Security and correctness

Faster only counts if it is still right

Turbokern changes the code that runs your models. These are the rules it follows so that every kept change is correct, accurate within your budget, and traceable.

Guards on every runDesign rules
Numerical check
Every kernel
Accuracy check
Every precision change
Audit log
Every decision
Data pooling
Opt-in, features only
SOC 2 Type II
Planned

Principles

Four rules the agent cannot break

01

Measure before keeping

No prediction, heuristic or model output is ever kept without a measurement on the target GPU.

02

Correct before fast

A kernel is benchmarked only after it passes a numerical equivalence check.

03

Your budget, your eval

Accuracy is judged on your evaluation set against the budget you set.

04

Everything on the record

Every candidate, result and decision is written to the run report.

Method

How correctness and accuracy are checked

Kernel changes are compared to the reference implementation on representative and edge-case inputs, within a numerical tolerance per data type.

Precision and sparsity changes are scored on your evaluation set. Each score carries a confidence interval, and results the eval set is too small to resolve are flagged rather than kept.

  • Tolerances are set per data type and can be tightened per model.
  • Combined changes are re-measured together before packaging.
  • A person approves deployment unless you choose otherwise.
  1. C-041Attention, 64×64 tiles3-stage pipelineMax error above BF16 tolerance on long sequences.fails checkRejected
  2. C-066INT4 on all layersper-group 128Accuracy −2.8%, budget 1.0%.over budgetRejected
  3. C-067FP8, 76 of 80 layers4 attention layers in BF16Accuracy −0.3% (±0.1), inside budget.−19.6 msKept

Example run on one H100. Candidates, speedups and reasons are illustrative, not measured results.

Standards context

How Turbokern relates to the frameworks you report against

Turbokern can produce evidence that supports your own compliance work. It is not a certification body and holds no certifications today.

Turbokern does not certify customer systems. Statuses reflect plans at pre-launch and will change.
FrameworkWhat it asksHow Turbokern can helpStatus
SOC 2Controls over security and availabilityTurbokern's own SOC 2 Type II auditPlanned
ISO/IEC 27001An information security management systemPolicies designed to align; certification not yet soughtNot started
EU AI ActDocumentation and quality management for some AI systemsRun reports document model changes and their measured effectEvidence only
ISO 26262, IEC 62304Functional safety for automotive and medical softwareOn-device measurements for your validation; no safety claimsEvidence only

Audit trail

Every decision, with its evidence

The run report is an append-only log. Each entry records the candidate, the hardware, the software versions, the measurement and the verdict. Reports can be exported to your tracking system.

  • GPU model, driver, CUDA and TensorRT versions per entry.
  • Hashes of input model and output engine.
  • Who approved deployment and when.
run_07.report.jsonl
{"id": "C-067", "stage": "quantra",
 "gpu": "H100-SXM-80GB", "cuda": "12.6",
 "p95_ms": 41.8, "acc_delta": -0.003,
 "verdict": "kept"}
// example entry, illustrative values

Security

Controls for your models and data

IsolationPer-tenant storage and compute for runs, traces and artifacts.
EncryptionData encrypted in transit and at rest in the managed service.
Your environmentVPC and on-prem modes keep weights and data on your infrastructure.
AccessSSO and role-based access planned for general availability.
Data poolingOff unless you opt in; only anonymous performance features are pooled.
DisclosureReport vulnerabilities to security@turbokern.com.

FAQ

Questions engineers ask

Will Turbokern see our model weights?

In the managed mode, yes, inside your isolated tenant. In VPC and on-prem modes the weights never leave your infrastructure.

Are you SOC 2 certified?

No. A SOC 2 Type II audit is planned before general availability and has not started.

Can an optimized engine behave differently from the original?

Within the tolerances and accuracy budget you set, yes; that is what optimization means. Anything outside them is rejected and logged.

Who is responsible for deploying the engine?

You are. Turbokern packages and documents it; by default a person on your team approves deployment.

Next step

Ask us anything about data handling

Our security contact will answer questions before you share any model.