0. Cybernetic Scope Note

This entry is conceptual and systems-oriented.

It does not treat metrics, targets, benchmarks, KPIs, scores, models, rewards, indicators, incentives, or measurement systems as inherently failed. Systems need indicators. Measurement can reveal state. Targets can focus effort. Reward functions can coordinate action. Benchmarks can expose capability. Metrics can preserve auditability when they remain subordinate to reality.

The failure begins when the proxy becomes the object.

The issue is not measurement.

The issue is optimization pressure that breaks the relationship between the measure and the state it was meant to represent.

Goodhart Collapse occurs when a system improves the indicator while damaging, ignoring, hiding, or decoupling from the thing the indicator was meant to protect.

1. Definition

Goodhart collapse occurs when a metric, proxy, target, score, benchmark, incentive, KPI, risk model, reward function, or observable indicator becomes optimized so intensely that it stops preserving the underlying state, value, function, risk, coherence, or repair condition it was meant to represent.

The system may improve:

• scores
• ratings
• benchmarks
• risk indicators
• safety metrics
• compliance metrics
• productivity measures
• incident counts
• engagement signals
• revenue metrics
• turnaround times
• satisfaction surveys
• audit pass rates
• model rewards
• output quality scores
• restoration milestones

But the underlying state may not improve.

It may even degrade.

The core failure is:

proxy selected
optimization pressure↑
proxy improves
proxy / state fit↓
H↑

Goodhart Collapse is the cybernetic failure where the control system learns how to satisfy its signal while abandoning the reality that made the signal meaningful.

The metric remains.

The meaning decays.

2. Core Pattern

The core pattern is:

1. A system needs to evaluate a state, value, function, risk, success condition, repair condition, or coherence level.
2. It selects a proxy, metric, score, target, benchmark, or reward signal.
3. The proxy initially correlates with the real state.
4. The system begins optimizing the proxy.
5. Actors adapt behavior toward the metric.
6. The proxy becomes easier to improve than the underlying state.
7. Optimization pressure increases.
8. Proxy performance improves.
9. The relationship between proxy and state weakens.
10. The system treats proxy improvement as real improvement.
11. Hidden debt accumulates in the unmeasured state.
12. Failure appears when measured success no longer predicts actual function.

This failure mode often appears as:

the number improved, so the system improved

or:

the benchmark is higher, so capability is higher

or:

we hit the target, so the purpose was fulfilled

The restorative question is:

does improvement in this measure still imply improvement in the state?

Goodhart Collapse begins when the system stops asking that question.

3. Failure Signature

Typical signature:

metric pressure↑
proxy performance↑
state correspondence↓
gaming / adaptation↑
auditability↓
H↑
real function drift↑

Extended signature:

the score improves while the system worsens
the target is met while purpose is missed
the benchmark rises while capability narrows
the reward is maximized while value is lost
the risk model improves while risk shifts elsewhere

Common forms include:

teaching to the test
security teams reducing incident counts by reducing detection
AI models optimizing benchmark performance while losing real-world robustness
customer support optimizing ticket closure while users remain unresolved
justice systems optimizing case throughput while repair fails
platforms optimizing engagement while meaning degrades
economies optimizing GDP while local coherence declines
health systems optimizing billing or throughput while care quality suffers
restoration systems optimizing closure milestones while burden remains
organizations optimizing productivity while hidden fatigue accumulates

The defining condition is not that a metric is imperfect.

The defining condition is that the metric continues receiving authority after optimization has degraded its relationship to the underlying state.

4. Primary U-Layer Origin

Common origin layers:

• U1 — Power / Budgets: incentives, funding, status, authority, legitimacy, or survival depend on measured outcomes.
• U2 — Configuration / Boundaries: the metric boundary excludes relevant costs, affected nodes, externalities, or repair conditions.
• U3 — Execution / Runtime: actors optimize behavior toward the metric.
• U4 — Information / Truth: metric value substitutes for state truth.
• U5 — Coordination / Time: proxy-state correlation decays over time as actors adapt.
• U6 — Coherence Field: improved scores create the feeling of success or safety.
• U7 — Memory / Recurrence: the metric remains trusted because it used to work.
• U8 — Environment / Field: environmental adaptation, adversarial behavior, or scale pressure breaks proxy validity.

