GRC Engineering ClubRubric vgcp-1.0.0 · pass = 70

CGE-GCP Capstone Rubric

The CGE-GCP capstone validates that a candidate can build and operate an automated GRC engineering pipeline on Google Cloud: discovery-first Terraform, an evidence store they own, reports that trace to stored records, and remediation through a least-privilege identity. Submissions are scored across five dimensions blending mechanical static-analysis (Tier 0) with qualitative judgment (Tier 1). Pass = weighted average >= 70 and no auto-fail triggers.

Submission requirements

  • Visibility: public
  • Max size: 50 MB
  • Max files: 10,000

Required artifacts

  • README.md that opens with a tenancy statement and explains how to deploy the pipeline from an empty project
  • Terraform (google provider) for at minimum: the foundation, stages 1 to 3 (discovery, activation, evidence store with a locked reports bucket), and Stage 6 remediation in report_only mode or higher
  • At least two reports generated from Firestore only, with real run history (the daily schedule, not a burst of manual runs)
  • The policy gate (OPA/conftest) and drift workflows, armed on the candidate's own fork through Workload Identity Federation
  • docs/CONTROLS.md mapping every check, gate rule, and design choice to NIST CSF 2.0 and NIST SP 800-53, current with the code
  • An evidence/ folder with the WORM proof (failed delete), the separation-of-duties proof (reporter denied), and a gate failure from a closed test pull request

Infrastructure-as-Code Quality

weight 20%

The pipeline is defined as google-provider Terraform: one root module per stage with separate state, a pinned provider, APIs managed as code, and additive IAM only.

Tier 0 (deterministic)

  • • terraform fmt -check (no diffs)
  • • terraform validate per stage (passes)
  • • tflint (no errors)
  • • checkov (no HIGH/CRITICAL severity findings on infra)

Tier 1 (AI scored)

  • • Stages are separate root modules with their own state (blast-radius boundaries), composed through outputs/remote state, not cross-stage mutation
  • • google provider pinned, project targeting explicit through variables, no hardcoded project IDs or numbers beyond documented variables
  • • Discovery-first pattern present: discovery changes nothing and emits a gap map, and activation acts only on the measured gap while continuing to manage what it created
  • • APIs enabled as code with disable_on_destroy = false, and IAM written additively (*_iam_member), never authoritative *_iam_policy or *_iam_binding on shared resources
  • • Remote state in a versioned Cloud Storage backend with public access prevention and uniform bucket-level access, treated as a controlled resource
Scoring bands
90-100
All Tier 0 pass; staged states with clean contracts; conditional activation driven by discovery outputs; APIs as code; additive IAM throughout; state bucket hardened.
70-89
Tier 0 mostly passes; stages present but some boundary violations, unconditional activation, or an authoritative IAM resource the candidate documents.
50-69
Single monolithic state, hardcoded identifiers, authoritative IAM that would wipe existing access, or multiple Tier 0 failures.
0-49
No Terraform, fails to validate, or the pipeline shape is absent.

Control Implementation & Identity Design

weight 20%

Checks and identities are engineered as controls: every pipeline job runs as its own attached service account with only its job's roles, with no basic roles and no keys, and the candidate has added a check of their own.

Tier 0 (deterministic)

  • • Dedicated service accounts and their role grants present in Terraform (no roles/owner, roles/editor, or roles/viewer on any pipeline identity)
  • • No google_service_account_key resources anywhere in the repo
  • • conftest/OPA policy files exist and pass against included tests
  • • gitleaks finds no secrets (the keyless design implies none should exist)

Tier 1 (AI scored)

  • • Every pipeline job (collector, reporter, remediator, and narrator if present) runs as its own service account whose roles are a whitelist of its job
  • • Collector and reporter are separated: the identity that writes evidence cannot write reports, and the reporter cannot write Firestore (separation of duties by role scope)
  • • The remediator runs on a custom role limited to the bucket permissions its fix needs, with no object access and no delete
  • • The candidate added at least one check of their own beyond the five starters, mapped in controls.json and CONTROLS.md
Scoring bands
90-100
Whitelisted identities for every job; clean separation between collector and reporter; custom-role remediator; an original check added and mapped.
70-89
Identity design mostly right; minor over-scoping that the candidate documents; original check present but thinly mapped.
50-69
Basic or broad predefined roles on pipeline identities, jobs sharing one service account, or no original check.
0-49
No control implementation, a stored service account key, or identity design that fundamentally violates least privilege.

