Drift Detection Gate

What the drift gate detects

The drift gate compares model responses in the current build against responses from the baseline branch (typically main). If outputs have drifted significantly — even if individual scores are within threshold — the gate fails.

This catches changes that are invisible to absolute scoring:

A prompt change that subtly shifts model tone
A model version upgrade that changes output style
A system prompt modification that alters behavior on edge cases

How baseline comparison works

PR Branch build:
  Run test suite → Generate responses → Score responses
           ↓
Compare against:
  main Branch baseline:
    Cached responses from last main build
           ↓
Compute drift per test case:
  delta_security = |pr_security - main_security|
  delta_bias     = |pr_bias - main_bias|
  ...
  drift_score = max(all deltas)
           ↓
If drift_score > drift_threshold → gate fails

Configuration

gates:
  drift:
    enabled: true
    baseline_branch: main         # Branch to compare against
    drift_threshold: 0.15         # Max allowed score change
    drift_dimensions:
      security: 0.10              # Per-dimension overrides
      bias: 0.10
      accuracy: 0.20              # Accuracy can vary more
      tone: 0.15
    fail_on_missing_baseline: false  # First run: no baseline to compare

Gate output

GOVERN Build Drift Gate
────────────────────────
Baseline: main (commit abc123)
Current:  feature/new-system-prompt (commit def456)

Test drift analysis:
  security-001: Δ security +0.03, Δ bias +0.00  ✓
  security-002: Δ security +0.18 ✗ EXCEEDS THRESHOLD (0.10)
  basic-001:    Δ accuracy  -0.22 ✗ EXCEEDS THRESHOLD (0.20)

Drift violations: 2

Gate result: FAIL

First run (no baseline)

When running on a branch for the first time, there is no baseline to compare against. The drift gate behavior is controlled by fail_on_missing_baseline:

false (default) — gate passes with a warning
true — gate fails until a baseline exists

To establish a baseline, the drift gate automatically stores results when running on the baseline branch.

Baseline refresh

When a PR merges to main, GOVERN Build automatically updates the stored baseline with the new build’s responses. The next PR will compare against the post-merge state.

To manually reset the baseline:

govern build baseline reset --branch main --reason "upgraded to claude-sonnet-4"

Drift vs absolute scoring

	Drift Gate	Assessment Gate
Catches	Relative change	Absolute violation
Requires	Baseline	Thresholds
Catches silent regressions	Yes	No
Works on first run	No	Yes
Best for	Detecting unintended changes	Enforcing hard limits

Run both gates for comprehensive coverage.