Common manifestation layers:

• U1 — Budgets: incentives reward the proxy.
• U3 — Execution: behavior shifts toward metric maximization.
• U4 — Truth: proxy replaces state.
• U5 — Time: correlation decays through adaptation.
• U6 — Coherence Field: measured success feels coherent.
• U7 — Memory: historical metric trust persists.

Goodhart Collapse is primarily a U4 / U1 proxy-incentive failure.

The system lets measurement authority outrun state validity.

5. Typical Development Sequence

A common development sequence is:

1. A useful metric is created.
2. The metric becomes trusted.
3. The metric is tied to reward, funding, status, safety, compliance, ranking, legitimacy, or control.
4. Actors learn the metric.
5. Optimization pressure rises.
6. The metric becomes a target.
7. Behavior adapts to improve the target.
8. Unmeasured dimensions are neglected or exploited.
9. The metric improves.
10. The underlying state diverges.
11. The system celebrates success.
12. Hidden debt appears later as reduced function, brittle capability, trust loss, exploitation, or repair burden.

The loop often looks like:

state → proxy → target → optimization → proxy-state drift → false success

Another common loop is:

metric challenged → improved score cited → underlying state unaudited → collapse deepens

Goodhart Collapse becomes self-protective because improved numbers are often used as evidence that the metric remains valid.

6. Diagnostic Markers

Diagnostic markers include:

• Metric improvement no longer predicts field improvement.
• Actors optimize for visible scores rather than real function.
• Unmeasured dimensions degrade.
• Edge cases are excluded to preserve performance.
• Benchmarks improve while robustness declines.
• Incident counts fall while risk remains or increases.
• Targets are met through burden shifting.
• Users, affected nodes, or downstream systems report mismatch with official success.
• The metric becomes difficult to question.
• Proxy performance rises after incentives change.
• Short-term scores improve while long-term debt accumulates.
• Metric definitions are adjusted to preserve success appearance.
• The system cannot explain why the measure still maps to the state.
• Good outcomes require actions that the metric does not reward.
• Repair improves when attention returns to state rather than target.

Useful diagnostics:

• Metric / Reality Fit: Tests whether the metric still corresponds to state.
• Proxy Validity: Measures whether the proxy remains valid under optimization pressure.
• Target Drift: Tracks divergence between target and purpose.
• Reward Integrity: Tests whether reward preserves value and repair.
• Goodhart Risk: Measures susceptibility to metric capture.
• Measurement Back-Action: Tests how measurement changes behavior.
• Success / State Fit: Compares measured success with actual success.
• Auditability: Determines whether metric claims can be inspected.
• Hidden Debt: Tracks unmeasured cost.
• Restoration Linkage: Tests whether metrics route to real repair.

7. Related Gates

Relevant gates include:

• Metric Gate: Fails when metric authority exceeds metric validity.
• Proxy Gate: Fails when proxy fit is not revalidated.
• Target Gate: Fails when target replaces purpose.
• Truth Gate: Fails when score substitutes for state truth.
• Auditability Gate: Fails when metric claims cannot be inspected.
• Observability Gate: Fails when state disappears behind measurement.
• Optimization Gate: Fails when pressure degrades the state being optimized.
• Restoration Gate: Fails when target achievement does not route to repair.

The first common gate failure is usually the Metric Gate.

The system lets the measure become the object of control.

