HARDSEAL EDGE · AI RUNTIME EVIDENCE LAYER

Prove the exact bytes your AI produced
from declared run artifacts
under declared conditions.

Hardseal Edge produces SHA-256 hash-chained, offline-verifiable evidence packets for AI inference running at the edge — on Jetson, on air-gapped servers, on any system where you cannot phone home. The packet captures device state, model identity, benchmark telemetry, sensor lineage, and explicit limitations into a single SHA-256 hash chain that anyone can independently verify with the stdlib-only Python verifier or the reference browser verifier on /verify.html. A passing verification proves the packet is internally consistent and free of accidental change; defeating a deliberate re-package requires the operator-held signing secret (Ed25519 attribution is on the roadmap).

Runs offline. No phone-home. No SaaS dependency. Built for defense and aerospace suppliers, DIB primes, autonomy operators, and regulated AI deployments where "trust us" is not an answer.

// WHY EDGE

Cloud-rooted AI evidence breaks where defense-grade systems live.

Standard AI observability stacks assume a steady connection to a SaaS endpoint. They log to centralized buckets, attest via cloud KMS, and require a phone-home for verification. That model fails the moment your inference runs on a denied network, an air-gapped environment, a moving platform, or any setting where a defense customer asks: can you prove what the AI did, without trusting your vendor's cloud?

Hardseal Edge inverts the model. The packet is the artifact. The verifier is open and runs anywhere. The chain is reproducible byte-for-byte. There is no standing SaaS dependency, no required API call, and no vendor-controlled service needed to recompute the chain. The evidence is the same whether you read it tomorrow or in a litigation hold seven years from now.

// WHAT IT IS

One packet schema. Five sections. One chain root.

01 · DEVICE

Hardware identity, frozen at run-time.

Platform, model number, JetPack/L4T version, power mode, clock pinning state, thermal floor, MAC address fingerprint. Captures the silicon the AI ran on, not the silicon you wish it ran on.

02 · MODEL

Engine binary, hashed.

SHA-256 over the deployed inference engine (TensorRT, ONNX, OpenVINO). Precision class. Input shape. Build commit. The exact bytes that produced the inference, not a "model card" describing them.

03 · BENCHMARK

Statistical telemetry, not a single number.

Per-block min/mean/p50/p99/max/stdev across the full run. Million-iteration soaks. Thermal saturation traces. Throughput, not best-case marketing FPS.

04 · SENSORS

Power, thermal, governance state.

Live power-rail samples, thermal zones, DVFS state, governor mode. Captures the conditions the AI was actually operating under. Hash-sealed alongside the inference into the same SHA-256 chain, not in a separate observability silo.

05 · LIMITATIONS

Scope of assurance — declared, not implied.

Explicit statement of what the packet does not certify. Not a CMMC C3PAO assessment. Not a guarantee of mission outcomes. Operator-defined truth boundary.

∞ · INTEGRITY

SHA-256 hash chain over canonical JSON.

Per-section hash + chain root. Banned-phrase scan to keep packets free of legal overclaims. Verifier reproduces the chain client-side. Mutate any byte → integrity fails.

// PROOF

The evidence is on disk. Run the verifier yourself.

Hardseal Edge has shipped real evidence packets from a Jetson Orin Nano Super, including thermal-saturation soaks at one million inferences per packet across multiple precision classes and power modes. Each packet has a SHA-256 chain root that is byte-reproducible from the data inside the packet, using a verifier you read in full before you run.

trophy_case_bundle: 11 sealed packets · 269 KB
bundle_sha256: f6fdc719edc303eae774177f265fe86c6a55162c73d2540b5626ab58189ce7c9
workloads: yolov8n / yolov8m / yolov8L INT8 + FP16
power_modes: 25W · MAXN_SUPER (pinned clocks, 3× replicate chain)
soak_depth: 1,000,000 inferences per packet (steady state)
verifier: verify_standalone.py (Python stdlib only) · /verify.html (browser)
tamper_test: cross-packet section-hashes swap caught by red-team suite
// DEPLOYMENT

Deploy Edge where AI evidence has to survive offline.

Hardseal Edge is built for teams running AI inference in environments where cloud observability is unavailable, unwanted, or insufficient — Jetson devices, isolated networks, test ranges, air-gapped labs, autonomy stacks, and regulated supplier environments.

Edge does not replace a Cyber AB Registered Provider Organization (RPO), a C3PAO, or a CMMC readiness engagement. It produces verifiable AI-runtime evidence packets. When that evidence is relevant to a broader compliance program, your RPO and assessor interpret it under their own judgment.

All pricing shown below is pilot anchor pricing; engagements are subject to final MSA execution and founder approval.

EDGE · MONTHLY

$495 per environment / month

Hardseal Edge CLI, hash-chained sample packets (SHA-256 chain root + sidecar digest), browser verifier, onboarding documentation, and standalone Python verifier for downstream technical review. Deploys offline. Generates evidence packets from your own inference runs.

$495 / mo per environment
EDGE · PILOT

First-environment validation

Founder-led pilot for one AI runtime environment: packet generation, verifier walkthrough, trophy-case comparison, evidence-boundary review, and deployment notes for your technical or compliance team. Scope quoted after environment review.

Pilot scope quoted after review
EDGE · ENTERPRISE

Multi-environment SOW

Portfolio deployments across devices, models, power modes, sensor pipelines, and supplier environments. Designed for primes, autonomy teams, aerospace suppliers, and regulated AI operators that need repeatable packet evidence across many runs.

Contact for SOW
Adjacent product: Hardseal Core is a separate CMMC Level 2 evidence engine with 17 collectors for Windows, Active Directory, M365/Entra, network architecture, filesystem, audit logs, and system state. Core and Edge share Hardseal's evidence-integrity primitive but use distinct schemas and serve distinct scopes. Edge is not a general CMMC tool; Core is not an AI runtime tool. Pricing and SOWs for direct sales as shown; bundled and Enterprise engagements negotiated separately.
// SCOPE OF ASSURANCE

What Hardseal Edge does not claim.

Hardseal Edge produces integrity evidence, not a certification. It does not certify the underlying AI system as compliant with CMMC, NIST 800-171, ISO 27001, FAA airworthiness, NIAP, or any other regulatory regime. It does not replace a C3PAO assessment, an RPO engagement, an FAA type certification, or any human assessor judgment. It does not guarantee mission outcomes, model accuracy, or safety properties of the AI system.

What it does: prove that the bytes you see in a packet are the bytes that were hash-sealed into the SHA-256 chain from declared run artifacts on a Hardseal-instrumented engine, on a declared device, under declared sensor and power conditions — and that a naive or accidental modification to the packet is caught when the chain is recomputed. A deliberate re-package by an adversary with the hashing tools is not caught by the verifier alone; that is what the operator-held signing secret (and the Ed25519 signature on the roadmap) is for. The truth of the underlying measurement is the operator's responsibility under the limitations section of every packet. The evidence is the artifact. The interpretation is yours.

If you ever see a broader claim in any slide, document, or email — treat that as a bug in our language, not a property of the verifier. The code only enforces the narrow integrity claim described here.