Turn one AI task into a report your team can review.

• Six artifacts from one command: a normalized failure record, source-linked evidence, bounded classification candidates, a next-check proposal, and a triage handoff as Markdown and as a standalone HTML page.
• A stop rule: incomplete, blocked or instruction-bearing input returns BLOCKED with the stop reason and the missing evidence listed.
• Five single-purpose skills for your AI assistant: failure intake, evidence preserver, failure classifier, proposal reviewer and ticket drafter. Each carries an expert playbook with a procedure, a checklist, common mistakes and a worked example on the bundled data.

Best for: QA engineers and SDETs who hand failures to developers every week and want the same evidence standard each time. You need: Node.js 22 and a terminal.

Next: L2 UNDERSTAND, AI Application Evaluation, applies the same evidence discipline to an AI feature: criteria, sample cases and a result review the team can inspect.

• An evaluation contract, a fixture plan, a result review and an open-risk list, written as JSON and as an HTML report.
• Deterministic checks, semantic criteria and prohibited outcomes kept apart, each with an owner and pass, fail and unresolved definitions.
• A CI evaluation gate for GitHub Actions, GitLab CI or make that returns a verdict per critical segment, with two policies (default and exploratory).

Best for: QA leads and engineers who must judge an AI feature before release and need more than a demo to point to. You need: Node.js 22 and a terminal.

Next: L3 APPLY, AI Test Automation Review, helps you choose which AI-assisted tests are worth writing, propose the smallest change and name who maintains it.

• A test contract, an automation proposal, a validation checklist and a maintenance plan from one command, as JSON, Markdown and HTML.
• A target register and policy rubric for picking stable, valuable targets before anyone generates test volume.
• An ADR (architecture decision record) template to record why a test change was accepted or rejected, plus a RACI and a review checklist.

Best for: Test-automation engineers whose AI-assisted suite is growing and who want quality and cost to stay under human ownership. You need: Node.js 22 and a terminal.

Next: L4 SCALE, Enterprise AI Testing Operating Blueprint, turns these practices into an owned, observable path across teams, with gates and a staged rollout.

• A five-check record, a variation summary, a risk list and a next-check proposal, plus an HTML report.
• A stop at the first unresolved risk, so you know whether the demo deserves deeper testing before you invest in it.
• Five single-purpose skills: scope intake, check selector, trial runner, variation reviewer and risk recorder. Each carries an expert playbook with a procedure, a checklist, common mistakes and a worked example on the bundled data.

Best for: QA engineers and technical leads starting LLM testing who want a short, repeatable routine. You need: Python 3.10 or newer and a terminal.

Next: L2 UNDERSTAND, Non-Deterministic Evaluation Design, defines how much variation is acceptable and compares runs on evidence the team can defend.

• An evaluation contract, a comparison plan, an evidence summary and an uncertainty register, as JSON, Markdown and HTML.
• A contract template, gate policy, decision register, trials template and rubric to fill in with your own system.
• An evaluation gate for GitHub Actions, GitLab CI, pre-commit or make.

Best for: Engineers who own AI quality criteria and regression decisions and are tired of subjective arguments. You need: Node.js 22 and a terminal.

Next: L3 APPLY, Production AI Evaluation System, moves evaluation out of the notebook into a dataset, an evaluator, thresholds, CI and monitoring that owners can act on.

• An evaluation system blueprint, a threshold map, a CI proposal and a monitoring loop, as JSON, Markdown and HTML.
• A warranty-assistant example with a failing run and a fixed run side by side, in English and Polish.
• Offline routing and tool-recovery sample inputs, so you can see how the evaluation behaves when a step fails.

Best for: Teams that run AI evaluations beyond a one-off experiment and need a signal that survives the demo. You need: Node.js 22 and a terminal.

Next: L4 SCALE, Enterprise Evaluation Platform Blueprint, adds shared governance across teams: risk tiers, an evaluator registry, gates and staged adoption.

• A test scope, a source-linked injection evidence pack, a finding draft and an owner action, plus an HTML report.
• 24 inert prompt-injection cases in 6 categories, each with a unique canary token, a recording sheet and a converter that turns the filled sheet into the review input. The library sends nothing: you send the inputs to a system you are authorized to test and paste the answers.
• An authorization gate that comes before any check, and a blank input template with a field guide for your own scope.

Best for: Security-minded QA engineers who test an LLM application and need a finding they can defend, with the limits of the evidence stated. You need: Python 3.10 or newer and a terminal.

Next: L2 UNDERSTAND, LLM Security Review, starts from a threat model and turns it into authorized checks and findings with owners.

