Skip to content
Menu

TECHNICAL REVIEW / VERISTADE

Understand the setup.
Know what to test.

Before a test, agree which actions or records the product can check, where information goes and what result would be useful. This page gives your technical team the questions, setup details and evidence to examine.

Product descriptions and evaluation guidance · Updated October 4, 2026 · Not a public API or deployment specification

Permission checks.
Investigation records.

Agent Defender asks whether an action may proceed. Swarm Hunter asks what supplied records support. The diagrams below describe those paths; they do not show a working integration between the products.

Agent Defender / controlled local pilot

A check before
a connected action.

The documented pilot checks the state of a practice file project and creates a branch only after exact approval. It uses Git, a tool for tracking file changes, in one disposable local project.

  1. Proposed actionIdentify the exact action, resource and connected tool.
  2. Permission checkpointCompare the request with the authority granted for the task and any required human approval.
  3. Connected resultAn allowed action can proceed through the connected tool; a request may instead be refused or sent for review.
  4. Review recordInspect the request, decision and saved result before expanding the workflow.

An unconnected or bypassing action is outside the described coverage. This is not antivirus, general computer isolation or a universal agent integration.

Read the dated technical brief
Swarm Hunter / local review build

A case built from
permitted records.

The dated brief describes bringing in activity records, following related events and saving a reviewed case. Maps and replay help a reviewer examine the sequence; they do not authenticate the records.

  1. Explicitly supplied sourceUse supported records you are permitted to inspect; identify their source and coverage.
  2. Related events and findingsExamine advisory findings alongside the supporting records, timing and gaps.
  3. Human assessmentConsider ordinary explanations and separate observed events from inference.
  4. Saved or exported caseCheck retained source labels, original input, settings and assessment.

No automatic agent blocking, discovery of unknown targets or authenticated AI attribution is established. Missing input limits the investigation.

Read the dated technical brief

Version matters. The published Swarm Hunter brief describes 0.6.0. The dated product status record distinguishes local 0.6.1 from staged 0.6.2. Confirm the exact build and agreed setup before testing; these descriptions are not interchangeable release specifications.

Know where information goes.
Know who can act.

A local workflow is not, by itself, a privacy or security guarantee. Agree these details for the specific build before introducing sensitive information.

Summary of the dated public briefs. Not a current deployment specification.
Review questionAgent DefenderSwarm Hunter
InputAction requests, connected resources, permissions and approval information.Supported supplied activity records, public report leads and reviewer assessments.
Described setupA Windows local pilot with a browser interface and a disposable Git project.A portable HTML workspace; the documented Windows launcher uses Python and a local SQLite store.
Stored informationLocal request, permission and records of saved results in the documented pilot.Browser cases or storage in the local service, depending on setup. Exports can include original input and should be treated as sensitive.
External accessDepends on the explicitly connected tool and its authorized setup.A configured public report source can make bounded requests. A source reference alone does not authenticate the record.
Still to agreeExact build, supported environment, access, encryption requirements, retention, deletion, performance needs, support arrangements and any commercial terms. No public guarantee is offered here.

Preserve the source.

Keep original inputs, timestamps and missing intervals visible. A hash is a fingerprint of a file or record. Comparing hashes can help detect changes; it does not prove who created the record.

Check the failure path.

Agree how refusal, changed requests, unavailable components and incomplete records should be handled. Then observe and document the actual behavior.

Define the exit.

Agree how to stop, recover, export permitted results and handle retained data before the evaluation begins.

Agree what would
count as useful.

Agree the question, the expected result and the conditions that would count as failure. Then use fresh cases to test actual behavior. A website example cannot establish product effectiveness.

  1. Choose one bounded workflow.

    Name the product, exact build, responsible reviewer, permitted inputs and actions outside the agreed scope. Start with harmless records or a disposable project.

  2. Predefine success and failure.

    For permission checks, include normal, forbidden, changed, repeated and cases where the checkpoint is unavailable. For investigations, include ordinary activity, an agreed issue and incomplete context.

  3. Keep fresh evaluation cases separate.

    Do not tune a fix using the answers to the evaluation set. Record false warnings, missed issues, unresolved cases, task completion, reviewer disagreement and effort, not just favorable examples.

  4. Inspect the saved result.

    Reopen the record or case and compare the output with the original input, settings and assessment. State what the evidence supports and what remains uncertain.

A worksheet your team can review.Blank fields for scope, data handling, expected outcomes, observed results and the final decision. It contains no test results or implied authorization.

Download Markdown · planning template ↓

Review basis: Agent Defender technical brief, revision 1.0; Swarm Hunter technical brief, revision 1.0; status reported by the company dated October 4, 2026. No SDK, supported public API, independent certification or product benchmark is established by this page.

Does the plan cover the work?
Could it use fewer reviews?

In a fictional workflow, choose review points so every listed connection gets attention. Compare a simple suggestion with the smallest set for that example, then inspect what each choice covers. The research explains why finding the best plan can be much harder than checking one.

Explore the research and working example

A planning model, not a threat detector or a product integration. Permission checks remain a separate requirement.

A specific question.
A reviewable next step.

Describe the workflow and the decision you need to make. Leave credentials, confidential records and suspicious files out of the initial inquiry.

Discuss your setup