Six frameworks. One infrastructure.

Live badge status pulled from /api/compliance-badges. Color reflects current observation state, not aspiration.

🛡️
● In Observation
SOC 2 Type II
Day 134 of 180 — Observation window active

Type II proves the design operates correctly over the 180-day observation window, not just that it was designed correctly. Scope: security, availability, confidentiality. Audit start: Jan 1 2026. No shortcuts, no vendor attestations.

CC6.1 · CC6.6 · CC7.2 · CC8.1 implemented
📋
● Built-in, not bolted on
SEC 17a-4(f)
WORM audit log — compliant by architecture

Every compliance record is written to an append-only, HMAC-SHA256 signed table. PL/pgSQL triggers block UPDATE and DELETE at the database level — no application path to mutation exists. 7-year retention enforced. Designed for RIA, broker-dealer, and CCO use.

HMAC-SHA256 · PL/pgSQL WORM · 7yr retention · Neon PostgreSQL
🔒
● Data safeguard compliant
Reg S-P
Data handling and breach notification ready

Customer NPI is never stored in cleartext. AES-256-GCM encryption at rest, TLS 1.3 in transit. Breach notification procedure documented with 30-day reporting target. Data Processing Agreement available on request.

AES-256-GCM · TLS 1.3 · Breach procedure documented · DPA available
🇪🇺
● 113 articles mapped
EU AI Act
Articles 12, 14, 15 — fully addressed

Article 14 human oversight: hallucination-blocked events recorded, budget exhaustion stops intent. Article 15 accuracy: MARCH verification gate + adversarial regeneration. Article 12 logging: WORM append-only audit trail. Enforcement: Aug 2 2026.

Art. 12 logging · Art. 14 human oversight · Art. 15 accuracy
🏥
● Safeguards implemented
HIPAA Technical Safeguards
PHI handling, BAAs, access controls

PHI never processed without explicit BAA. Per-tenant data isolation enforced at query level — no tenant can access another's data. Access audit log captures every PHI read. BAA template available on request. Healthcare state layer: WA MHMD, TX HB 300, 21st Century Cures Act.

Per-tenant isolation · Access audit log · BAA available · No PHI without agreement
⚛️
● NIST FIPS 203/204/205
PQC-Ready
Post-quantum cryptography infrastructure

ML-KEM-1024 (CRYSTALS-Kyber), ML-DSA-87 (CRYSTALS-Dilithium), SLH-DSA (SPHINCS+) deployed per NIST FIPS 203/204/205. Quantum-safe signature infrastructure is live in the audit chain. No dependency on RSA-2048 or ECDSA for audit integrity.

ML-KEM-1024 · ML-DSA-87 · SLH-DSA · NIST FIPS 203/204/205

Live, not screenshots.

Every indicator below pulls live data. If anything is degraded, it shows here.

Encryption Status
● Operational
AES-256-GCM
At rest · TLS 1.3 in transit · CMEK-ready
Audit Log Write Status
● Writing
Loading…
Last 5 entries — live from compliance_audit_log
Tenant Isolation
● Verified
Per-query verified
Row-level isolation · No cross-tenant read path
System Uptime
● Live
Loading…
From synthetic monitor · Updated every 60s
Recent audit log entries Loading…
Loading live entries…

How verification actually works.

No black box. Three steps, cryptographic proof at each boundary.

1
Agent runs
446+ specialist agents bid on every task. The winning agent executes against your regulatory context. Every intermediate state is captured — not just the final output.
2
Triple-Gate verifies
MARCH adversarial checkers challenge the output across factual accuracy, logical consistency, and regulatory compliance. Platt scaling calibrates confidence. Zeroth Law consistency check blocks divergent outputs.
3
Transparency Card signs
The output, all agent IDs, confidence scores, gate verdicts, and citations are bundled into a Transparency Card and HMAC-SHA256 signed. Appended to the immutable audit log. Verification is math, not trust.
📄
Transparency Card 2.0 — anonymized sample
10-field structure · HMAC-SHA256 signed · Appended to audit log
● CHAIN INTACT
thread_id
thd_demo_8f3a9c1e…
winner_agent
sec-exam-specialist-v4 (bid: 0.847)
march_verdict
PASS (9/9 checkers)
gsar_score
0.94 — grounded citations
factual_confidence
0.91 (Platt calibrated)
citations
SEC Release No. 34-38245 · 17 CFR 240.17a-4
consistency_check
PASS — triangulation_hash verified
context_integrity
PASS — 0 context leaks
executed_at
2026-05-14T18:22:07.441Z
hmac_signature
a3f7…9c2e (SHA-256, truncated)
Immutable audit trail — how decisions flow
Intent submitted
Agents compete
Triple-Gate checks
HMAC-SHA256 signed
WORM audit log

PL/pgSQL triggers block UPDATE/DELETE on compliance_audit_log at the database level. Not configuration — code enforced by the database engine itself.

