Trust
Nine US patent applications, 495 claims, 62 independent claims. Methodology validated at scale on public corpora; production accuracy pending Q3-Q4 2026 customer pilots. Honest framing, not marketing inflation.
Patent register
All 9 provisionals filed at USPTO between March and April 2026 under micro-entity status ($585 total). PCT international protection deadlines noted below.
| # | Patent | Application | Filed | Claims | IC | PCT deadline |
|---|---|---|---|---|---|---|
| 1 | G — AI authorship attribution | US 64/009,864 | 2026-03-18 | 32 | 4 | 2027-03-18 |
| 2 | J — Single-pass multi-domain scan | US 64/022,466 | 2026-03-30 | 50 | 5 | 2027-03-30 |
| 3 | K — Real-time visualization (Dracula) | US 64/030,731 | 2026-04-06 | 77 | 8 | 2027-04-06 |
| 4 | H — AI Artifact Audit (HAES) | US 64/030,752 | 2026-04-06 | 56 | 8 | 2027-04-06 |
| 5 | A — Tiered LLM remediation | US 64/030,762 | 2026-04-06 | 58 | 8 | 2027-04-06 |
| 6 | F — PredOpt backlog optimizer | US 64/030,773 | 2026-04-06 | 52 | 7 | 2027-04-06 |
| 7 | B — CI/CD gate w/ policy DSL | US 64/033,022 | 2026-04-08 | 51 | 6 | 2027-04-08 |
| 8 | D — Cross-tool canonical scoring | US 64/033,058 | 2026-04-08 | 71 | 9 | 2027-04-08 |
| 9 | C — Regression detection | US 64/033,063 | 2026-04-08 | 48 | 7 | 2027-04-08 |
| Σ | 9 filed (Patent E defensive-published only) | — | — | 495 | 62 | — |
Validation evidence
Patent G PoC v3.0 (March 2026). Methodology cross-referenced against 6,439,303 code samples spanning 64 distinct AI models (Llama, Qwen, DeepSeek, Gemini, Phi, GPT-4, Claude, IBM Granite, StarCoder, …) across 13 programming languages and 9 published datasets (AICD-Bench T1/T2/T3, DroidCollection, CodeMirage, GPTSniffer, DevGPT, SWE-bench, Stack Overflow Survey). All 7/7 Patent G claims CONFIRMED with STRONG-to-VERY-STRONG evidence at scale.
Three confirmed claims
- Multi-signal detection necessary. Single-feature classifiers (e.g. code length alone) are insufficient — Cliff's δ 0.07-0.28 (negligible-to-small effect size across 6M+ samples). Patent G's multi-signal architecture is necessary, not optional.
- Per-model fingerprinting feasible. Distinct line/character/stdev signatures per LLM (e.g. o3-mini avg 169.9 lines vs llama3.3 avg 83.2 — a 2× difference). Enables per-tool quality scoring on customer codebases.
- AI adoption is exponential. DevGPT corpus shows 145% growth in 77 days (Spearman rs = 0.98, p < 0.001); power users dominate (Gini 0.68 — top 7% of authors produce 41% of AI commits).
Honest framing
The PoC validates the methodology at scale on public corpora. End-to-end ariada.ai classifier accuracy on production customer code is pending Q3-Q4 2026 customer pilots (PoC v4.0 in scoping). We do not claim "validated on N customer sites today."
Patent F + Patent J research record
- Patent F (PredOpt). arXiv methodology paper
build prompt queued at
product/plans/2026-04-27-patent-f-arxiv-paper-build-prompt.md; target Q2-Q3 2026 (post-PCT). - Patent J differentiation analysis (CANTOR, 2026-04-17). Cross-referenced Siteimprove estate (US 11,995,091, US 11,409,832, App-Pub 2019/0073365). Siteimprove is multi-modal on one regulatory domain; Patent J is multi-regulatory-domain in single pass — architecturally distinct. Q2-2026 (RAMANUJAN) confirms the Feb-2026 Siteimprove release does not collapse the distinction.
Sources: research/poc/patent-g/PATENT_G_POC_V3_REPORT.md,
umbrella PRD §13, research/output/competitive-Q2-2026.md.
Security & compliance footprint
- SOC 2 Type II — in flight (Phase 2 target Q4 2026)
- ISO 27001 — planned (Phase 3)
- GDPR Article 28 DPA — available
- EU residency by default — Hetzner FSN1 + HEL1; Phase 2 multi-region
- EU AI Act Article 50 — HAES emits the Art. 50 evidence package per artifact
- EAA / WCAG 2.2 AA — this site is itself axe-clean (cobbler's-shoes rule)
- Append-only audit ledger — tamper-evident hash chain, daily Merkle anchor
Risks (honest framing)
Top risks from the umbrella PRD §14, abridged. Full risk register lives in the internal product spec.
| # | Risk | Impact | Mitigation |
|---|---|---|---|
| 1 | Enterprise sales cycle (180-360d) > runway | Critical | Land-and-expand from blamer/clamper Pro; partner agency program; founder-led outbound to 50 named accounts |
| 2 | Incumbent (Deque, Siteimprove) adds 1-2 ariada.ai capabilities | High | 9-patent stack means single-capability copy non-fatal; arxiv-monitor weekly + competitor quarterly |
| 4 | Pilot accuracy below 80% canonical scoring agreement | High | Phase 1 limited to high-confidence rule mappings; iterative D-weight tuning; manual override |
| 7 | LLM cost overrun on Tier-2/3 remediation | Medium | Tier cascade + cache + similarity reuse; per-org budget; auto-downgrade; dual-vendor |
| 8 | Founder bandwidth — 1-2 ppl can't ship 6 capabilities + sales | Critical | Phase 1 narrowed to 6 capabilities, not all 9; build-prompts dispatched in fresh sessions |
| 10 | Patent infringement claim (AudioEye 10-K IP-litigation posture) | Critical | 9 ADOPTA provisionals + CANTOR / Evinced / AudioEye prior-art scans + threat-triage; PCT on time; design-around reserve |
Cobbler's shoes
ariada.ai detects accessibility violations. ariada.ai itself must be impeccable. Every page on this site is checked with axe-core (zero violations target) and Lighthouse (a11y score 100). If you find an accessibility issue, email a11y@ariada.ai and we will fix it.