Glossary

The identity and runtime governance layer for AI systems. Encodes who you are — so AI acts in coherence with your voice, constraints, and decision logic.

Your compressed identity signature. Encodes tone, values, decision architecture, and behavioral boundaries in a portable symbolic format.

Your personal ArcKernel configuration. Includes your IDNA, boundary rules, tone weights, and drift thresholds — generated via calibration.

A deterministic safety circuit. Measures semantic drift in real time and blocks outputs that cross predefined risk thresholds. Triggers human handoff if needed.

The gradual erosion of AI behavior away from your encoded identity. Often manifests as increasingly generic or sycophantic output.

Continuous monitoring of AI responses against your IDNA signature. Alerts or blocks when output exceeds semantic drift tolerance.

The symbolic constants of your identity — tone, decision priorities, emotional triggers — that remain stable across context and time. IDNA encodes these.

Structured transition protocol from AI to human agent. Includes compressed context: intent, emotional state, friction point, and communication preferences.

The constraint that all identity and drift parameters must compress into 3–8KB. Enables fast transfer, predictable performance, and constant memory overhead.

ArcKernel operates independently of model provider (Claude, GPT, Gemini). IDNA persists across platforms — ensuring portability and continuity.

Symbolic memory compression module. Encodes episodic identity-relevant information into compressed form for reuse.

Predictive recursion module. Forecasts conversational drift and trajectory misalignment before it manifests — enabling preemptive interruption.

Method Constraint Enforcement module. Governs HOW the agent reasons, not just WHAT it outputs. Catches violations where scope is fine but reasoning methodology is wrong. Finance: 99.2% compliance. Legal: 97.6%.

Output register modulation system with six discrete levels (O0–O5). Controls the abstraction depth and creative latitude of agent output. Same agent, different voice, enforced by the kernel. O1 (Analyst): 86–100% compliance across 12 models.

Three-stage structural coherence monitor: detect, patch, realign. Catches structural degradation that HALT alone misses (57 events). Operates independently from HALT (Pearson r = 0.40). The module that watches the watchers.

Output fidelity verifier. Produces three-dimension scoring (policy adherence, intent fidelity, behavioral consistency) with a configurable verdict gate (verified / requires review / rejected). 201 complementarity events with HALT.

Full-stack orchestration module. Coordinates all governance modules into a single validated loop snapshot. Zero interaction failures (0/192). Order-independent (delta 0.0007). All modules contribute positive marginal value.

Trust ledger. Tracks identity-aligned decisions across time. Produces per-action fidelity scores, drift flags, and loop closure reports. Exportable compliance evidence for Article 12 record-keeping.

Immutable audit trail. Creates cryptographically verified per-action records: IDNA state, drift score, decision vector, tool authorization, and kernel witness hash. Tamper-evident and forensic-grade.

Area Under the Precision-Recall Curve. Measures how well a system distinguishes real events from false alarms under class imbalance. mOm5 achieves 0.54 vs 0.37 baseline (breach base rate ~18%).

Cryptographic hash verification used by mOm4 to ensure canonical identity state has not been tampered with or corrupted. 100% integrity across 1,857 checks.

Reduction in total tokens a model generates under governance. mOm4 achieves 7% cumulative compression on GPT-4o-mini at 75-turn depth — reducing cost and latency simultaneously.

Net latency impact where governed responses are faster than ungoverned. ArcKernel's identity compression narrows the model's search space, producing equivalent output faster. Range: −70ms (Mistral) to −2,424ms (Grok).

Correlation coefficient between −1 and 1. Values near 0 indicate statistical independence. DriftDefenseStack and HALT show r = 0.40, confirming they detect different failure types rather than duplicating each other.