8. Related Operators

Relevant operators include:

• Γ — Selection: Selects what is rewarded, amplified, counted, or optimized.
• Ψ — Observation / Interface: Receives the metric as an observable state signal.
• Au — Auditability: Determines whether proxy-state fit can be inspected.
• O — Coherence: Appears high when metrics improve.
• H — Hidden Debt: Accumulates in unmeasured dimensions.
• Λ — Compatibility: Tests whether metric remains compatible with real state.
• K — Constraint / Load: Rises where actors must satisfy targets even when they conflict with reality.
• R — Restoration Capacity: Declines when optimization does not route to repair.
• Τ — Trajectory / Time: Reveals proxy-state drift over recurrence.
• Φ — Flow / Resource Movement: Routes resources, status, funding, attention, and authority through metric values.
• BΣ — Boundary Integrity: Determines whether metric boundaries include relevant costs and affected nodes.
• G — Gain: Amplifies metric pressure.
• D — Damping: Can reduce optimization pressure or suppress signals challenging the metric.

Common operator pattern:

Γ selects metric
Φ routes reward through metric
G increases optimization pressure
Ψ reads improved score
O appears improved
Λ proxy fit declines
BΣ excludes unmeasured costs
Au fails to challenge proxy-state drift
R does not reach affected state
H accumulates
Τ reveals delayed divergence

The core operator inversion is:

metric improves → state improves

instead of:

metric improves → proxy validity audit → state verification

Goodhart Collapse turns measurement into reality substitution.

9. Related Laws and Invariants

Related Laws

• Goodhart Collapse: optimized measures lose state correspondence.
• Success Proxy Substitution: proxy success replaces real success.
• U4 Truth Substitution: measurement substitutes for truth.
• Hidden Debt Accumulation: unmeasured costs accumulate beneath target success.
• Auditability Collapse: metric claims cannot be inspected against reality.
• Observability Collapse: state disappears behind indicator.
• Measurement Back-Action: measurement changes behavior and state.
• Instrumentation Theater: metrics and dashboards display control without state contact.
• Pseudo-Coherence: improved metrics create apparent coherence.
• Runaway Optimization Trap: optimization continues after value fit collapses.

Related Invariants

• Metrics Must Remain Subordinate to State: measure must serve reality, not replace it.
• Targets Must Not Replace Purpose: goals must preserve function and meaning.
• Optimization Requires Reality Audit: pressure must trigger proxy revalidation.
• Reward Functions Must Preserve Repair Linkage: reward must not bypass affected-state repair.
• Proxy Fit Must Be Revalidated Under Pressure: optimization changes proxy validity.
• Indicators Must Remain Falsifiable: metric claims must be challengeable.
• Success Must Survive Affected-State Audit: measured success must match lived or actual state.

10. Common False Positives

Not every metric failure is Goodhart Collapse.

Common false positives include:

• Metrics used as partial indicators.
• Targets kept subordinate to broader judgment.
• Benchmarks validated against real-world performance.
• KPIs with explicit limits and counter-metrics.
• Reward functions tested for unintended behavior.
• Metrics that trigger inquiry rather than close it.
• Measurement systems that track unmeasured costs.
• Targets used temporarily without becoming authority.
• Indicators that remain falsifiable and audited.
• Measured success that consistently matches real state under pressure.

Clarifying rule:

This is not Goodhart Collapse unless a metric, proxy, target, score, benchmark, incentive, KPI, risk model, reward function, or observable indicator continues receiving success, safety, legitimacy, or coherence credit after optimization pressure has degraded its relationship to the underlying state.

11. Common False Repairs

Common false repairs include:

• adding more metrics without restoring state contact
• creating a composite score that hides proxy drift
• changing target definitions to preserve success
• adding dashboards around the same corrupted proxy
• optimizing harder after mismatch appears
• punishing actors for revealing metric failure
• treating metric gaming as individual misconduct only
• adding incentives to counter incentives without state audit
• replacing one proxy with another unvalidated proxy
• using benchmark diversity without real-world validation
• using confidence scores as truth
• rewarding repair theater because it meets milestones
• adding compliance checks that verify target achievement only
• declaring the metric fixed because performance improved

False repair often produces the loop:

metric drift exposed → metric system expanded → measured success restored → state remains divergent

Another common loop is:

target gaming found → anti-gaming metric added → actors optimize new metric → state still ignored

The repair fails because it adds proxy structure without restoring the relation between proxy and reality.

