Justify

The simplest change that fixes the problem is the safest change to merge. Adding code is easy. Removing the need for code is engineering.

Justify

The Additive Bias Problem

AI models are trained to be helpful, which creates a systematic bias toward adding code rather than fixing root causes:

This skill audits changes for these patterns and requires explicit justification for each.

When To Use

• After completing implementation work
• Before committing or creating PRs
• When reviewing your own changes for quality
• When scope-guard flags RED/YELLOW zone

When NOT To Use

• Before writing the code, because this audits work already done (use imbue:karpathy-principles)
• Deciding whether a feature belongs in scope (use imbue:scope-guard)

Audit Protocol

Step 1: Gather the Delta

# Determine base branch
base=$(git merge-base master HEAD 2>/dev/null \
  || git merge-base main HEAD 2>/dev/null)

# Get change statistics
git diff "$base" --stat
git diff "$base" --shortstat
git diff "$base" --diff-filter=A --name-only  # new files
git diff "$base" --diff-filter=M --name-only  # modified files
git diff "$base" --diff-filter=D --name-only  # deleted files

Step 2: Compute Additive Bias Score

Score each dimension 0-3 (0 = clean, 3 = high bias):

Line Ratio Scoring:

Aggregate Score:

bias_score = sum(signal_score * weight) / sum(weights)

Step 3: Iron Law Compliance Check

The Iron Law states: tests drive implementation, not the other way around. Check for violations:

# Find test files that were modified
git diff "$base" --name-only | rg "test_|_test\.|spec\." \
  || git diff "$base" --name-only | grep -E "test_|_test\.|spec\."

# For each modified test file, check what changed
git diff "$base" -- <test_file> | rg "^[-+].*assert|^[-+].*expect|^[-+].*should"

Violation patterns (test logic was tampered):

• Assertion values changed (expected output modified)
• Test cases removed or commented out
• @skip or @pytest.mark.skip added
• Error expectations weakened (broad exception types)
• Mock return values changed to match new behavior
• Test renamed to no longer describe original behavior

Each violation requires explicit justification:

"I changed this test assertion because the requirement changed, not because my implementation couldn't meet the original requirement."

If the requirement didn't change, the test should not change. Fix the implementation instead.

Step 4: Minimal Intervention Analysis

For each changed file, answer:

1. Was this change necessary? Could the goal be achieved without touching this file?

2. Was this the minimal change? Could fewer lines achieve the same result?

3. Did this change add or remove complexity? New functions, classes, or control flow = added complexity that needs justification.

4. Is there a subtraction-first alternative? Could removing code fix the problem instead of adding code?

Step 4.5: Invariant Impact Analysis

Changes can be minimal and still catastrophically wrong if they silently revise a load-bearing design decision. For each changed file, check whether it touches a design invariant:

What counts as an invariant:

• Architectural patterns (module boundaries, layer separation, data flow direction)
• Data structure choices (why a map vs list, why normalized vs denormalized)
• API contracts (public interfaces, protocol formats)
• Error handling strategies (fail-fast vs recovery)
• Concurrency models (single-threaded assumption, actor model, shared-nothing)

Detection heuristic:

# Check for structural changes (new modules, moved
# boundaries, changed interfaces)
git diff "$base" --name-only | rg "(interface|abstract|base|core|types|schema|model)" \
  || git diff "$base" --name-only | grep -E "(interface|abstract|base|core|types|schema|model)"

# Check for pattern-breaking changes
git diff "$base" -U5 | rg "(TODO.*refactor|HACK|WORKAROUND|XXX)" \
  || git diff "$base" -U5 | grep -E "(TODO.*refactor|HACK|WORKAROUND|XXX)"

When an invariant conflict is detected:

Do NOT silently pick a resolution. Present the three options to the human:

Add to Justification Report:

### Invariant Impact: NONE / DETECTED

[If DETECTED:]
- **Invariant**: [name the design decision]
- **Conflict**: [what change clashes with it]
- **Option chosen**: Preserve / Layer / Revise
- **Justification**: [why this option, not the others]
- **Human reviewed**: YES / NO — if NO, flag as
  requiring review before merge

Compounding risk warning: Bad invariant decisions accumulate. If this branch has multiple invariant revisions, flag the entire branch for architectural review. Each silent invariant change multiplies the probability of an unsalvageable codebase.

Step 5: Generate Justification Report

Output a structured report:

## Justification Report

**Branch**: feature/xyz
**Base**: master
**Delta**: +N/-M lines, X files changed

### Additive Bias Score: X.X (ZONE)

| Signal | Score | Detail |
|--------|-------|--------|
| Line ratio | N | +A/-D = R:1 |
| New files | N | [list] |
| Test changes | N | [list] |
| New abstractions | N | [list] |
| Workarounds | N | [list] |

### Iron Law Compliance: PASS/FAIL

[List any test logic modifications with justification]

### Change-by-Change Justification

#### file.py (+N/-M)
- **What**: [description]
- **Why**: [root cause this addresses]
- **Alternatives considered**: [what else could work]
- **Why this is minimal**: [why fewer changes won't work]

#### test_file.py (+N/-M)
- **What**: [description]
- **Justification**: [why test logic changed, if it did]
- **Iron Law status**: PASS/VIOLATION

### Risk Assessment

| Factor | Rating |
|--------|--------|
| Lines changed | LOW/MED/HIGH |
| Files touched | LOW/MED/HIGH |
| Test modifications | NONE/JUSTIFIED/VIOLATION |
| New abstractions | NONE/JUSTIFIED/UNNECESSARY |
| Overall merge risk | LOW/MED/HIGH |

### Recommendations

[List any changes that should be reconsidered,
simpler alternatives, or unnecessary additions]

Decision Weights

When evaluating competing approaches, weight these factors:

The Subtraction Test: Before accepting any change, ask: "Could I achieve this by removing code instead of adding it?" If yes, prefer the subtractive approach.

Integration with Proof of Work

Justify extends proof-of-work with change-level accountability:

• proof-of-work: "Did it work?" (evidence)
• justify: "Was this the right way?" (reasoning)

Both are required before claiming work is complete. Run proof-of-work first, then justify.

Anti-Patterns to Flag

1. Test Mutation

Changing test expectations to match broken code. Fix: Revert the test change, fix the implementation.