Update README.md
Browse files
README.md
CHANGED
|
@@ -1,12 +1,11 @@
|
|
| 1 |
---
|
| 2 |
-
title: Agentic Reliability Framework
|
| 3 |
-
emoji:
|
| 4 |
colorFrom: blue
|
| 5 |
-
colorTo:
|
| 6 |
sdk: static
|
| 7 |
pinned: false
|
| 8 |
license: apache-2.0
|
| 9 |
-
short_description: Bayesian governance for agentic systems
|
| 10 |
---
|
| 11 |
|
| 12 |
# Agentic Reliability Framework (ARF) – Technical Overview
|
|
@@ -19,29 +18,46 @@ short_description: Bayesian governance for agentic systems
|
|
| 19 |
|
| 20 |
## 🧠 Core Engine (Protected)
|
| 21 |
|
| 22 |
-
|
| 23 |
|
| 24 |
-
|
|
|
|
|
|
|
| 25 |
\text{risk} = w_{\text{conj}}\cdot\frac{\alpha}{\alpha+\beta} \;+\; w_{\text{hmc}}\cdot p_{\text{hmc}} \;+\; w_{\text{hyper}}\cdot \mu_{\text{hyper}}
|
| 26 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
|
| 28 |
-
|
| 29 |
-
- Posterior variance → 90% HDI for uncertainty quantification.
|
| 30 |
|
| 31 |
-
|
| 32 |
|
| 33 |
-
|
| 34 |
\begin{aligned}
|
| 35 |
L_{\text{approve}} &= \text{COST\_FP}\cdot R + \text{COST\_IMPACT}\cdot b_{\text{mean}} + \dots \\
|
| 36 |
L_{\text{deny}} &= \text{COST\_FN}\cdot(1-R) + \text{COST\_OPP}\cdot v_{\text{mean}} \\
|
| 37 |
L_{\text{escalate}} &= \text{COST\_REVIEW} + \text{COST\_UNCERTAINTY}\cdot\psi
|
| 38 |
\end{aligned}
|
| 39 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
|
| 41 |
-
|
| 42 |
-
**Lyapunov stability** – Quadratic candidate \(V(x,r) = \alpha r^2 + \beta\|x-x_{\text{des}}\|^2\) ensures healing actions converge.
|
| 43 |
|
| 44 |
-
|
| 45 |
|
| 46 |
---
|
| 47 |
|
|
|
|
| 1 |
---
|
| 2 |
+
title: Agentic Reliability Framework (ARF) – Technical Overview
|
| 3 |
+
emoji: 🧠
|
| 4 |
colorFrom: blue
|
| 5 |
+
colorTo: indigo
|
| 6 |
sdk: static
|
| 7 |
pinned: false
|
| 8 |
license: apache-2.0
|
|
|
|
| 9 |
---
|
| 10 |
|
| 11 |
# Agentic Reliability Framework (ARF) – Technical Overview
|
|
|
|
| 18 |
|
| 19 |
## 🧠 Core Engine (Protected)
|
| 20 |
|
| 21 |
+
### Bayesian Risk Fusion
|
| 22 |
|
| 23 |
+
Conjugate Beta priors (online) + HMC logistic regression (offline) + optional hierarchical hyperpriors.
|
| 24 |
+
|
| 25 |
+
$$
|
| 26 |
\text{risk} = w_{\text{conj}}\cdot\frac{\alpha}{\alpha+\beta} \;+\; w_{\text{hmc}}\cdot p_{\text{hmc}} \;+\; w_{\text{hyper}}\cdot \mu_{\text{hyper}}
|
| 27 |
+
$$
|
| 28 |
+
|
| 29 |
+
Weights are dynamic:
|
| 30 |
+
|
| 31 |
+
$$
|
| 32 |
+
w_{\text{hmc}} = \min\left(0.6,\; \frac{n}{n_0}\right), \qquad
|
| 33 |
+
w_{\text{hyper}} = \min\left(0.3,\; \text{base}\cdot(1-w_{\text{hmc}})\right)
|
| 34 |
+
$$
|
| 35 |
+
|
| 36 |
+
Posterior variance → 90% HDI for uncertainty quantification.
|
| 37 |
|
| 38 |
+
### Expected Loss Minimisation
|
|
|
|
| 39 |
|
| 40 |
+
Chooses `APPROVE`, `DENY`, or `ESCALATE` by minimising:
|
| 41 |
|
| 42 |
+
$$
|
| 43 |
\begin{aligned}
|
| 44 |
L_{\text{approve}} &= \text{COST\_FP}\cdot R + \text{COST\_IMPACT}\cdot b_{\text{mean}} + \dots \\
|
| 45 |
L_{\text{deny}} &= \text{COST\_FN}\cdot(1-R) + \text{COST\_OPP}\cdot v_{\text{mean}} \\
|
| 46 |
L_{\text{escalate}} &= \text{COST\_REVIEW} + \text{COST\_UNCERTAINTY}\cdot\psi
|
| 47 |
\end{aligned}
|
| 48 |
+
$$
|
| 49 |
+
|
| 50 |
+
### Execution Ladder (Rust)
|
| 51 |
+
|
| 52 |
+
Mechanical gates: `license`, `confidence`, `risk`, `rollback`, `causal`.
|
| 53 |
+
|
| 54 |
+
### Lyapunov Stability
|
| 55 |
+
|
| 56 |
+
Quadratic candidate $V(x,r) = \alpha r^2 + \beta\|x - x_{\text{des}}\|^2$ ensures healing actions converge.
|
| 57 |
|
| 58 |
+
### Cryptographic Signing
|
|
|
|
| 59 |
|
| 60 |
+
Ed25519 signatures for `HealingIntent` (Python `cryptography`, Rust `ed25519_dalek`).
|
| 61 |
|
| 62 |
---
|
| 63 |
|