Change Control Process
ZenPort AI Technologies Inc. (Kaspen) · Version 1.0 · Effective October 6, 2026 · Owner: CEO
1. Purpose
This process defines how Kaspen proposes, reviews, tests, approves, releases and reverses changes to production systems. Its goal is simple: no change reaches customers unless a second person has reviewed it, automated checks have passed, and we know how to undo it. It supports our Security commitments and the change management criteria of SOC 2.
Kaspen is an AI agent that acts on behalf of property managers. A change to a prompt, playbook or model can alter what Kaspen says to a tenant as much as a change to code. This process treats both the same way.
2. Scope
This process covers every change to systems that store, process or transmit customer data, including:
- Application source code and dependencies;
- Infrastructure, cloud configuration and infrastructure-as-code;
- Database schemas and migrations;
- CI/CD pipelines, secrets and access permissions;
- PMS, messaging and other third-party integrations;
- AI behavior: system prompts, default playbooks, tool definitions, safety gate rules, model selection and model version.
It applies to all employees and contractors who contribute to Kaspen's software or infrastructure.
Out of scope. Changes customers make through the product (their settings, approval limits, playbook edits and autonomy levels) are governed by the product's own permissions, dry-run and audit controls described in section 10.
3. Definitions
- Change. Any addition, modification or removal of code, configuration, AI behavior, dependency or infrastructure that affects production.
- Production branch. The `main` branch, from which production is built and deployed. It is protected and cannot be pushed to directly.
- Standard change. A low-risk, pre-approved type of change, such as an automated dependency patch or a documentation fix. It still goes through a pull request and review.
- Normal change. The default for features, fixes and AI behavior changes. Requires full review, passing checks and staging verification.
- Emergency change. A change needed to resolve an active production incident or a serious security vulnerability, where waiting would cause greater harm. Follows section 9.
- Author. The person who makes the change.
- Reviewer. An engineer, other than the Author, qualified to approve the change.
- Replay set. A library of real, de-identified past tasks with known correct outcomes. Kaspen's behavior on these tasks is compared before and after a change.
- Safety set. Tests that Kaspen must always pass, such as never leaking one party's data to another, catching emergencies, and never taking a locked action.
4. Roles
| Role | Responsibilities |
|---|---|
| Author | Opens the pull request, explains what changed and why, adds or updates tests, and resolves review comments |
| Reviewer | Reviews for correctness, security, privacy and AI safety; approves or requests changes. Never the Author |
| Engineering Lead | Reviews high-risk changes (security, data access, safety gate, model changes); approves emergency changes and check exceptions |
| Release Manager | Confirms the release checklist, triggers or supervises the production release, and decides on rollback. May be the Engineering Lead |
| CEO | Owns this process, approves exceptions, and is the escalation point for violations |
No single person may author, approve and release the same change to production.
5. Source control
All code and AI configuration live in our source control system, our single source of truth. Prompts, default playbooks, safety rules and model settings are stored as versioned files in the repository, not edited live in production. Repository access follows least privilege, and every commit is attributable to a named person.
- Work happens on short-lived feature branches.
- Feature branches merge into `staging` through a reviewed pull request, which deploys to the staging environment.
- Production releases happen only by a reviewed pull request from `staging` into `main`.
Branch protection on `main` and `staging` requires a pull request, at least one approving review from someone other than the Author, passing required checks, and dismissal of approvals when new commits are pushed. Force pushes and direct pushes are blocked.
6. Pull requests and review
Every change reaches a protected branch through a pull request. Each pull request must:
- Describe what changed, why, and link the related task or incident;
- Flag any effect on security, privacy, customer data, database schema or AI behavior;
- Show how the change was tested.
Reviewers check correctness, security, data handling, test coverage, migration safety and that no secrets are committed. For AI behavior changes, reviewers also read the replay and safety results (section 7) and sample changed outputs. AI-assisted code review may support but never replaces human review. Significant changes after approval require re-approval.
7. Automated checks
Required checks run on every pull request and must pass before merge:
- Build, lint and formatting;
- Unit and integration tests;
- Static application security testing;
- Dependency scanning for known vulnerabilities;
- Secret scanning;
- Replay set: for changes that can affect AI behavior, no regression in correct outcomes against the current production version;
- Safety set: zero failures. A single failure blocks the merge.
Bypassing a required check needs a recorded exception approved by the Engineering Lead or CEO (section 12).
8. Environments and release
Kaspen runs three separate environments with separate credentials: development for building and local testing, staging for integrated testing on de-identified or test data, and production for customers. Production credentials are never available in development or staging.
Changes move from development to staging to production. Before a production release, the Release Manager confirms:
- An approving review from someone other than the Author;
- All required checks passed;
- The change was verified in staging;
- Database migrations are reviewed, backward compatible and reversible;
- Security, privacy and AI safety effects are reviewed and any findings fixed;
- A rollback plan exists;
- Monitoring and alerts cover the change;
- Customers are told in advance about changes that affect how Kaspen behaves for them.
Production deployments run through the CI/CD pipeline from `main`. Each deployment is logged with the commit, the person who triggered it and the time.
Model changes. Switching AI models or model versions is a normal change. It requires a replay comparison against the current model, a clean safety set, and a staged rollout before all customers move.
9. Emergency changes
Emergency changes are allowed only for a severity 1 or 2 production incident or a serious security vulnerability.
- Approve. The Engineering Lead or CEO approves before release, or immediately after if they cannot be reached in time.
- Keep it small. Ship the smallest change that fixes the problem. It may shorten staging time but is never released without testing. The safety set still runs.
- Document. Open the pull request, get review from a second engineer and complete documentation within one business day.
- Review. Label it as an emergency change and cover it in the incident review.
Pausing Kaspen for a customer, workflow or channel is an operational control, not a change, and may be used at any time to contain an incident.
10. Customer-side controls
Customers control how Kaspen behaves for their own account. These controls are separate from this process but work alongside it:
- New workflows start in a dry run, where Kaspen drafts and nothing is sent, until they meet readiness criteria and the customer's owner or manager turns them on;
- Workflows go live at an approval-required level, never fully autonomous;
- A wrong send or safety failure automatically returns that workflow to dry run;
- Every setting change is recorded in the customer's audit log with who made it and when.
11. Rollback
Every production release has a rollback plan. The default is to revert the merge or redeploy the last known good build through the pipeline. AI behavior changes can be rolled back by redeploying the previous prompt, playbook or model version. Database migrations are written so application code can roll back on its own; destructive migrations need separate approval and a verified backup first. The Release Manager or Engineering Lead decides on rollback and records the decision.
12. Exceptions
Any departure from this process must be requested from and approved by the CEO. Exceptions are recorded with a risk assessment and an end date, and reviewed until closed.
13. Audit trail
Change records serve as SOC 2 evidence. The source control system keeps an immutable history of pull requests, reviews, approvals, check results, merges and authors. The CI/CD system keeps build, test, security scan, replay and safety results and deployment logs. Protected branch history is never rewritten or deleted. Records are kept under our retention policy and provided to auditors on request.
14. Violations
Report suspected violations to the CEO. Violations may result in removal of system access and disciplinary action, up to and including termination of employment or contract.
Document control
| Version | Date | Change | Author | Approved by |
|---|---|---|---|---|
| 1.0 | October 6, 2026 | Initial version | Engineering | CEO |