Scenario Configuration
🔬 Inference pipeline (how variables would be computed in practice)
| r_sacred, r_coercion, r_extraction | Fine-tuned intent classifiers on labelled prompt datasets; probability outputs [0–1]. |
| MBI_t | Recurrent state model, EWMA + shock update each turn (see Auto-Update below). |
| p_t | Model calibration score; estimated from token probability distributions. |
| H_t | Shannon entropy over top-k token distribution at generation time. |
| z_t | Explicit consent flag from system prompt or user acknowledgment; binary. |
⚙️ MBI Auto-Update Mode (Stateful)
MBI_t = clip[(1−η)·MBI_(t-1) + η·(1−S_t) − κ·max(0, S_t − τ_s)]
Guardian Gate Cascade
Rule: IF (z_t = 0 AND sensitive) → PERMISSION_REQUEST
Rule: IF E_t ≥ τ_E → CONTAIN
Eₜ is a composite interaction risk proxy — it reflects model-observable interaction pressure, not normative ethical judgement. Inherently noisy; requires empirical calibration. See panel below for full definition.
Rule: IF MBI_t < τ_MBI → MIRROR_USER_FRAME
MIRROR_USER_FRAME: When boundary integrity erodes, the Guardian stops asserting its own frame and reflects the user's stated goals back — a de-escalation posture that re-establishes shared grounding before continuing.
📐 MBI Formal Definition ORIGINAL
defaults: w_s=0.4, w_c=0.3, w_e=0.3
(1−η)·MBI_(t-1) + η·(1−S_t)
− κ·max(0, S_t − τ_s)
)
⚠ Ethical Pressure (Eₜ) — Composite Interaction Risk Proxy
Eₜ is a composite interaction risk proxy, not a direct measure of ethical correctness or user intent. It is designed to capture structural signals of alignment instability within a model–user interaction.
Rather than attempting to evaluate morality or truth, Eₜ estimates the degree to which an interaction exhibits patterns historically associated with:
- · coercion or manipulation
- · boundary erosion
- · role or identity destabilisation
- · extraction or prompt injection attempts
- · escalating adversarial framing
These signals are derived from observable linguistic and contextual features using probabilistic classifiers and uncertainty estimates. Eₜ therefore reflects model-observable interaction pressure, not normative ethical judgement.
Because Eₜ is a proxy, it is inherently noisy and subject to calibration error. False positives and false negatives are expected, and threshold values must be tuned empirically using adversarial and benign datasets.
The goal of Eₜ is not to detect "good" or "bad" users, but to provide an interpretable state variable that enables dynamic, trajectory-aware governance during interaction.
α = [0.4, 0.3, 0.2, 0.1] — illustrative; reflects prior that identity destabilisation is highest-cost
Future work: dataset-driven calibration · uncertainty-aware thresholding · adaptive weighting across interaction contexts.