Hearth Insights

The control plane
for regulated AI

Hearth Insights enforces accountability at every AI decision: verifiable, immutable, impossible to bypass.

Cryptographically Verified · Air-Gap Ready · Auditable by design.

Request Platform Demo What can you show a regulator?

holt.yml: aml forensic pipeline

agents:

  # Data gatherer: zero network egress
  DataGathererKyc:
    image: aml-forensic/datagatherer-kyc:latest
    bidding_strategy:
      type: "claim"
      target_types: ["GoalDefined"]
    allowed_types: ["Failure", "KycData"]
    network_access:
      mode: restricted
      allowed_domains: []     # no external access
    workspace:
      mode: ro
    volumes:
      - "./data:/data:ro"

  # AI specialist: restricted to model endpoint
  L1KycSpecialist:
    image: aml-forensic/l1-kyc-specialist:latest
    bidding_strategy:
      type: "claim"
      target_types: ["KycData"]
    allowed_types: ["Failure", "KycAnalysis"]
    network_access:
      mode: restricted
      allowed_domains: ["host.docker.internal"]  # locally hosted LLM
    workspace:
      mode: ro

from /etc/holt/policy.yaml

GUT APPENDIX 2026 Clinical evaluation: Holt graph-based architecture outperforms 120B models on local infrastructure.

Review the evidence →

THE DIAGNOSIS

You were promised speed. You got a hidden factory.

Your board wants to know why the AI demo is not in production. The answer is not the model. The answer is that your most expensive engineers are building the plumbing of compliance: audit logs, security gates, access controls, and forensic trails, from scratch, instead of building your product.

This is the Hidden Factory: a second operation running inside your AI programme, invisible to the board, consuming budget that was approved for delivery.

The regulator will not ask how fast the model ran. They will ask what controls were in place, who approved them, and whether you can prove they worked.

THE SOLUTION

The safety chassis

01 THE GLASS BOX

A defensible record
of every decision

When a regulator or your board asks what your AI decided and why, you answer immediately. Every artefact includes the system_manifest_id — the cryptographic hash of the holt.yml that governed it. You can prove not just the decision, but the exact policy configuration in force at the moment it was made.

$ holt hoard 110aeaa9

{
  "id": "110aeaa9ffa7dfe4beb5490d9d9e50b7...",
  "header": {
    "type": "CalculatedMetrics",
    "produced_by_role": "L1BehavioralCalculator",
    "created_at_ms": 1781700566652,
    "parent_hashes": [
      "911aba1b383561bc4add422c3549059c..."
    ],
    "logical_thread_id": "53dc7f3e9607c79e...",
    "metadata": {
      "system_manifest_id": "7ca755cd3d501c5b..."
      // holt.yml hash active at decision time
    }
  },
  "relationships": {
    "produced_by": { "status": "complete" },
    "consumed_by": {
      "parallel_agents": ["L1BehavioralAnalyst"]
    }
  }
}

02 SOVEREIGNTY

Bring the AI to your data

Network isolation is enforced at the agent level, not the deployment boundary. Each agent's permitted external domains are declared in policy and enforced by the engine. Data gatherers operate with zero egress. AI specialists are constrained to named model endpoints. No agent can exceed its declared permissions.

holt.yml: per-agent network enforcement

DataGathererKyc:       # reads raw data
  network_access:
    mode: restricted
    allowed_domains: [] # zero egress

L1KycSpecialist:       # runs AI analysis
  network_access:
    mode: restricted
    allowed_domains: ["host.docker.internal"]  # locally hosted LLM

L1Gatekeeper:          # human review point
  network_access:
    mode: restricted
    allowed_domains: [] # zero egress

03 THE EDITOR MODEL

Codify risk decisions

In high-stakes industries, 99% accuracy is not enough. The gatekeeper agent fires on the completed AI dossier and can only produce a question, a decision to escalate, or a decision to discard. There is no automated approval path. The allowed_types constraint is enforced by the engine — not by convention — and cannot be bypassed.

holt.yml: human-in-the-loop enforcement

L1Gatekeeper:
  image: aml-forensic/l1-gatekeeper:latest
  bidding_strategy:
    type: "exclusive"
    target_types: ["L1TriageDossier"]

  # Engine-enforced output constraints.
  # There is no automated approval type.
  # Human must act.
  allowed_types:
    - "Question"                  # request human input
    - "Failure"                   # system error
    - "L1TriageDecisionEscalate"  # elevate to L2
    - "L1TriageDecisionDiscard"   # close the alert

  workspace:
    mode: ro
  network_access:
    mode: restricted
    allowed_domains: []

REGULATED INDUSTRIES

Built for regulated environments.

FINANCIAL SERVICES

SM&CR personal accountability is already in force.

The FCA Mills Review (January 2026) formalised individual executive accountability for algorithmic failures under SM&CR. The Treasury Select Committee confirmed there is no published safe harbour for AI assurance. You cannot delegate this responsibility.

Enforce immutable AML adjudication trails. Move from manual review to deterministic AI replay. When the regulator calls, you answer with a forensic artefact, not a log file.

ENFORCEMENT PRECEDENT

£44.1m FCA enforcement action (2025)

Governance and oversight failures

REGULATORY EVENT

FCA Mills Review (January 2026)

SM&CR personal liability for AI decisions confirmed

HEALTHCARE

The immutable ledger for clinical environments.

Current automated AI pipelines lack the comprehensive audit trails required for regulated clinical data. Holt enforces these controls. Operating in secure, air-gapped environments with minimal privileges, Holt records every text process, retrieval, prompt, and model execution to a cryptographically verifiable ledger. From automating clinical notes to orchestrating complex LC-MS analytical pipelines, patient data never leaves your perimeter, and every decision is fully reconstructable.

GOVERNMENT & LEGAL

Classified. Air-gapped. Zero egress.

Deploy in environments with no external connectivity. Process classified directives with zero external API calls and no licensing dependencies.

RESEARCH & LIFE SCIENCES

Featured in the Gut appendix.

In a recent peer-reviewed study (Gut Appendix, 2026), Holt was tested in UHSFT to detect HPO terms from gastrointestinal clinical notes using an agentic workflow. By decomposing complex clinical workflows into discrete steps and enforcing a strict review cycle, Holt allowed a 3-billion parameter edge model to outperform a 120-billion parameter alternative on the same local hardware. Superior precision with a drastically reduced carbon footprint.

SUPPLY CHAIN INTEGRITY

Trust is not a sentiment.
It is a signature.

We do not ask you to trust our code. We ask you to verify our artefacts. Every release of the Holt engine is cryptographically signed (Cosign) and published to an immutable transparency log. You run exactly what we built, with a mathematical chain of custody from our compiler to your air-gap.

The question is not whether your AI is accurate. The question is whether you can prove it.

Request Platform Demo Verify Releases

cosign verification: latest release

$ cosign verify-blob \
    --key cosign.pub \
    --signature checksums.txt.sig \
    checksums.txt

Verified OK

# holt-linux-amd64
b7233901ce42c70dc220b3c64eb8cc17
6e63d790c5db1b841c3ad97cb2a8c4bb

# docker image
ghcr.io/hearth-insights/holt-engine/
  holt-orchestrator:latest

sha256:ea41cdd3db3244e7ecd0177321e1120b
       ed8acb7da8f99eb9efb64923ca79d0d7

# built:  2026-06-15 20:37:18 UTC
# commit: ce2f3176cb8a338a1bd12d0f4808944

github.com/hearth-insights/holt