System Architecture

GAMP Technical Reference

The cryptographic infrastructure, AI systems, provisional patents, and API architecture that make GAMP the SSL for AI.

Six Interlocking Trust Layers

GAMP is not a single system — it is six mechanisms that together constitute a trust protocol for AI. Each layer addresses a different failure mode in AI deployment trust.

01
Identity

Cryptographic Passport

Every registered AI agent receives a unique Ed25519 keypair generated fresh at registration. GAMP signs the passport payload using RFC 8785 canonical JSON serialisation — deterministic byte representation regardless of key ordering. The signature covers all passport fields. Each update stores the previous passport hash, creating a tamper-evident chain. Any field change post-signing invalidates verification.

Ed25519 RFC 8785 JCS Hash chaining Per-agent keypairs Canonical JSON
02
Evidence

Verifiable Evidence Packets (VEP)

A VEP is a structured, signed, tamper-evident object that allows third parties to independently verify a claim about a model. The evaluator signs all fields except id, verification_status, and verification_notes using their own Ed25519 private key. GAMP re-runs a five-check verification suite on submission. Verified VEPs carry 1.5× the trust score penalty of unverified assertions — proven evidence is not equal to bare claims.

Evaluator Ed25519 signing 5-check verification 1.5× verified weight 0.5× unverified weight Replay prevention prev_vep_hash chaining
03
Monitoring

Semantic + Statistical Drift Detection

The composite drift score is the primary technical differentiator. A semantic layer (Claude AI, 60% weight) analyses behavioural dimensions — tone, reasoning style, safety profile, refusal patterns, domain focus. A statistical layer (deterministic, 40% weight) computes cosine n-gram similarity, refusal rate shift, Shannon entropy change, and structural profiling. The final score is fully reproducible — a SHA-256 reproducibility hash covers all inputs and outputs.

Drift = 0.6 × semantic_score // Claude AI analysis
+ 0.4 × statistical_score // deterministic fingerprint
// statistical_score =
0.35 × cosine_drift + 0.30 × refusal_drift
+ 0.20 × entropy_drift + 0.15 × structural_drift
Reproducibility hash Reason codes Benign / concerning / critical Cooldown system Versioned baselines
04
Intelligence

AI Audit Document Analysis

Uploaded audit documents are analysed by AI before any human reviewer sees them. Each document receives a quality score (0–100), completeness rating, key claims extraction, red flag identification, and placeholder detection. Cross-document analysis checks consistency across training data descriptions, architecture specs, safety evaluations, and compliance documentation. Results are stored on the AuditRecord and inform the tier recommendation.

Per-document quality scoring Cross-document consistency Gap analysis Placeholder detection Tier recommendation
05
Compliance

AI Industry Classification

An independent AI assessment analyses all registration fields — model name, description, intended use, architecture, training data, limitations — and produces a probability-weighted industry assessment. This detects mismatches between declared and actual use case: a financial advice model declaring "general productivity" to avoid the higher audit requirements of financial services. Triggered regulatory frameworks are identified automatically.

MiFID II FDA SaMD MDR (EU) FAIS Act (SA) EU AI Act HPCSA
06
Propagation

Live Badge + Webhook Network

One script tag embedded on any site makes a live call to GAMP on every page load. When a model is suspended, drift-flagged, or tier-changed, every embedded badge reflects the change within one page load — everywhere simultaneously. Outbound webhooks deliver HMAC-SHA256 signed event payloads to registered endpoints. Auto-disable after 5 consecutive failures prevents dead endpoint flooding.

Real-time propagation HMAC-SHA256 signing 9 event types Browser extension Python SDK JavaScript SDK
Provisional Patents Filed

Four provisional patent applications have been filed establishing priority date on the core innovations of the GAMP system. These represent the novel technical contributions that distinguish GAMP from all existing AI governance systems.

Patent 1 — Most Important

Evidence-Weighted Trust Scoring System

A method for computing a dynamic trust score for AI models wherein claims backed by cryptographic evidence carry a higher penalty multiplier than unverified assertions. The 1.5× verified / 0.5× unverified differential weighting tied to Ed25519 signature verification status is novel and not present in any existing AI governance or registry system.

  • Differential weighting of verified vs unverified AI model claims in trust score computation
  • Integration of Ed25519 cryptographic verification with dynamic trust scoring
  • Automatic trust score update upon VEP verification status change
  • Trust score propagation to embedded visual indicators in real time
Patent 2

Real-Time Distributed Trust Propagation

A method for simultaneously propagating AI model trust status changes to all distributed embedded instances within one page load cycle — without requiring the embedding site to update their code. SSL certificate revocation does not propagate to all websites simultaneously. GAMP's live badge mechanism does. This distinction is novel.

  • Single administrative action triggering simultaneous badge state change across all deployments
  • Live verification call on each page load against the central registry
  • Cryptographic verification status affecting distributed visual trust indicators
  • Real-time revocation propagation without embedding site code changes
Patent 3

Verifiable Evidence Packet System and Validation

The VEP architecture as a structured, signed, tamper-evident evidence object for AI model claims. The combination of RFC 8785 canonical JSON serialisation, independent evaluator Ed25519 signatures, prev_vep_hash chaining for evidence versioning, and integration with differential trust scoring is novel as a combined system applied to AI model certification.