The Five Shields are not aspirational. They are live infrastructure.

Sturna's architecture maps to the three principles of Tria Prima: Sulfur (ethical intent — the CMAG constitutional layer and COMBINED_ALERT override), Mercury (agent mesh — the 201-agent competitive mesh that self-organizes via Shiva Protocol), and Salt (infrastructure — the immutable audit trail, WORM enforcement, and HMAC-signed context manifests). Every compliance claim on this page traces to one of these three layers. The shields are not features. They are the enforcement mechanism that makes the claims credible.

SHIELD 1 — Immutability
Principle: Arrow of Time
What it protects: Tamper-evident audit trail. Every finding is timestamped, signed, and irreversible.
Live system: HMAC-SHA256 AuditLogger → GET /api/trust/audit-stats

Every audit entry carries a deterministic canonical payload signed with HMAC-SHA256. PL/pgSQL triggers block UPDATE and DELETE at the database level — there is no application-layer path to mutation. Verification is server-side: GET /api/trust/audit-stats recomputes HMAC for every entry on read. A mismatch is a production incident.

Evidence: Integrity score 100% | WORM enforcement active | SEC 17a-4(f) compliant | 7-year retention
→ View live audit health
SHIELD 2 — Coherence
Principle: Levin's Bioelectricity
What it protects: Phase-locked agents only. Disconnected agents are quarantined before they hallucinate.
Live system: Context Manifest Protocol → GET /api/live-metrics/context-integrity

Every multi-agent decomposition produces a context manifest. The manifest must be acknowledged (manifest_ack) before the next phase begins — this is a hard block, not a warning. Context leak detection runs in parallel; any cross-boundary data exposure triggers an immediate critical block and audit event.

Evidence: 100% manifest acknowledgment rate | 0 context leaks (7-day window) | context conservation score 100%
→ View live context integrity
SHIELD 3 — Adversarial Challenge
Principle: GAN Architecture
What it protects: Every HIGH+ finding is falsified by a discriminator agent before it reaches the user.
Live system: Transitive Consistency Verifier → GET /api/live-metrics/consistency

The Transitive Consistency Verifier triangulates agent outputs against source state. Triangulation hashes track clause-level divergence; sessions exceeding the divergence threshold (0.35) are flagged for human review. GAN discriminator queued for Week 3 rollout — currently handled by consistency threshold as the active defense.

Evidence: Triangulation hash 8ef70830d298d95580563f0c4c0ff0c8 active | 5 sessions monitored | divergence threshold 0.35
→ View live consistency metrics
SHIELD 4 — Human Sovereignty
Principle: Sacred Geometry
What it protects: HIGH-criticality findings cannot reach the user without human confirmation.
Live system: HITL escalation path via Transparency Card → /demo/live-bidding

HIGH-criticality findings trigger a mandatory human-in-the-loop gate. The Transparency Card shows all 5 confirmation fields: intent classification, agent identities, scoring rationale, execution output, and audit signature. The human review path is not a suggestion — it is a hard stop on the execution pipeline.

Evidence: Live Transparency Card (5 fields) active at /demo/live-bidding | AuditLogger HMAC-signed entries | Human review path enforced at pipeline level
→ See Transparency Card live
SHIELD 5 — Ethical Override
Principle: Tesla + Mercurial Bridge
What it protects: Consumer harm triggers an immediate override auction, regardless of intent classification.
Live system: COMBINED_ALERT P0 — wired, dogfood pending

The COMBINED_ALERT mechanism fires when consumer harm indicators are detected. It immediately pauses the current execution and opens a competing auction on ethical grounds. The override is not configurable — it is constitutional, hard-coded into the CMAG layer. Current status: mechanism is wired and deployed. Internal dogfood testing in progress before general availability.

Evidence: COMBINED_ALERT P0 wired in production | CMAG constitutional hard stop active | Dogfood pending

Security architecture

Current cryptographic and infrastructure configuration. Claims link to implementation.