12. Restoration Direction

Restoration requires deflating metric authority, reconnecting indicators to actual state, auditing optimization effects, and restoring purpose, affected-node reality, and repair linkage.

Primary restoration direction:

deflate proxy authority,
recontact the underlying state,
audit optimization pressure,
and rebind metrics to repair

A fuller restoration path includes:

1. Name the metric or proxy. Identify the score, target, benchmark, KPI, incentive, reward, model, or indicator.
2. Name the underlying state. Clarify what the metric was supposed to represent or protect.
3. Separate measure from reality. Treat the metric as evidence, not the state itself.
4. Audit proxy validity. Test whether the metric still correlates with the underlying condition.
5. Audit optimization pressure. Identify how actors changed behavior in response to the target.
6. Find unmeasured costs. Surface ignored dimensions, burden shifts, externalities, and affected-node reality.
7. Reduce target authority. Prevent the metric from closing inquiry or carrying proof status.
8. Restore direct state contact. Use field checks, qualitative feedback, causal audits, and affected-node review.
9. Repair hidden debt. Address damage caused by proxy optimization.
10. Redesign incentives. Reward state improvement, not isolated proxy improvement.
11. Install counter-audits. Periodically test metric-state fit under pressure.
12. Keep metrics provisional. Maintain falsifiability and review as conditions change.

A valid restoration path should reduce:

proxy authority
target drift
metric gaming
unmeasured cost
success / state gap
reward distortion
hidden debt
restoration bypass
optimization pressure

Goodhart Collapse is not repaired by measuring harder.

It is repaired by restoring the authority of reality over the measure.

13. Cross-Module Links

• Cybernetics: Core feedback and control failure; the system steers by a signal that no longer maps to state.
• Diagnostics: Requires metric/reality, proxy-validity, Goodhart-risk, target-drift, and measurement-back-action diagnostics.
• Scaling: Goodhart pressure intensifies when metrics scale across larger systems.
• Security: Security metrics can be captured through incident suppression, checkbox compliance, scan-score optimization, or detection avoidance.
• AI Governance: Reward functions, benchmarks, red-team scores, refusal rates, and safety metrics can diverge from real behavior.
• Economy: Economic systems often optimize visible indicators while exporting hidden debt.
• Restoration: Repair fails when closure metrics replace actual load reduction.
• Control Systems: Control based on corrupted feedback drives state divergence.
• Interfaces: Dashboards and scores can make proxy success feel like reality.
• Coherence: Goodharted metrics produce pseudo-coherence through measured success.

14. Relationship to Parent / Child Modes

Production treatment: Standalone Entry / Canon-aligned

This mode maps upward to:

• FM-CORE-003 — Success Proxy Substitution
• FM-CORE-006 — U4 Truth Substitution
• FM-C-020 — Measurement Back-Action Loop
• FM-CORE-001 — Pseudo-Coherence
• FM-CORE-002 — Hidden Debt Accumulation

Sibling or related Cybernetics modes include:

• FM-C-002 — Instrumentation Theater
• FM-C-012 — Gain Saturation
• FM-C-019 — Adversarial Reward Hacking
• FM-C-020 — Measurement Back-Action Loop
• FM-C-021 — Parasitic Extraction
• FM-C-022 — Dominance Masquerading as Control

Related cross-family modes include:

• FM-S-007 — Feedback Gaming
• FM-SEC-006 — Metric Capture / Reward-Hacked Security
• FM-PX-031 — Charismatic Goodhart
• FM-OMD-007 — Runaway Optimization Trap
• FM-AIX-011 — Epistemic Distortion
• FM-AIX-020 — Catastrophic Overweighting
• FM-ECOX-016 — Risk Model Theater
• FM-ECOX-027 — Growth Theater
• FM-JC-M-001 — Goodhart Justice
• FM-R-006 — Repair as Compliance

