Custom AI Filters with GHOST Architecture
Intent-layer AI safety via UIP behavioral probes and LHTM activation monitoring — custom policies for any organization.
If you are searching for custom AI content filters, intent-based moderation, or AI safety architecture beyond regex and LlamaGuard, you have found the right layer of the stack.
KosherChat is built on GHOST — Generative Harm Observation via Subspace Trajectories — a dual-mode safety architecture that filters at the intent layer, not surface text.
The problem with surface filters
Every mainstream guardrail — system prompts, embedding classifiers, NeMo Guardrails, LlamaGuard — reads text. Attackers modify surface phrasing while preserving harmful intent. The jailbreak literature proves this is not a calibration bug; it is structural.
GHOST changes the object being filtered: from "what does this sentence look like?" to "what is this model being steered toward producing?"
GHOST components
UIP — Unconstrained Intent Probe (black-box path)
The user's message runs through a small, safety-unaligned local probe. The probe fulfills the underlying request without being manipulated by jailbreak framing. The behavioral output reveals intent independent of polite wording, hypotheticals, or roleplay.
LHTM — Latent Harm Trajectory Monitoring (white-box path)
During main-model generation, GHOST monitors residual-stream activations for harmful trajectory signatures — the same internal signals studied in mechanistic interpretability research on refusal directions.
Why both?
They share no common failure mode. A jailbreak that fools surface text may still trigger LHTM. A jailbreak that suppresses internal signals may expose itself through UIP behavioral reconstruction.
USER MESSAGE │ ├── UIP Gate (behavioral intent) └── LHTM Monitor (activation trajectory) │ ALLOW / BLOCK
Custom filters for your organization
GHOST is not a single policy — it is an architecture you can map to community standards:
Implementation layers:
1. Default harm taxonomy — sexual content, violence, drugs, harassment 2. Community overlay — custom skills and policy templates 3. Probe calibration — tune UIP sensitivity for your risk profile 4. Monitoring — trajectory interrupts before users see violations
Comparison table
| Filter type | What it reads | Bypass strategy | GHOST |
| Keyword list | Tokens | Synonyms | Intent layer |
| Embedding classifier | Surface semantics | Paraphrase | Behavioral probe |
| LLM judge | Surface text | Adversarial prefix | Activation trajectory |
| System prompt | Instructions | Ignore / override | Dual independent gates |
Research and transparency
We publish openly in Sheva Studios Labs:
Who should contact us
- Product teams embedding AI who need licensable safety architecture
- School districts requiring policy-aligned filtering
- Faith organizations serving members via SMS or web
- Parents who want consumer-ready clean AI today without waiting for custom contracts
Consumers: start free. Organizations: Support with your use case.
Summary
Custom AI filtering is not a longer blocklist. It is a different layer of the stack. GHOST is KosherChat's answer — intent reconstruction plus trajectory monitoring — and the foundation for every clean channel we ship.