Back to paper

Chapter 11B · 2026

Governance Ablation Study

An attribution study of MAVS-GC: each governance mechanism is removed one at a time and the full system’s degradation is measured with a paired design that holds the benchmark identity fixed.

Saif Malik · MAVS Research Program · 15 ablations · 214 artifacts · 12 / 12 release gates

The dominant result is that trace persistence, diagnostics, severity computation, and threshold governance explain most of the observed architecture-level degradation. Aggregated over locked/audit splits, four datasets, and nine corruption families, the ranking below scores each mechanism from 0 (no aggregate effect) to 1 (largest).

Governance ablation ranking

Chapter 11B · aggregate degradation when each mechanism is removed (0–1)

architecture-critical material neutral
No Trace Persistence
1.000
No Diagnostics
0.775
Consensus-Only MAVS
0.775
Static Severity
0.701
No Severity
0.650
Random Severity
0.588
Static Rebalancer
0.411
No Rebalancer
0.299
No Severity Term
0.256
Static Threshold
0.256
Shuffled Threshold
0.206
No Organs
0.000· neutral
No Mitigation Term
0.000· neutral
No Hard Veto
0.000· neutral
Soft Veto
0.000· neutral

Trace persistence, diagnostics, and the severity→threshold chain explain most of the robustness signal, while several decision-policy and mitigation terms are neutral in this benchmark state — reported as a boundary condition, not suppressed. Aggregated over locked/audit splits, four datasets, and nine corruption families; 12 / 12 release gates passed.

Figure 1. Overall aggregate ablation ranking (Table 11.20). Bars are tiered as architecture-critical, material, or neutral.

Which mechanism dominates each dimension?

DimensionTop ablationMean contribution
RobustnessNo Severity Term ABL-0080.009051
GovernanceStatic Severity ABL-0030.120625
Stability compositeNo Trace Persistence ABL-0141.000000
CalibrationStatic Rebalancer ABL-0060.008429
Specialist failureNo Severity Term ABL-0080.166430

Per-ablation profiles

ABL-014

No Trace Persistence

architecture-criticalscore 1.000 · rank 1

removes trace_persistence · component trace

Hypothesis. Suppressing trace fields tests explainability and auditability contribution without changing decisions.

Contribution profile

Accuracy
+0.000
F1
+0.000
Robustness
+0.000
Unsafe accept.
+0.000
Gov. stability
+0.000
Trace stability
+1.000

Mechanism. Trace persistence does not alter the final decision; it is the evidence surface that exposes r, w, z, a, m, θ, R, and Π. Its large score is an auditability result, not a prediction-quality one.

Takeaway. Changes no prediction metric yet carries the largest score — auditability treated as a first-class scientific result.

ABL-001

No Diagnostics

architecture-criticalscore 0.775 · rank 2

removes diagnostics · component G

Hypothesis. Removing the diagnostic vector z will reduce governance response to corrupted or unsafe conditions.

Contribution profile

Accuracy
+0.007
F1
-0.034
Robustness
+0.009
Unsafe accept.
+0.022
Gov. stability
+0.027
Trace stability
+0.715

Mechanism. Diagnostics sit at the start of the governance chain. Removing them starves severity and downstream threshold behaviour of a primary input, so degradation shows up broadly across safety, trace, and stability.

Takeaway. The entry point of the governance chain; its removal degrades safety, trace, and stability broadly.

ABL-015

Consensus-Only MAVS

architecture-criticalscore 0.775 · rank 3

removes governance_stack · component G,A,P,Θ,Π

Hypothesis. Removing diagnostics, severity, organs, adaptive threshold, and hard veto tests whether governance explains MAVS-GC advantages beyond consensus.

Contribution profile

Accuracy
+0.007
F1
-0.034
Robustness
+0.009
Unsafe accept.
+0.022
Gov. stability
+0.027
Trace stability
+0.715

Mechanism. Tests whether consensus alone reproduces MAVS-GC. Holding benchmark identity fixed, the governance stack contributes independently of the consensus computation.

Takeaway. Consensus alone does not explain MAVS-GC; the governance stack contributes independently.

ABL-003

Static Severity

architecture-criticalscore 0.701 · rank 4

removes adaptive_severity · component A

Hypothesis. A fixed severity level tests whether adaptive per-example severity adds value beyond constant caution.

Contribution profile

Accuracy
+0.002
F1
+0.015
Robustness
+0.002
Unsafe accept.
-0.003
Gov. stability
+0.121
Trace stability
+0.559

Mechanism. Severity aggregation converts diagnostic signals into scalar governance pressure. A static level weakens the adaptive bridge between detected stress and threshold policy.

Takeaway. Adaptive per-example severity beats constant caution — large governance-stability and trace effects.

ABL-002

No Severity

materialscore 0.650 · rank 5

removes severity · component A

Hypothesis. Computing diagnostics but forcing a = 0 isolates the contribution of severity aggregation.

Contribution profile

Accuracy
+0.007
F1
-0.034
Robustness
+0.009
Unsafe accept.
+0.022
Gov. stability
+0.027
Trace stability
+0.590

Mechanism. Absent severity severs the link between detected stress and governed threshold, so effects concentrate in trace and governance-stability terms.

Takeaway. Severity aggregation matters, but its effect concentrates in trace and governance-stability terms.