• A threat and control map, a case plan, an assessment register, finding drafts and an owner retest and remediation plan.
• Templates for a decision record, RACI, risk register, review checklist and a review-input file.
• CI examples for GitHub Actions (including a scheduled run), GitLab CI, pre-commit and make.

Best for: Security engineers, QA practitioners and technical leads who review LLM risk and need each finding to have an owner and a retest. You need: Node.js 22 and a terminal.

Next: L3 APPLY, AI Security Test Implementation Workflow, runs your own authorized, inert checks against one endpoint you name, keeps hashed evidence and connects the results to your delivery process.

• An authorized local check runner. It sends the inert, canary-based checks your contract defines to one endpoint you name, only inside a valid authorization window and boundary, and records hashed evidence in the results schema. It sends nothing when authorization is missing, expired or out of scope. It runs only the checks you wrote: it does not scan, attack or certify.
• A bundled stub target with a hardened and a vulnerable mode, so you see a clean run and a failing run locally first, and a make and GitHub Actions chain that fails on missing or stale results.
• A security check contract, an evidence schema, an integration plan, a run record and a gate decision as reviewable documents, with templates for an authorization record, check catalog, cleanup runbook, findings register and results policy.

Best for: Engineers who want authorized AI security checks to run from one command and from CI, with the contract, the authorization and the evidence rules settled first. You need: Node.js 22 and a terminal.

Next: L4 SCALE, Enterprise AI Security Framework, makes AI security testing an owned operating model with a control library, monitoring and staged adoption.

• A tool inventory in which every tool has one declared purpose.
• A permission matrix that starts from least privilege, and a side-effect map that names an owner for every write or call.
• An approval handoff that proposes one bounded change and keeps access limited to it.

Best for: Developers and tech leads who choose what an AI agent may call, write or change. You need: Node.js 22 and a terminal.

Next: L2 UNDERSTAND, Orchestration Runbook, coordinates a multi-step agent workflow with visible state, handoffs and recovery points.

• A stage runbook, handoff contracts, a state record and a recovery plan, as JSON, Markdown and HTML.
• An event command that applies one declared event to a saved state, so a run can resume from a stable artifact.
• Templates for stages, events, a handoff register, an ADR, RACI and a risk register.

Best for: Engineers who coordinate multi-step agent runs and need to know which state is real after an interruption. You need: Node.js 22 and a terminal.

Next: L3 APPLY, Evaluated Delivery System, puts an evaluation gate, release evidence and rollback in front of every AI-assisted change.

• An evaluation contract, a release evidence plan, a recovery record and a decision note, as JSON, Markdown and HTML.
• A gate that compares baseline and candidate runs against your metric policy and returns a verdict.
• CI kit: a run-gate script, GitHub Actions and GitLab CI gates, a scheduled re-evaluation workflow, pre-commit and make.

Best for: Teams that want every AI-assisted change to pass an evaluation gate before it ships. You need: Node.js 22 and a terminal.

Next: L4 SCALE, Engineering Operations Blueprint, adds owners, signals, change control and a recovery route checked against your drill record, across teams.

• A research brief in which every fact is an exact quote linked to its source ID.
• A fact ledger that lists your own inferences and any declared conflicts between sources.
• An uncertainty list that shows which unknowns affect the decision, and one proposed next check.

Best for: Creators and operators who need research a colleague can check against its sources. You need: Node.js 22 and a terminal.

Next: L2 UNDERSTAND, Knowledge to Decision Workflow, turns approved sources into a decision note with options compared on the same criteria.

• A source map, an option table, a decision note and a recovery record, as JSON, Markdown and HTML.
• Options scored against your criteria and against hard minimums you define; scores are labeled as your judgments.
• Templates for a decision charter, criteria library, decision log, hard-gate register, a RACI (who is responsible, accountable, consulted and informed) and a review checklist.

Best for: Managers and specialists who turn gathered knowledge into a decision with options, evidence and a recorded choice. You need: Node.js 22 and a terminal.

Next: L3 APPLY, Repeatable Business Process, writes a weekly process down with triggers, approvals, exception routes and duplicate protection.

• A process map, an approval matrix, an exception route and a recovery checklist, as JSON, Markdown and HTML.
• A ledger that stops a duplicate run: you supply the record of what already ran.
• Twelve English templates, with Polish variants: an SOP (standard operating procedure), steps, payload, exceptions, approval matrix, ledger, review checklist and more.

Best for: Business owners and team leads who repeat the same process every week and want it written down, reviewed and repeatable. You need: Node.js 22 and a terminal.

Next: L4 SCALE, Work Operating System, sets up one reviewed operating system for the recurring work of a whole team.