TLS grade
A+
Render TLS termination · HTTPS enforced · HSTS
Encryption at rest
AES-256-GCM
Per-save 12-byte IV · per-tenant key prefix · versioned keychain
Encryption in transit
TLS 1.2+
Render termination · Neon PostgreSQL SSL required
KMS provider
Render Env Secrets
STURNA_ENCRYPTION_KEY + AUDIT_HMAC_KEY injected at runtime
Post-quantum crypto
ML-KEM-1024
Hybrid: ML-KEM-1024 + AES-256-GCM · 1568-byte KEM ciphertext
Audit log retention
365 days
TTL index enforces cleanup · WORM: no UPDATE/DELETE allowed
Cryptographic controls summary
Control Implementation Status Evidence
State encryption AES-256-GCM, 12-byte random IV per operation Active src/lib/secure-sturna-checkpointer.js
Audit log signing HMAC-SHA256, deterministic canonical payload, verified on every read Active src/lib/audit-logger.js
Post-quantum key exchange ML-KEM-1024 + AES-256-GCM hybrid · 0.98ms avg perf Active src/lib/secure-sturna-checkpointer.js §PQ
Zero-trust agent tokens HS256 + per-tenant HMAC rotation (30d) · 5-min TTL Active services/zero-trust-token-service.js
Audit log immutability PL/pgSQL trigger blocks UPDATE/DELETE on compliance_audit_log Active migrations/046_*.js
Tenant isolation Per-tenant Redis key prefix · cross-tenant throw + audit event Active src/lib/redis-checkpointer.js
Key versioning keyVersion field on every STATE_SAVED event · rolling rotation Active audit log keyVersion field
Live pilot signal
active pilot workspaces with team members reviewing compliance outputs
Live count from pilot_workspace_members table · updates on page load
Loading…

Compliance program status

Real status. Observation period started 2026-05-03. Type II requires 90–180 days of operating evidence. Not complete — target Q4 2026.

SOC 2 Type II — Observation Period 1 of 1
Auditor: Not yet engaged · Target: 2026-11-01 · Compliance platform: Drata (recommended, trial not yet started)
In observation
Day 10 of 180 ~5.6% complete

Observation started 2026-05-03T07:40:00Z. The security architecture (Type I — design) is complete. Type II proves it operates correctly over the observation window. No certification exists yet — this tracker reflects where we actually are.

CC6.1 Logical Access Controls
CC6.6 Encryption in Transit & At Rest
CC7.2 System Monitoring
CC8.1 Change Management
~
CC6.1 gap Access provisioning policy — In progress Q3 2026
~
CC7.4 Incident response policy — In progress Q3 2026
~
CC8.1 gap GitHub branch protection — In progress Q3 2026
~
CC7.2 gap Datadog HMAC failure alerts — In progress Q3 2026

Verify Audit Chain

Paste any Sturna thread ID from a Transparency Card or audit log reference. Server-side HMAC recomputation proves chain integrity without exposing keys. GRC reviewers: download the PDF report to attach to your audit file.

🔐
HMAC Audit-Chain Verifier
Server recomputes HMAC-SHA256 for every entry — no trust required, math verifies
Audit chain entries

Audit log health

Real-time aggregates from the live compliance_audit_log table. Cached 5 minutes. Integrity score below 100% would be a production incident.

Total audit entries
All HMAC-signed log records
Integrity score
% entries with valid HMAC
Oldest entry
Proves continuous operation
Retention policy
WORM — append-only, no UPDATE/DELETE
Last entry written
Most recent log activity
Jurisdictions covered
What does integrity score mean? The server samples the 100 most recent audit entries and recomputes each HMAC-SHA256 signature server-side. Comparing stored vs recomputed — any mismatch indicates post-write tampering. 100% = all entries intact. PL/pgSQL triggers block UPDATE/DELETE at the DB level as defense-in-depth.

Article-by-article attestation

Sturna maps its architecture against EU AI Act Articles 10–15 (High-Risk AI obligations). Full Article-by-Article analysis available at the readiness tool. Below: Sturna's own posture on each key obligation.

Article 10
Data and data governance
Addressed

ComplianceClassifier tags PII/MNPI inline. Hard MNPI reject. Per-tenant data isolation enforced.

Article 11
Technical documentation
In progress — Q3 2026

Architecture doc + data-flow diagram drafted. Formal Annex IV package in progress.

Article 12
Record-keeping
Addressed

Append-only HMAC-signed audit log. WORM enforcement. 365-day retention. SEC 17a-4 aligned.

