Oracle & AI Validation

The Trayon Oracle is a decentralized network of AI agents that capture data from primary sources, process it through machine learning models, and submit it for validator consensus before it is committed on-chain. The guiding principle is Zero Trust Data: no single source is ever trusted by default.

Data pipeline

Data Sources (official APIs, transparency portals, market feeds)
        │
        ▼
AI Data Ingestion Layer
   ├─ Web scrapers / API connectors
   ├─ Schema & checksum validation
   └─ Statistical anomaly detection
        │
        ▼
AI Processing & Feature Engineering
   ├─ NLP (text and context analysis)
   ├─ Time series analysis
   ├─ Fraud detection models
   └─ Predictive / forecasting models
        │
        ▼
Consensus & Validation Layer
   ├─ Validator nodes (2/3+1 honest majority)
   ├─ BLS threshold signing
   └─ Reputation scoring & slashing detection
        │
        ▼
Blockchain Settlement (Merkle root commitment)

AI ensemble agreement

Each data point is scored by an ensemble of independent models — typically combining an Isolation Forest and Gradient Boosting classifier for fraud detection, alongside an LSTM/ARIMA model for time series forecasting. A value is only proposed to the network once the ensemble reaches statistical agreement, reducing the risk that a single flawed model introduces bad data.

Why two layers of verification

AI ensemble agreement alone is not enough — models can share blind spots. Decentralized validator consensus alone is also not enough — validators can only check what they receive. Combining both layers means:

• The AI layer catches statistical anomalies and known fraud patterns before data ever reaches consensus.
• The validator layer independently re-executes and cross-checks proposed data against a 2/3+1 quorum requirement.

This is what allows Trayon to flag manipulation in real time, rather than discovering it in a retrospective audit months later.

On-chain commitment

Once validated, data is committed as a Merkle root with an aggregated BLS signature via the TrayonOracle contract. See Smart Contracts for the interface reference.

The AI model ensemble, in detail

Different data domains call for different model types. Trayon runs a small set of specialized models per data category rather than one general-purpose model:

Fraud detection

Input:   Financial data, transaction history
Output:  Fraud probability score (0–100)
Model:   Isolation Forest + Gradient Boosting
Accuracy: 94–96%

Example:
  R$1M transaction from a historically small account
  → Fraud score: 87%
  → Requires manual sign-off from 2 validators
  → If approved, committed on-chain; if not, discarded

NLP & sentiment analysis

Input:   News text, judicial decisions, official statements
Output:  Sentiment (bullish/bearish) + extracted entities
Model:   Transformer-based (BERT/GPT-class)

Example:
  Headline: "Government announces new import tariff"
  → Sentiment: bearish (82% confidence)
  → Entities: [Government, Tariff, Imports]
  → Downstream signal: -3–5% expected on exporter equities

Time-series forecasting

Input:   30–90 days of historical price/value data
Output:  24-hour forecast with confidence interval
Model:   LSTM + Prophet
Accuracy: 72–78%

Example:
  USD/BRL history: [5.20, 5.18, 5.22, 5.19, ...]
  → 24h forecast: 5.21 (range 5.19–5.23)
  → Confidence: 74%
  → Validators vote on whether the forecast is admissible

Market manipulation detection

Input:   Volume, price, and wallet-concentration data
Output:  Manipulation score (0–100)
Model:   Anomaly detection + graph analysis
Detects: Pump-and-dump patterns, wash trading, whale coordination

Case study: government procurement audit

A concrete walkthrough of the full pipeline, end to end, for a government procurement dataset:

1. Ingestion
   Trayon pulls the public procurement portal feed (bid amounts,
   winning vendor, contract value) alongside the Treasury's
   published budget execution data for the same period.

2. Cross-source validation
   The two independent sources are diffed; if the procurement
   amount and the budget execution entry disagree by more than a
   configured tolerance, the entry is flagged for anomaly review.

3. AI ensemble scoring
   The fraud-detection model scores the flagged entry (e.g. a
   contract awarded 40% above the median for comparable bids
   scores 91/100 for fraud risk).

4. Validator consensus
   High-risk entries require an explicit 2/3+1 validator quorum
   vote rather than default auto-approval; validators can pull the
   underlying source documents via the same ingestion connectors.

5. On-chain commitment
   Approved entries are committed as a Merkle root via
   TrayonOracle.commitData(), with the AI risk score and validator
   vote tally stored alongside the data for future audits.

6. Query & billing
   Government auditors or journalists query the committed record
   through the data marketplace (50,000 TRAY per full audit — see
   TRAY Tokenomics), with the query fee split 70/20/10 across
   validators, burn, and treasury.

This is what "detect manipulation in real time" means concretely: the discrepancy is flagged during ingestion — before settlement — rather than surfacing months later in a retrospective audit.

Sources by sector

Each vertical draws from multiple independent sources so no single feed can unilaterally determine a committed value: