PATTERNWALL

Constitutional governance middleware that intercepts adversarial prompts before the model sees them. Deterministic. Auditable. Model-independent.

Read the Paper Try Live Demo

LAYER −1

PatternWall enforces governance before inference — operating at Layer −1 between user input and the model . It blocks malicious input before compute is spent, applies identical logic across providers, and produces an auditable trail for every decision.

Input

User Prompt

Raw input, any source

Layer −1

PatternWall

Content · Context · Behavior

Decision

Tiered Response

Verdict + audit hash

Output

AI Model

Any provider, same rules

Deterministic

Identical input produces identical scores, identical evaluations, identical responses. No probabilistic components.

Auditable

Every decision emits an audit log with explicit match attribution. Reproducible compliance evidence.

Model-Independent

Tested across 5 distinct architectures — frontier and open-source. Same rules, same outcomes.

BENCHMARK DATA

28 multi-turn adversarial campaigns across medical, legal, financial, psychological, and infrastructure domains. 280 test runs. 1,180 total turns. 5 frontier models.

100%

Per-Attack Detection

28 of 28 campaigns intercepted before downstream model receives the disallowed turn

96.4%

Hard Block Rate

27 of 28 hard blocked. 1 received graduated enforcement

55.1%

Per-Turn Interception

65 of 118 turns caught. Benign setup turns correctly allowed

False Positives

0 false positives across 265 benign control turns (53 per model × 5)

Models Tested

GPT-5.2, Claude Opus 4.5, Grok, Mistral Large 3, Qwen 3 VL

280

Test Runs

140 raw + 140 gated, 1,180 total turns

Raw vs. Gated

Model-native defense vs. infrastructure governance. Average native resistance: 4.3%.

Model	Without PatternWall	With PatternWall
GPT-5.2	14.3% native block	55.1% catch rate
Claude Opus 4.5	0% native block	55.1% catch rate
Grok	0% native block	55.1% catch rate
Mistral Large 3	3.6% native block	55.1% catch rate
Qwen 3 VL	3.6% native block	55.1% catch rate

Development Trajectory

5.4× improvement through systematic gap analysis. No version accepted without full-corpus validation.

Version	Hard Block	All Catches
v3.2	10.2%	10.2%
v3.2.1	36.4%	36.4%
v3.3	22.9%	23.7%
v3.4.2	28.0%	39.0%
v4.2	43.2%	54.2%
v4.3	52.5%	55.1%

LIVE DEMO

Paste any prompt below. PatternWall screens it through multiple detection layers with stateful threat tracking — before the model ever sees it.

PatternWall v4.0 —