Skip to content

Governance as Meaning-Alignment

Governance failures are not primarily technical failures. They are reference frame translation failures. When the meaning encoded in governance language diverges from operational reality, the result is compliance theatre — governance that looks correct but governs nothing.

This page explains why governance language drifts, how to detect it, and what the plugin provides to prevent it.

The Core Problem

Governance language carries different meanings in different reference frames. Consider "ensure human oversight of AI-generated code":

Frame Interpretation
Regulatory A qualified human evaluates whether the AI decision is appropriate before it takes effect
Engineering A human clicks "approve" on a pull request that contains AI-generated code
AI system A boolean gate prevents merge without at least one approval

All three frames can be satisfied simultaneously while governance fails semantically. The approval happens, but the oversight is absent. The PR was approved, the audit trail exists, the gate passed — and nobody checked whether the code was correct.

This is not a hypothetical. It is the default state of most governance constraints. The language sounds precise. The implementations diverge. The audits confirm the divergence. The risk persists.

Semantic Drift

Governance language drifts through five predictable stages:

Stage What happens Example
1. Coinage A term is introduced with specific meaning "Meaningful human oversight" means active evaluation by a qualified human
2. Adoption without frame The term enters governance documents without its original context Policy says "require meaningful human oversight" without defining "meaningful"
3. Implementation from a different frame Engineers implement from their own reference frame Team adds a PR approval gate — technically "human oversight"
4. Audit from yet another frame Compliance verifies from the institutional frame Audit confirms approval records exist — "oversight requirement met"
5. Crisis The gap becomes visible through real-world harm AI-generated code introduces a vulnerability despite "oversight" passing at every stage

Most governance constraints in the wild are at Stage 2 or 3. The language has been adopted, the implementation has diverged, and nobody has noticed because the audits check the wrong frame.

The plugin's /governance-audit command detects which stage each constraint is in. Constraints at Stage 3 or higher need immediate attention.

Governance Debt

Governance debt is the accumulation of governance mechanisms whose language no longer corresponds to the risks they address. It is the fourth form of debt, alongside technical, cognitive, and intent debt.

The four debts reinforce each other in a vicious cycle:

  1. Governance debt → intent debt: governance language obscures the purpose of controls
  2. Intent debt → cognitive debt: teams cannot form accurate mental models of what the governance actually requires
  3. Cognitive debt → technical debt: developers implement controls they do not understand
  4. Technical debt → governance debt: the gap between governance language and operational reality widens further

Unlike technical debt, which becomes visible in hours or days, governance debt manifests over months or years — but when it does, it manifests as institutional crisis, not a failed build.

The Three-Frame Translation Problem

Every governance constraint operates at the intersection of three reference frames:

Interface Failure mode Detection
Human ↔ AI "It did what I said, not what I meant" Tests, verification, falsification
Human ↔ Institution "The regulation does not match how we actually work" Governance specification exercises
AI ↔ Institution "The system is compliant but not safe" Adversarial compliance testing

The most dangerous failures occur when all three frames are misaligned simultaneously: the human understands the risk, the governance language does not capture it, and the AI system can demonstrate compliance while the risk persists.

The plugin's /governance-constrain command includes a three-frame alignment check that surfaces these divergences before a constraint is written to HARNESS.md.

Falsifiable Governance

A governance constraint must answer three questions to be falsifiable:

  1. What do you verify? — the specific observable condition
  2. What counts as evidence? — what artefacts demonstrate compliance
  3. What happens on failure? — the response when verification fails

If a constraint cannot answer all three, it is governance language pretending to be a constraint. It belongs in a policy document, not in HARNESS.md.

Before (governance language):

Human oversight is required for all AI-generated code.

After (falsifiable governance constraint):

Every PR containing AI-generated code must have at least one review that includes a substantive comment on design intent and confirmation that tests cover the changed behaviour. Verified by agent review of PR comments. Evidence: PR review record. Failure action: PR cannot merge until substantive review is added.

The difference is not length. The difference is that the second version can be checked, can fail, and specifies what happens when it does.

How the Plugin Helps

The plugin provides three governance commands that address different stages of the governance lifecycle:

Command Purpose When to use
/governance-constrain Translate governance language into a falsifiable constraint with three-frame alignment When encoding a new governance requirement
/governance-audit Detect semantic drift, inventory governance debt, check frame alignment Quarterly, or when governance feels stale
/governance-health Quick pulse check on governance metrics and trends Between quarterly audits

The governance-auditor agent owns the deep investigation work. Existing agents (harness-enforcer, harness-gc, assessor) gain lightweight governance awareness through the governance skills.

A Stop hook nudges /governance-audit when governance-related files change during a session and the last audit is stale.

For a hands-on walkthrough, start with the Governance for Your Harness tutorial. For specific tasks, see the how-to guides: