additive-bias-defense

The default answer to "should we add this?" is no. The burden of proof is on the addition.

Additive Bias Defense

Note (#444): Frontmatter declares provides.guidance, not provides.contract. The scrutiny questions are consumed by partner skills (pensive:code-refinement, conserve:unbloat, imbue:scope-guard) that voluntarily embed them; no validator or hook in leyline enforces them directly. If a future contributor adds enforcement, restore the contract label and link the validator path here.

When NOT To Use

• Removing bloat that already landed (use conserve:bloat-detector)
• Scoring whether a feature is worth building (use imbue:scope-guard)

The Problem

LLMs are additive by nature. They reinvent wheels, add unnecessary complexity, hallucinate issues and modify tests to justify them, and deviate from priorities. This contract provides a systemic defense.

The Scrutiny Questions

Applied to every proposed addition (code, files, abstractions, error handling, configuration):

1. Priority alignment: Is this a deviation from the current priority?
2. Criticality: Is it critical to implement at this juncture?
3. Simplicity: Does a simpler or more elegant solution exist?
4. Evidence: What evidence proves this is needed (not assumed)?
5. Consequence: What breaks if we do not add this?

If the proposer cannot answer questions 4 and 5 with concrete evidence, the addition is unjustified.

Anti-Pattern Detection

Pattern	Signal	Challenge
Wheel reinvention	New utility/helper overlapping existing code	"Does X already do this?"
Hallucinated issues	Fix for a bug with no reproduction evidence	"Show the failing test before the fix"
Test manipulation	Test changed to match behavior rather than spec	"Did the spec change, or did you change the test?"
Complexity creep	Abstraction introduced for single use case	"Is this the 3rd use, or the 1st?"
Priority deviation	Work not traceable to current task/spec	"Which requirement does this serve?"
Gold plating	Error handling or flexibility beyond need	"What breaks without this?"

Burden of Proof Verdict

After applying scrutiny questions and anti-pattern detection, produce a verdict:

Verdict	Meaning	Action
justified	Evidence supports the addition	Proceed
needs_evidence	Plausible but unproven	Provide evidence or remove
unjustified	No evidence, likely bias	Remove or justify

Integration Contract

Review-oriented skills MUST consult this contract by:

1. Applying the 5 scrutiny questions to each addition
2. Scanning for the 6 anti-patterns
3. Producing a burden-of-proof verdict
4. Including the verdict in their output

Consuming Skills

Skill	Integration Point
attune:war-room	Prosecution Counsel role uses scrutiny questions
sanctum:pr-review	Every added file/function challenged
pensive:code-refinement	Refactors pass "3rd use" test
conserve:unbloat	Findings feed removal candidates
attune:mission-orchestrator	Plan sections scanned before user review
imbue:justify	Scrutiny questions extend audit protocol

Related Skills

• imbue:karpathy-principles - "Simplicity First" and "Surgical Changes" principles invoke this contract from a higher-level four-principle synthesis
• See docs/quality-gates.md#skill-level-quality-gate-composition for the full gate-skill federation graph

The Subtraction Principle

Rely less on AI and initial lines of thinking. Challenge yourself to think of a more elegant implementation or a simpler solution.

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

Exit Criteria

• All 5 scrutiny questions applied to every proposed addition; questions 4 (evidence) and 5 (consequence) answered with concrete evidence, not assumptions
• All 6 anti-patterns scanned (wheel reinvention, hallucinated issues, test manipulation, complexity creep, priority deviation, gold plating); any match named with the pattern label
• A burden-of-proof verdict produced for each addition: justified, needs_evidence, or unjustified; the verdict appears in the consuming skill's output
• Subtraction principle applied: at least one check for "could this be achieved by removing code instead?" before accepting any net-positive line-count change