TruthNexus
EU AI Act Article 13 — enforcement begins August 2026CMS Prior Auth Rule — AI rationale traceability required January 2027Check your compliance posture →
Medical · Legal · Government · Compliance

AI decisions that survive a regulatory audit

Every Skippy output carries source lineage, calibrated confidence, and a cryptographic audit trail — traceable from the response all the way back to the primary source. Built for the EU AI Act, FDA, CMS, OMB M-24-10, and court-of-record evidence standards.

Nature Communications 2025
50–90% of medical AI claims unsupported by their cited sources
ECE 0.07
Calibration error
5 / 5
PCCP gates
143
Medical sources
HIPAAGDPRSOC 2FedRAMP
75%
Of PA denials overturned on appeal — each one a defensibility failure
$100B+
Annual healthcare fraud — pre-payment screening is the only structural fix
ECE < 0.10
Calibrated confidence on every finding — measured, not claimed
3% revenue
EU AI Act penalty for non-compliant AI — August 2026 enforcement
The Structural Problem

What LLMs can't provide — and why bigger models won't fix it

These are not accuracy bugs. They are properties of how transformer training works. Regulators don't ask “how accurate is this?” — they ask “show me your evidence chain.”

No evidence chain

When an LLM produces a regulated claim — clinical, legal, or financial — there is no traceable path to a source. The claim came from blended training data, millions of documents indistinguishably mixed. There is no source document ID to show a regulator, judge, or auditor.

No calibrated confidence

LLM confidence is expressed through hedging language, not evidence. A claim backed by 1,000 randomized trials and a claim backed by one case report — or a controlling Supreme Court holding versus a single district-court dictum — produce equally confident-sounding outputs. There is no way to tell them apart.

No knowledge boundary

LLMs have no structural "I don't know" state. They generate plausible text past the edge of their knowledge, silently. In any regulated context — clinical, legal, financial, federal — confident hallucination is more dangerous than an honest refusal.

The Answer

What a Skippy response provides

Every output is gated by a verifier before delivery. The citation target is not a URL or a document chunk — it is a specific, versioned finding that can be walked all the way back to the evidence it rests on.

Verifier-gated output
Ungrounded output is rejected before delivery — not flagged, not softened. A hard contract, not a filter.
Calibrated confidence
ECE-validated against held-out data. 0.85 confidence means ~85% accurate — a measurement, not a posture.
Full source lineage
Source document ID, version date, evidence count, convergence pattern, and any contradictions — on every finding.
First-class knowledge boundary
NOT_COVERED is a real API response. Skippy is honest about what it doesn't know. No fabrication past the evidence boundary.
POST /v1/ground/verify · 200 OK · 11ms
// Request
"claim": "Warfarin + amiodarone co-administration"
// Response
"verdict": "CONTRAINDICATED"
"confidence": 0.97
"evidence_pattern": "convergent"
"sources": 14
"contradictions": 0
"lineage": [
"PMID:28823332",
"DrugBank:DB00682",
"FDA label · 2024-03-15"
]
"audit_id": "sk_a3f92b1c4d7e"
"signed": true
Verifier: PASSECE: 0.07Signed: Ed25519Retraction check: clean
Who It's For

Built for the person who gets audited

Compliance & Legal

Audit-ready output, every time

  • Every claim carries a source ID and evidence count — no assembling evidence after the fact
  • EU AI Act Article 13 and CMS-0057-F documentation generated automatically
  • Cryptographic audit trail satisfies court-of-record and regulatory review standards
See compliance docs →
Clinical Teams

Evidence you can show the clinician

  • Recommendations link to primary sources — PMID, guideline version, and date
  • Calibrated confidence distinguishes a strong RCT from a case report
  • NOT_COVERED returns honestly when evidence is insufficient — no fabrication
See Medical products →
Engineering & API

Drop-in verifier, not a model replacement

  • REST API with structured JSON — verdict, confidence, source lineage, audit ID
  • Open-source skippy-verify runs offline to check any response cryptographically
  • P99 latency < 80ms; no PHI retained beyond the API call
View API docs →
Hard Limits

What Skippy will not do

Trust is built as much by what a system refuses as by what it claims. These are not configuration options — they are structural constraints baked into the verifier contract.

No output without evidence

When evidence is insufficient, Skippy returns NOT_COVERED — a real API response, not a hedged paragraph. It does not generate plausible-sounding text past the evidence boundary.

No autonomous decisions

Skippy verifies and grounds claims — clinical, legal, financial. It does not create de novo recommendations or final decisions. Professional judgment stays with the clinician, lawyer, or regulator — always.

No bypass on verifier failure

Output that fails the verifier gate is blocked at the API level — not softened, not flagged for review. The contract is hard. There is no override path.

No request-data retention

Request context is not stored beyond the API call. PHI, MNPI, privileged communications, and classified content stay in the request — they do not enter Skippy's logs, training pipeline, or storage layer.

Integration

From API key to audit-ready in a week

Skippy is a REST API — not a platform migration. You bring the query; Skippy returns the verified finding, confidence score, source lineage, and signed audit trail. No model fine-tuning, no RAG pipeline to build, no evidence base to maintain.

01
Get your API key
Sign up and receive credentials. Sandbox environment active immediately — no approval gate.
02
Send your first query
POST your claim to /v1/ground/verify. Receive verdict, confidence, and source lineage in the response body.
03
Verify the audit trail
Run skippy-verify locally (pip install skippy-verify) to cryptographically confirm any response. Open source, Apache 2.0.
04
Deploy to production
Add the audit_id to your records. BAA executed, evidence chain documented, compliance posture upgraded.
Regulatory & Compliance

Active deadlines across regulated domains.
The nearest is 3 months away.

EU AI Act Article 13 enforcement begins August 2026. These are not future risk — they are active enforcement timelines with penalties up to 3% of global annual revenue. LLMs cannot structurally satisfy any of these requirements.

HIPAA ReadySOC 2 Type II · Q4 2026GDPR CompliantBAA Available7-Year Audit RetentionFDA-Traceable Citations
Security & compliance documentation →

EU AI Act — Article 13

Enforcing
August 2026

Technical documentation of knowledge base, how outputs are derived, and audit records for every decision. Non-compliance: up to 3% of global annual revenue.

CMS Prior Authorization Rule

Active
January 2027

AI-driven PA rationale must be traceable to specific, publicly available, versioned clinical criteria. A black-box model cannot satisfy this requirement.

OMB M-24-10

Active
Active

Federal agencies deploying high-risk AI must maintain a compliant audit trail and provide citizens the right to request human review and the evidence basis for AI-assisted decisions. Applies to CMS, VA, NIH, FDA, OPM, and 10+ other agencies.

FDA SaMD AI/ML Action Plan

Active
Ongoing

Software as a Medical Device must support independent clinician review of the evidence basis. A system that cannot show its evidence chain loses the CDS exemption.

Domains

Explore by domain

Medical, Legal, and Government are live with verified evidence bases. Finance, Scientific Research, and others are in active development.

Ready to see it in your domain?

We work with health systems, life sciences companies, legal teams, federal agencies, and financial institutions. Let's talk about your evidence problem.