Our approach

Confidence comes from how the system works.

Six principles behind IRONWARD, each stated with its honest limit.

Coordinated intelligence

Specialist agents share a defined task and evidence, with separate responsibilities for investigation, action planning and verification.

LimitMore agents is not the goal. We will measure whether role separation improves outcomes and operator effort.

Permission outside the model

The execution layer checks permitted operations independently of an agent’s language output. Customer controls determine what may happen.

LimitPlanned capability. Policy, isolation and effect tests are required before any safety claim.

Evidence behind the result

A useful outcome identifies what was checked, what changed and what remains uncertain. We are building verification into the workflow rather than adding it as a final report.

LimitMissing evidence keeps a task open. No green status without the record.

Deployment choice

Local processing and approved cloud routes are design options, selected around customer requirements and validated before release.

LimitA local model keeps model processing on your premises. Logs, backups, updates and support access are agreed separately.

Measured improvement

Pilots will compare verified task completion, operator effort, regressions and cost per accepted result against the partner’s current workflow.

LimitNo pilot results exist yet. Failed and abandoned cases will be kept, not hidden.

Open where it helps the ecosystem

We intend to publish our evidence-record and approval-policy formats so agentic security tools can be tested on the same terms.

LimitThe formats, not the product. Published during the development period.

Both sides of the outcome

In the planned lab scenario, the verification step checks that the threat activity was blocked and that a normal reference transaction still completes. Either one failing fails the task.

containment · approved scopeactor Aactor Bdeny witnessbenign clientconfirmed

Where the models run

Routing follows your data policy. Local-only tasks wait or fail openly if local AI is unavailable; they never quietly switch to the cloud.

local boundary · data policylocal AI nodes · plannedapproved API providereligible data only · logged

Trust is earned through observable outcomes, not promised in copy.