  • RFC 8785 canonicalisation for deterministic, reproducible payload hashing
  • Hash chaining for tamper-evident evidence versioning (prev_vep_hash)
  • Five-check automated verification: fields, hash integrity, signature, proof, timestamp
  • Unique canonical hash constraint for replay attack prevention
Patent 4 — Most Novel

Semantic Behavioural Deviation Detection Tied to Trust Scoring

A composite AI and statistical drift detection system where detected behavioural deviation automatically updates a trust score and propagates to embedded status indicators. The 60/40 composite of semantic AI analysis and deterministic statistical fingerprinting producing a reproducible, auditable, reason-coded drift score tied to automatic enforcement is novel.

  • Composite semantic AI + deterministic statistical drift score for deployed AI models
  • SHA-256 reproducibility hash enabling independent audit verification of drift scores
  • Structured reason codes (SAFETY_PROFILE_SHIFT, REFUSAL_RATE_DROP) for regulatory explainability
  • Cooldown system with multi-check confirmation preventing false positive penalties
  • Versioned baselines distinguishing legitimate model updates from unauthorised changes
What the System Provides
🔐

Per-Agent Ed25519 Keypairs

Every agent gets a unique keypair at registration. The private key is returned once and never stored by GAMP. Outputs signed with the key can be verified against the registered public key — proving they came from the genuine registered agent.

⚖️

Live Trust Score (0–100)

Dynamic score computed from tier base, co-signature bonuses, recent verification bonus, VEP claim penalties (weighted by verification status), drift reports, and status penalties. Suspended models score 0 regardless of tier.

🔗

Certifier Co-Signing Network

Independent certifiers register their own Ed25519 keypairs. Two co-signatures auto-upgrade a model to Audited tier. The chain of trust — who certified this model, who verified the certifier — is publicly readable and cryptographically re-verified on every call.

📡

Public Transparency Log

All registrations and tier changes are publicly readable in append-only form. Modelled on certificate transparency logs in TLS. Anyone can audit the full history of any model without an account. Machine-readable revocation feed for security scanners.

🏭

Industry Minimum Tier Enforcement

15 industry codes with enforced minimum tier requirements. Healthcare requires Audited. Defence requires Accredited. Finance requires Verified. Non-compliance shown immediately on declaration. The AI classification layer detects undeclared regulated use.

🔑

Master Key Rotation

Versioned key registry with configurable grace period (default 90 days). Rotate generates a new keypair, archives the old one. Existing passports remain verifiable against the old key during the grace window. Bulk re-signing via resign-all endpoint.

🏢

Organisation Accounts

Multi-member organisations with four roles: owner, admin, member, auditor. Invite system with email confirmation. Organisation-level agent registration. Agents can be associated with an organisation rather than an individual account.

🔄

GAMP-Signature HTTP Header

Standard: v=1; model=id; ts=unix; sig=base64; hash=sha256. Covers body hash integrity, timestamp freshness (5-minute window), and Ed25519 signature. API consumers can verify agent outputs inline without a separate verification call.

☸️

Kubernetes Helm Chart

Full Helm chart with HSM/KMS provider selector: local PEM, AWS KMS (IRSA), HashiCorp Vault, GCP Cloud KMS. Pod anti-affinity, HPA autoscaling, cert-manager TLS, rate limiting annotations. Enterprise-ready for on-premise deployment.

107 Production Endpoints

Full interactive API documentation available at /docs (requires backend running). All endpoints are documented with request/response schemas, authentication requirements, and example payloads.

107
Total routes
148
Tests passing
38
Python modules
22
API modules
ModuleMethodEndpointDescription
AuthPOST/auth/registerRegister account — sends welcome email
AuthPOST/auth/loginLogin — returns JWT token
AuthPOST/auth/forgot-passwordSend password reset email
ModelsPOST/models/registerRegister AI agent — issues Ed25519 keypair and passport
ModelsGET/models/{id}/verifyVerify passport signature — public, no auth
ModelsGET/models/{id}/badge.svgLive SVG badge — reflects real-time trust status
VEPPOST/vep/submitSubmit Verifiable Evidence Packet with evaluator signature
VEPPOST/vep/{id}/verifyAdmin re-verification or manual override
SemanticPOST/semantic/models/{id}/set-baselineBuild semantic behavioural fingerprint
SemanticPOST/semantic/models/{id}/checkRun composite drift detection — semantic + statistical
SemanticPOST/semantic/models/{id}/verify-publicPublic trustless verification — no account required
AI AuditPOST/ai-audit/models/{id}/run-full-analysisFull AI analysis of all uploaded documents
AI AuditPOST/ai-audit/models/{id}/classify-industryAI industry classification and mismatch detection
TransparencyGET/transparency/feedAppend-only public audit feed — no auth
TransparencyGET/transparency/revokedMachine-readable revocation feed for scanners
WebhooksPOST/webhooksRegister HMAC-signed webhook endpoint
WebhooksPOST/webhooks/{id}/testSend test ping to verify endpoint reachability
RegistryGET/registry/searchSearch registered agents — public, filterable
KeysPOST/keys/rotateRotate master keypair with configurable grace period
CertifiersPOST/certifiers/cosignCo-sign passport — triggers tier upgrade at 2 signatures

Showing 20 of 107 endpoints. Full documentation at /docs while backend is running, or at gampid.co.za/docs on production.