Threats and Failure Modes

[!IMPORTANT] ODS exists because software delivery has predictable failure modes — and they get worse when AI generates code at speed. This page maps each ODS module to the governance failure it prevents.

Failure Modes

# Failure Mode What Goes Wrong ODS Mitigation
01 Unstructured branch names Can’t categorize branches for CI workflows, auto-deploy, or changelog generation Branch Naming enforces <type>/<description> format
02 Commit messages lack intent Can’t determine if a change was a fix, feature, or breaking change from the commit alone Commit Message requires Conventional Commits with optional AI attribution
03 PR descriptions are inconsistent Reviewers can’t quickly find Summary, Changes, or Testing sections; AI disclosures are missing PR Description mandates structured sections including AI Disclosure
04 AI-generated code is invisible No record of which PRs contain AI-generated code, making review and liability tracking impossible AI Change Review requires disclosure protocol: L1 (AI-assisted), L2 (AI-generated, human-reviewed), L3 (AI-generated with mandatory human review)
05 CI failures are opaque Failures produce unstructured logs; can’t programmatically determine if failure is AI-related or infrastructure-related CI Failure produces machine-parseable reports with AI explanation
06 Release readiness is gut-feel “Is this release ready?” answered orally or via Slack, with no evidence trail Release Readiness requires evidence-based gates with scoring
07 Approval rules are informal Who needs to approve what is tribal knowledge, not documented policy Approval Workflow makes approval rules declarative and version-controlled
08 No rollback plan exists Production incident occurs; no documented path to revert Rollback Plan requires structured rollback plans with verification steps
09 No deployment evidence After a release, no immutable record of what was deployed, when, or whether gates were met Production Release Evidence creates auditable, timestamped bundles

AI-Specific Threats

AI coding tools introduce new failure modes that ODS directly addresses:

AI Threat Why It Matters ODS Response
AI writes code with no disclosure Team can’t distinguish human-written from AI-generated changes PR Description: mandatory AI Disclosure section; Commit Message: AI-assisted: true footer
AI changes are never reviewed High-volume AI PRs go unreviewed because “AI wrote it, it’s probably fine” AI Change Review L1/L2/L3 protocol; review records are machine-verifiable
AI introduces subtle bugs AI-generated code passes CI but contains logic errors CI Failure: structured reports flag AI-related failures; release readiness gates catch patterns
AI makes deployment decisions Without evidence gates, AI-assisted deployments are opaque Release Readiness: scored gates prevent unverified deployments
No audit trail for AI-era releases Compliance auditors ask “how do you control AI risks?” but no structured answer exists Production Release Evidence: immutable deployment bundles include AI review records

Governance Failures (Non-AI)

Traditional delivery governance also has well-known failures:

Failure Example ODS Module
Branch naming chaos 47 branches named fix, test, wip 01 Branch Naming
Meaningless commits "fix stuff", "update" as commit messages 02 Commit Message
Unreviewed changes 300-line PR merged with no description 03 PR Description
Undocumented releases “We deployed last Tuesday, I think” 09 Production Evidence
Absent rollback plans 4-hour outage because no one knew how to revert 08 Rollback Plan

Threat Model Principles

ODS’s threat model is not about security vulnerabilities. It’s about delivery governance failures:

  1. Opacity: Changes happen without structured records.
  2. Inconsistency: Every team uses different conventions.
  3. Untraceability: Can’t answer “was this reviewed?” or “what evidence existed for this release?”
  4. AI blind spots: Can’t tell if AI wrote code, reviewed it, or influenced a deployment decision.

ODS mitigates these by requiring machine-readable evidence at each governance checkpoint.

Further Reading