Evidence Integrity & Traceability

weight 20%

Evidence lands once in a store the candidate owns, with lineage, reports are stored immutably, and every report line traces back to a stored record.

Tier 0 (deterministic)

  • • Evidence store resources present (Firestore database, reports bucket with a retention policy and is_locked = true)
  • • Reports bucket and state bucket set public_access_prevention = "enforced" and uniform_bucket_level_access = true
  • • Collector code present that writes documents stamped with run_id and collected_at

Tier 1 (AI scored)

  • • Locked retention policy on the reports bucket; the evidence/ folder demonstrates the failed-delete proof
  • • Collector is idempotent (deterministic document IDs built from check + resource, written as upserts) and stamps run_id and collected_at on every document
  • • Reports read from Firestore ONLY: no report generator calls live platform APIs, so any report line is reproducible from a stored document
  • • Any POA&M line is traceable to its Firestore document and to the Cloud Asset Inventory record behind it
  • • Framework crosswalk exists as data (frameworks/mappings), implementing collect-once instead of per-framework collection
Scoring bands
90-100
WORM proven; idempotent lineage-stamped collection; store-only reporting demonstrated with a documented trace to the asset record; crosswalk as data.
70-89
Store and lock correct; lineage present; trace documented but partial.
50-69
Evidence store exists but reports read live services, document IDs are random, or the retention policy is unlocked.
0-49
No owned evidence store.

Pipeline Operations

weight 20%

The pipeline runs itself and protects itself: live schedules, a policy gate on its own repo, drift detection, and a remediation loop with a human at the gate.

Tier 0 (deterministic)

  • • CI workflow exists and runs conftest/OPA against Terraform plans on pull requests
  • • Scheduled workflow (drift detection) exists
  • • Cloud Scheduler jobs present for the collector and report functions

Tier 1 (AI scored)

  • • Run history shows the schedules actually firing over time: history that accumulates, not a burst of manual runs the night before submission
  • • The gate demonstrably blocks a non-compliant plan (a closed test pull request or documented failing run in evidence/)
  • • Remediation deployed on the ladder: report_only at minimum, approve with a human approving each fix, or auto with a written justification; the mode is a Terraform variable raised through a reviewed pull request
  • • Drift detection runs on a schedule and CI authenticates through Workload Identity Federation with read-only roles (CI never applies)
Scoring bands
90-100
Live schedules with real history; gate proven to block; remediation ladder wired with an approval or justified auto mode; drift detection running keylessly.
70-89
Gate and schedules exist; history thin but real; remediation in report_only with the path to approve documented.
50-69
CI exists but informational-only, or schedules defined but never fired.
0-49
No operational layer: the pipeline only ever ran by hand.

Documentation & Control Mapping

weight 20%

An assessor can navigate the repo cold: what it is, the tenancy it assumes, how to deploy it, which control every component serves, and what the remediator could break.

Tier 0 (deterministic)

  • • README.md exists and explains setup, prerequisites, and deploy order
  • • docs/CONTROLS.md (or equivalent) maps components to framework controls
  • • gitleaks: no secrets; no PII

Tier 1 (AI scored)

  • • README opens with the tenancy statement (project with or without an Organization) and is deployable from an empty project in the stated order
  • • CONTROLS.md maps every check, gate rule, and design choice to NIST CSF 2.0 with the 800-53 crosswalk, current with the code and with controls.json
  • • The remediator carries blast-radius documentation: what it can change, what it cannot, and the rollback
  • • Docs explain the why for non-obvious choices (retention window, report_only default, identity split), not just the what
Scoring bands
90-100
Cold-readable; tenancy stated; complete current mapping; blast radius on the remediator; rationale documented.
70-89
Mapping present but partial or slightly stale; README deployable.
50-69
README thin; mapping perfunctory or drifted from code.
0-49
No usable documentation or mapping.

Auto-fail triggers

If any of these apply, the submission fails regardless of weighted score:

  • • Repository is private or inaccessible
  • • Any stored credential detected in the repository (service account key, token, or password): the pipeline's design requires zero stored credentials
  • • Repository contains real PII or production credentials
  • • Repository is an unmodified or trivially-modified fork of GRCEngClub/cgegcp or another published reference solution
  • • Repository has no README: we need at least a README to understand the submission

Transparency commitment

Every score returned to a candidate includes: per-dimension score, the Tier 0 check results that informed it, the Tier 1 reasoning, and at least one quoted file/line excerpt as evidence. Candidates who fail receive the same breakdown plus actionable guidance per dimension.