MODALITIES
3
modalities
Text, image, and video each have their own detector. Audio is analysed as part of a video file, not as a standalone upload.
NEURAL INTERROGATION ENGINE ONLINE
ULTIMATE DETECTION PROTOCOL
What the engine actually does
Three modalities, one verdict
Every scan runs through the same forensic pipeline: ingest the file, extract modality-specific signals, fuse them with a quality bound, then return a probability with the evidence that produced it.
LANGUAGES
EN
/
AR
English and Arabic prose are scored with separate signal sets, and mixed Arabic/English input is detected and weighted accordingly.
OUTCOMES
3
verdicts
The engine is allowed to decline. When the evidence is weak or the detectors disagree, it returns Inconclusive instead of a confident guess.
AI generated
Human
Inconclusive
Inside the pipeline
How a verdict is reached
No single model decides the answer. Independent detectors report what they see, and the fusion step refuses to turn weak agreement into certainty.
-
01
Ingest and normalise
Text, documents, images, and video are decoded into the feature form the matching detector expects. Unsupported or empty input is rejected before any analysis runs.
-
02
Run the specialist detectors
Text reads stylometry, burstiness, phrase patterns, and humanizer traces. Images read pixel, compression, and metadata signals. Video adds frame-to-frame motion and codec consistency.
-
03
Fuse with a quality bound
Results are combined by evidence weight, and confidence is capped by the average quality of the inputs. Disagreement lowers confidence rather than hiding it.
-
04
Return the evidence, not just the label
You get a probability, a confidence level, the reasons behind the call, and any conditions that limited it. If the engine cannot justify a verdict, it says so.
"classification": "Human",
"ai_probability": 0.16,
"confidence": 0.63,
"verdict_status": "decisive",
"evidence_strength": "MODERATE",
"reasons": [
"Irregular sentence lengths",
"Domain-specific vocabulary"
],
"conditions": [
"Statistical signals only"
]A representative payload. The probabilities are the model's, the reasons and conditions come from the fusion step, and nothing is filled in after the fact.
System capabilities
What runs under the hood
Four things the engine genuinely does today — no speculative roadmap, no model counts we cannot back up.
Structure-aware parsing
Sentence rhythm, lexical patterns, and structural signals are read together rather than scored in isolation. A single unusual word cannot move the verdict on its own.
Local processing
Inference runs on our own hardware. Nothing about your file is sent to a third-party model provider.
Private by default
History is tied to your account and readable only by you through row-level security. We do not sell personal data.
Multi-modal forensics
Text, image, and video each run their own specialist detectors, and the fusion step refuses to turn weak agreement into a confident call. When the signals disagree, you get a low confidence score instead of a guess.
PHASE 01: INGESTION
Target Acquisition
Input target data vector. Accepts raw text, documents, images, and video within the limits of your plan.
01
PHASE 02: ANALYSIS
Signal Analysis
Combining stylometric, structural, and forensic signals. Results include confidence, warnings, and an evidence summary for review.
02
PHASE 03: EXECUTION
Terminal Verdict
Generate a probability, confidence, and evidence summary so you can review the reasoning before making a decision.
03