Article 13
Transparency
Addressed

Triple-Gate verification. GSAR grounding. Citations traced to source on every output.

Article 14
Human oversight
Addressed

Hallucination-blocked events recorded. Budget exhaustion stops intent. Human review path available.

Article 15
Accuracy & robustness
Addressed

MARCH verification gate. §13A GSAR adversarial regeneration. Post-quantum crypto. Self-healing router.

Full Article-by-Article Checklist
113 articles analyzed — Sturna's posture mapped to each obligation
View full checklist →

EU AI Act Articles 10–15 enforcement begins Aug 2, 2026. All operators of High-Risk AI systems in regulated verticals must demonstrate conformity before that date.

Last 90 days

No greenwashing. Dips included, fixes documented.

90-day uptime
99.1%
Excluding scheduled maintenance windows
Incidents (90d)
4
2 P1, 1 P2, 1 P3 — all resolved
MTTR
~47 min
Mean time to recovery across all incidents
Incident log — last 90 days
2026-05-01
OpenAI proxy crash loop — service degraded
LLM proxy entered crash loop after upstream API change. Intent execution degraded (~40% error rate). Fix: Proxy client reset on 429, exponential backoff added, startup self-test added to /health. Root cause: missing retry logic on proxy 429 responses.
P1 · 68 min
2026-04-18
Actuarial vertical 502 under concurrent load
Concurrent pilot load from /dogfood testing caused 502s on actuarial, medical, family-office, and pilots endpoints. Fix: DB connection pool increased, intent execution queue depth reduced. 4 intermittent 502 windows total.
P1 · 31 min
2026-04-02
vs-ChatGPT 429 rate-limit with no fallback
Rate-limit condition with no fallback caused persistent STURNA-DOWN alerts. Status: Known P0 — OPENAI_BASE_URL fix required. Route returns degraded responses while fix is pending.
P0 · Ongoing
2026-03-15
Memory verification degraded — embedding backfill blocked
Memory verification at 0% recall@5. Blocked by OPENAI_BASE_URL configuration gap preventing embedding generation. Status: Degraded — fix in progress with proxy team.
P0 · Ongoing
Feb–Apr 2026
All other systems — operational
No incidents affecting core intent execution, audit logging, compliance classification, or security controls. All 8 launch URLs green.
Operational

Downloadable due-diligence materials

Standard enterprise procurement documents. Current status noted. Email security@polsia.app for the complete ZIP or to request a security questionnaire.

📋
Security questionnaire (CAIQ Lite)
Cloud Security Alliance CAIQ Lite pre-fill — email to request
Available on request
🗂️
Data-flow diagram
Tenant data paths, encryption boundaries, sub-processor touchpoints
In progress — Q3 2026
🔬
Penetration test summary
Scheduled Q3 2026 — no pen-test completed yet
Scheduled — Q3 2026
📜
BAA / DPA template
Business Associate Agreement + Data Processing Agreement templates — enter email to download
⬇ Download (email required)
🛡️
SOC 2 readiness assessment
Full Type II readiness doc including gap analysis and control mapping
Available now
📊
Compliance audit log export
Live evidence export via API — HMAC-signed JSON bundle for Drata
Live API

Live audit evidence bundle: GET /api/compliance/evidence-export/bundle (admin key required). Formatted for Drata Custom Connection daily pull.

Actual list, actual regions

Every service that processes tenant data. No omissions.

Sub-processor Purpose Data region Data type processed DPA
Neon Primary database (PostgreSQL serverless) US-East (AWS us-east-1) Intent data, audit log, tenant config Available
Render Compute / hosting US-West (Oregon) Application runtime, env secrets Available
Upstash Redis Encrypted state storage (per-tenant) US-East-1 AES-256-GCM encrypted agent state Available
OpenAI (via Polsia proxy) LLM inference US (OpenAI data centers) Intent text, agent prompts (no PII policy enforced upstream) Pending
Polsia (proxy) OpenAI proxy routing & cost tracking US LLM request routing (no content storage) Available
Stripe Payment processing US Payment metadata only — no financial instrument data stored Available
Postmark Transactional email US Email addresses, notification content Available

All tenant data remains in US regions. No cross-border data transfer to EU/EEA at this time. EU tenants requiring data residency in-region: contact us — roadmap item Q4 2026.

Ready to talk to a compliance expert?

CCOs and GRC reviewers closing the June 3 RIA pilot round get a dedicated compliance walkthrough. We cover your specific regulatory obligations — not a generic demo.