How ariada.ai works

One scan emits evidence for every regulator. One score normalizes across vendors. One CI gate blocks regressions before they merge. One audit trail satisfies EU AI Act Article 50. One contract, one platform, nine capabilities — each mapped to a US patent application.

The integrated pipeline

Every ScanEvent emitted by Patent J flows downstream to G (who wrote it), D (canonical score), and H (AI Act compliance). C trends regressions, B blocks PRs, A suggests source patches, F schedules the backlog under sprint capacity, and K turns the numeric report into stakeholder-comprehensible narrative.

J (multi-domain scan) ─▶ G (attribution) ─▶ K (visualization)
        │                       │
        ▼                       ▼
   D (canonical          H (AIAS / Art. 50
    scoring + cert)        AI artifact registry)
        │                       │
        ▼                       ▼
   C (regression          B (CI/CD gate
    trend + cluster)        policy DSL)
        │                       │
        └──────────┬────────────┘
                   ▼
        A (auto-remediation, tiered LLM cascade)
                   │
                   ▼
        F (PredOpt backlog, MIP + ML warm-start)

Reference: umbrella PRD §4.2; scanner architecture v1.3 at product/microservices/ARIADA_SCANNER_ARCHITECTURE_v1.md.

Nine capabilities, one stack

J — Single-pass multi-domain scan (US 64/022,466)

Single scan emits conformance evidence across multiple regulatory domains. Canonical engine: Rust scanner on Hetzner. Output: locked ScanEvent stream over NATS → Node SSE → web/CLI consumers.

Domains: WCAG 2.1 AA, WCAG 2.2 AA (Phase 1.5), EN 301 549, EAA, Section 508, ADA II, DOS-lagen.
Rules: axe-core 4.11+ + custom Rust rule packs per regulation.
Targets: <30s/page; <10min for a 500-page property.

A — Tiered LLM remediation (US 64/030,762)

Source-code patches, not runtime overlay. Cheap-to-expensive cascade with cache + similarity reuse. Framework-aware diffs (React / Vue / Angular / Svelte / HTML).

Tier	Engine	Coverage	Unit cost
0	Deterministic rules (alt-text-from-filename, ARIA defaults)	~30%	$0
1	Cerebras / Gemini Flash (context-light)	~50%	~$0.002/fix
2	Claude Sonnet (form labels with copy, semantic landmarks)	~15%	~$0.02/fix
3	Claude Opus (modal focus traps, custom widget ARIA)	<5%	~$0.20/fix

B — CI/CD gate with policy DSL (US 64/033,022)

GitHub App webhook on PR open/sync (Phase 1); GitLab CI Phase 1.5; CircleCI / Jenkins via CLI. YAML policy DSL with differential thresholds for AI-authored code (Patent B Cl. 4 + Patent G integration).

version: 1
gate:
  budget:
    critical: 0
    serious: 5
    moderate: 20
  ai_authored_diff:                 # Patent B + Patent G
    critical: 0
    serious: 2
  domains: [wcag_2_1_aa, eaa_chap_iii]

D — Cross-tool canonical scoring (US 64/033,058)

Inputs: axe (canonical), Lighthouse, Pa11y, WAVE, Siteimprove API import, Deque DevTools manual import. Per-rule severity normalized to a unified 0-100 score plus WCAG-SC sub-scores. Signed cert per-domain (JSON+PDF, Ed25519, revocable).

Distinguishes from Siteimprove US 11,995,091 (single-tool SEO+a11y+QA): Patent D normalizes across N≥2 tools. See the Trust page for the CANTOR differentiation analysis (2026-04-17).

H — AI Artifact Audit (HAES) (US 64/030,752)

Append-only event ledger: every scan, violation, fix, override, cert issued/revoked. AI artifact registry with provenance (tool, time, approver). EU AI Act Art. 50 transparency record per artifact (model, training-data class, output marker). 7-year retention; tamper-evident hash chain; daily Merkle anchor.

C — Regression detection (US 64/033,063)

Cross-deploy diff engine; clusters root causes; per-component trend; sprint-level regression summary. Roadmapped to Phase 2 (Q4 2026).

G — AI authorship attribution (US 64/009,864)

Multi-signal classifier identifies Copilot, Cursor, Claude Code, Windsurf, Devin, CodeWhisperer, Tabnine. Methodology validated against 6.4M code samples / 64 AI models / 13 programming languages / 9 datasets (PoC v3.0). Production fingerprint engine in Phase 3 (Q1 2027).

F — PredOpt backlog optimizer (US 64/030,773)

Mixed-integer programming + ML warm-start over the violation backlog under sprint capacity, severity, dependency, and budget constraints. arXiv methodology paper queued. Phase 3 SaaS endpoint optional (only if OR/ML co-founder lands).

K — Dracula visualization (US 64/030,731)

Character-themed scanner visualization for stakeholder reports. Same engine as the draculascan.com viral demo. Embedded in dashboard from Phase 2.

MVP Phase 1 (Q3 2026)

Phase 1 launches with six of nine capabilities. C, K, AIAS expansion, and GitLab gate roadmapped to Phase 2 (Q4 2026). G full + F + self-hosted in Phase 3 (Q1 2027).

#	Component	Patent	Marketplace counterpart
1	Cross-domain scanner	J	(no standalone)
2	Tiered LLM remediation	A	reverter.ai (subset, IDE/MCP)
3	CI/CD gate w/ policy DSL	B	clamper.ai (subset, GH+Vercel)
4	Cross-tool canonical scoring	D	(no standalone)
5	Executive dashboard	—	(subset reports in standalones)
6	Audit trail (HAES = H + Art. 50)	H	(no standalone)

See pricing → Trust & patents