problem

Modern AI systems — especially autonomous and agentic ones — produce decisions without producing trustworthy confidence about those decisions. Softmax probabilities are miscalibrated. Self-reports from LLMs are aspirational. Ensembles average away the signal that matters most: when not to act. In high-stakes deployments, the absence of a defensible confidence layer is the gap between “deployable AI” and “auditable AI.”

The framework

VERDICT WEIGHT closes that gap by composing eight independent evidence streams into a single confidence score:
1

Evidence aggregation (Stream 1)

Combines model outputs, retrieval signals, and structured priors using uncertainty-aware fusion rather than naive averaging.
2

Uncertainty quantification (Stream 2)

Decomposes total uncertainty into aleatoric and epistemic components, exposing what the system cannot know.
3

Temporal stability (Stream 3)

Penalizes confidence that fluctuates across semantically equivalent inputs — a known LLM failure mode.
4

Cross-source coherence (Stream 4)

Cross-checks the decision against independent signal sources; rewards corroboration, surfaces contradiction.
5

Calibration (Stream 5)

Applies post-hoc reliability correction so reported confidence matches empirical correctness.
6

SIS / Curveball detection (Stream 6)

Detects adversarial inputs designed to flip a system’s confidence without flipping its prediction.
7

CPS / hash-chain integrity (Stream 7)

Cryptographically chains every scoring event to its predecessor, producing a tamper-evident audit log.
8

RIS / registry kill switch (Stream 8)

A binary, registry-level abort condition that overrides composed confidence when integrity is compromised.

What you get

  • A single calibrated confidence score suitable for thresholding, gating, or escalation.
  • A decomposed evidence trail showing which streams agreed, disagreed, or abstained.
  • A cryptographic provenance chain suitable for after-action review, regulatory audit, or legal discovery.
  • A kill switch that fires deterministically when adversarial or integrity conditions are detected.
VERDICT WEIGHT is a scoring layer, not a model. It composes signals from whatever model stack you already run. It is model-agnostic, vendor-agnostic, and has no external runtime dependencies on cloud services.