Evaluate Kurahi

This pack is what an evaluator needs to decide, on evidence, whether Kurahi fits. It takes about half a day for the core exercises; the rest of the evaluation is your own workflow. Everything in the pack runs on the Free plan, which has no time limit. For an evaluation that needs Enterprise features (digital staff at L2 and L3, governance, air-gap), contact us for written evaluation terms.

1. Prerequisites

• One machine: Linux x64, macOS on Apple silicon, or Windows 10/11 x64. No Node.js needed; the download carries its own runtime.
• A model: a local model through Ollama for the fully local configuration, or an API key for a curated provider. See Deployment and data boundaries to choose.
• Optional: a chat workspace (Slack, Teams, LINE) for approvals on a phone.

2. Install

curl -fsSL https://kurahi.com/install.sh | bash        # Linux, macOS

irm https://kurahi.com/install.ps1 | iex               # Windows

Then kurahi health should report the gateway, integrity and audit chain as ok.

3. Sample policy

Save as evaluation.yaml, then kurahi policy apply evaluation.yaml --sign.

version: 1
name: evaluation
description: Starter policy for a kurahi evaluation. Tighten it before production.
agents:
  "*":
    tools:
      deny: ["browser"]
    limits:
      maxSteps: 40
      wallClockPerJobMs: 900000
budgets:
  currency: JPY
  fleet:
    tokensPerDay: 2000000
    toolCallsPerHour: 600
    onBreach: pause
kill:
  mayFlip: [owner, admin]
approvals:
  # channel: "slack:#kurahi-approvals"   # uncomment to receive approvals in chat
  timeoutMs: 600000
  require:
    - class: exec
    - class: write
network:
  egress:
    deny: ["169.254.169.254"]
data:
  neverRead: ["~/.ssh/**", "**/.env", "**/*.pem"]
  neverWrite: ["/etc/**"]
  redact:
    fields: ["@personal", "inspector"]
    restoreTo: [message]
shield:
  mode: stop
attestation:
  required: true

Every file write and every shell command now needs your approval, personal data in tool results is masked before the model sees it, and ransomware detection is armed.

4. A business workflow, end to end

This exercise uses illustrative data you create; it is not a customer system.

1. In the agent’s workspace, create inspections.csv:

   asset,inspector,reading,limit
   Press 4,Sato Ken,92,80
   Press 7,Tanaka Yui,61,80
   Conveyor 2,Ito Mai,85,80

2. Ask the resident:

   kurahi agent --message "Read inspections.csv. For every asset over its limit, draft a work order in work-orders.md with the asset, reading and a recommended check. Do not include inspector names."

3. Watch what happens, and check each step:

   Step	What you should see
   Retrieval	The agent reads the file. With the context guard on, inspector names reach the model as [[PII:INSPECTOR:1]] placeholders.
   Proposed action	The agent asks to write work-orders.md. The policy holds it for approval.
   Approval	Approve in the terminal (or in your chat channel). Try it once with deny first: the job reports the refusal and stops cleanly.
   Outcome	work-orders.md lists Press 4 and Conveyor 2 only.
   Evidence	kurahi evidence list, then kurahi evidence export <job> --out job.zip and kurahi evidence verify job.zip. The bundle shows what was read, the approval, the write and the redaction counts.

5. Failure and recovery drills

Drill	How	Pass when
Provider outage	Configure a fallback (kurahi models fallbacks add ollama/<model>), then break the primary key	The job completes on the fallback, and the evidence records which model answered
Rejected approval	Deny the write in step 4	The job stops with a clear refusal; nothing is written; the evidence shows the denial
Stop everything	kurahi policy kill --all --reason drill, then send a message	New work is refused and running jobs are cancelled; kurahi policy release --all resumes
Ransomware drill	Create HOW_TO_DECRYPT_FILES.txt in the workspace	Every agent stops; kurahi audit tail shows shield.trigger; release resumes
Tampered evidence	Change one byte inside an exported bundle	kurahi evidence verify fails and names the broken record
Continuity	kurahi continuity run --throwaway --disruption all --out ./continuity-out	The report passes and kurahi continuity verify re-checks it offline
Inventory	kurahi inventory export -o inventory.json, then kurahi inventory verify inventory.json	The signed inventory lists every agent, model and policy rule and verifies

6. Success measures

Agree these at the start, and record the result of each:

1. Every drill in section 5 passes on your machine.
2. Your own workflow (chosen in the kick-off call) runs end to end with the approvals your organisation needs, and every run leaves a bundle that verifies.
3. Your security reviewer can answer, from the evidence and the deployment page, what data left the host and why.
4. Time from request to approved outcome for your workflow, measured against how it is done today.

7. Support during the evaluation

• E-mail support@kodasoken.com; we reply within two business days.
• A kick-off call to choose your workflow, and a review call at the end.
• Security questions: security@kodasoken.com.
• Not included during evaluation: production support hours, custom development, or a service level. Those are part of Kurahi Enterprise.

See also: Capabilities in this release.