Aliases preserved from source material:

• Goodhart Collapse
• Goodhart Failure
• Metric Capture
• Target Corruption
• Proxy Optimization Collapse
• KPI Capture
• Reward Proxy Collapse
• Measure-Target Collapse
• Indicator Corruption
• Metric Reality Divergence

15. Minimal Entry Version

Definition: Goodhart collapse occurs when a metric, proxy, target, score, benchmark, incentive, KPI, risk model, reward function, or observable indicator becomes optimized so intensely that it stops preserving the underlying state, value, function, risk, coherence, or repair condition it was meant to represent.

Signature:

metric pressure↑
proxy performance↑
state correspondence↓
gaming / adaptation↑
auditability↓
H↑
real function drift↑

Restoration direction:

• name the metric or proxy
• name the underlying state
• separate measure from reality
• audit proxy validity
• audit optimization pressure
• find unmeasured costs
• reduce target authority
• restore direct state contact
• repair hidden debt
• redesign incentives
• install counter-audits
• keep metrics provisional

16. Machine-Readable Summary

failure_mode:
  id: "FM-C-018"
  name: "Goodhart Collapse"
  family: "Cybernetics"
  production_treatment: "Standalone Entry / Canon-aligned"
  parent_modes:
    - "FM-CORE-003 — Success Proxy Substitution"
    - "FM-CORE-006 — U4 Truth Substitution"
    - "FM-C-020 — Measurement Back-Action Loop"
  primary_failure: "A metric, proxy, target, score, benchmark, incentive, KPI, risk model, reward function, or observable indicator continues receiving success, safety, legitimacy, or coherence credit after optimization pressure has degraded its relationship to the underlying state."
  source: "UTS — Failure Modes Registry"
  source_id: "FM-C-018"
  scope_note: "Conceptual and systems-oriented; does not treat metrics, targets, benchmarks, KPIs, scores, models, rewards, indicators, incentives, or measurement systems as inherently failed."
  aliases:
    - "Goodhart Collapse"
    - "Goodhart Failure"
    - "Metric Capture"
    - "Target Corruption"
    - "Proxy Optimization Collapse"
    - "KPI Capture"
    - "Reward Proxy Collapse"
    - "Measure-Target Collapse"
    - "Indicator Corruption"
    - "Metric Reality Divergence"
  signature:
    - "metric pressure↑"
    - "proxy performance↑"
    - "state correspondence↓"
    - "gaming / adaptation↑"
    - "auditability↓"
    - "H↑"
    - "real function drift↑"
  primary_layers:
    origin:
      - "U1 — Power / Budgets"
      - "U2 — Configuration / Boundaries"
      - "U3 — Execution / Runtime"
      - "U4 — Information / Truth"
      - "U5 — Coordination / Time"
      - "U6 — Coherence Field"
      - "U7 — Memory / Recurrence"
      - "U8 — Environment / Field"
    manifestation:
      - "U1 — Budgets"
      - "U3 — Execution"
      - "U4 — Truth"
      - "U5 — Time"
      - "U6 — Coherence Field"
      - "U7 — Memory"
  state_variables:
    - "Γ"
    - "Ψ"
    - "Au"
    - "O"
    - "H"
    - "Λ"
    - "K"
    - "R"
    - "Τ"
    - "Φ"
    - "BΣ"
    - "G"
    - "D"
  first_gate_failure: "Metric Gate"
  restoration:
    - "Metric Reality Audit"
    - "Proxy Deflation"
    - "Target Rebinding"
    - "Reward Integrity Repair"
    - "Goodhart Pressure Reduction"
    - "State Recontact"
    - "Affected-Node Outcome Audit"
    - "Optimization Boundary Repair"
    - "Restoration Linkage Recovery"