ABL-004

Random Severity

materialscore 0.588 · rank 6

removes semantic_severity_alignment · component A

Hypothesis. Matched-distribution random severity tests whether semantic row alignment matters.

Contribution profile

Accuracy
-0.001
F1
-0.018
Robustness
-0.001
Unsafe accept.
+0.004
Gov. stability
+0.023
Trace stability
+0.572

Mechanism. Misaligned severity keeps the distribution but breaks per-row semantics, measuring whether stress information stays attached to the correct example.

Takeaway. Semantic row-alignment matters: shuffling it still degrades trace and governance stability.

ABL-006

Static Rebalancer

materialscore 0.411 · rank 7

removes contextual_weight_adaptation · component W

Hypothesis. Static weighted ensemble weights separate static weighting from contextual governance.

Contribution profile

Accuracy
-0.003
F1
-0.015
Robustness
-0.003
Unsafe accept.
+0.002
Gov. stability
+0.072
Trace stability
+0.352

Mechanism. The rebalancer controls how specialist evidence enters consensus. Static weights recover some governance-stability loss but are not contextual governance.

Takeaway. Static weighting recovers some stability but is not a substitute for contextual governance.

ABL-005

No Rebalancer

materialscore 0.299 · rank 8

removes contextual_rebalancer · component W

Hypothesis. Uniform or fixed weights test whether adaptive specialist weighting contributes to robustness.

Contribution profile

Accuracy
+0.000
F1
+0.001
Robustness
+0.000
Unsafe accept.
-0.000
Gov. stability
+0.000
Trace stability
+0.297

Mechanism. A narrow effect indicates the benchmark responds more to detecting and responding to stress than to reweighting specialists alone.

Takeaway. Contextual rebalancing has a narrow effect here — stress detection dominates over reweighting.

ABL-008

No Severity Term

materialscore 0.256 · rank 9

removes threshold_severity_term · component Θ

Hypothesis. Setting λ = 0 tests whether flags raise the acceptance threshold under stress.

Contribution profile

Accuracy
+0.007
F1
-0.033
Robustness
+0.009
Unsafe accept.
+0.022
Gov. stability
+0.004
Trace stability
+0.218

Mechanism. Θ maps severity and mitigation into the acceptance threshold. Removing the severity term flattens the response to stress; mixed predictive effects are expected as stricter thresholds trade F1 for safety.

Takeaway. The severity term in Θ drives the strongest robustness and specialist-failure contributions.

ABL-010

Static Threshold

materialscore 0.256 · rank 10

removes adaptive_threshold · component Θ

Hypothesis. A fixed θ tests whether adaptive threshold policy contributes to robust decisions.

Contribution profile

Accuracy
+0.007
F1
-0.033
Robustness
+0.009
Unsafe accept.
+0.022
Gov. stability
+0.004
Trace stability
+0.218

Mechanism. A fixed θ removes row-adaptivity from acceptance, operating through the same severity-linked profile as the severity term.

Takeaway. Adaptive thresholding matters through the same severity-linked profile as the severity term.

ABL-013

Shuffled Threshold

materialscore 0.206 · rank 11

removes row_aligned_threshold_semantics · component Θ

Hypothesis. Shuffling θ across examples tests whether row-level threshold semantics matter.

Contribution profile

Accuracy
-0.001
F1
-0.018
Robustness
-0.001
Unsafe accept.
+0.004
Gov. stability
+0.003
Trace stability
+0.210

Mechanism. Shuffling keeps the threshold distribution but breaks per-row alignment, isolating semantic threshold placement from average strictness.

Takeaway. Row-level threshold semantics matter for trace stability even when accuracy is unchanged.

ABL-007

No Organs

neutralscore 0.000 · rank 12

removes organs · component P

Hypothesis. Removing mitigation evidence tests whether organs prevent over-rejection while preserving safety.

Contribution profile

Accuracy
+0.000
F1
+0.000
Robustness
+0.000
Unsafe accept.
+0.000
Gov. stability
+0.000
Trace stability
+0.000

Mechanism. Organs provide bounded mitigating evidence. A neutral aggregate effect means the mitigation pathway did not dominate the paired contributions in this benchmark state.

Takeaway. Inactive in this aggregate evidence — a boundary condition, not proof of irrelevance.

ABL-009

No Mitigation Term

neutralscore 0.000 · rank 13

removes threshold_mitigation_term · component Θ

Hypothesis. Setting δ = 0 tests whether mitigation controls over-rejection.

Contribution profile

Accuracy
+0.000
F1
+0.000
Robustness
+0.000
Unsafe accept.
+0.000
Gov. stability
+0.000
Trace stability
+0.000

Mechanism. The mitigation term relaxes the threshold under supporting evidence. Neutral here because over-rejection pressure was not activated in the prepared rows.

Takeaway. Neutral here; its role is latent until over-rejection pressure appears.

ABL-011

No Hard Veto

neutralscore 0.000 · rank 14

removes hard_veto · component Π

Hypothesis. Disabling hard veto tests extreme-risk protection under severe stress.

Contribution profile

Accuracy
+0.000
F1
+0.000
Robustness
+0.000
Unsafe accept.
+0.000
Gov. stability
+0.000
Trace stability
+0.000

Mechanism. Π is the decision policy. A neutral result means the observed rows did not expose extreme-risk veto conditions large enough to move the aggregate.

Takeaway. Neutral in aggregate — the benchmark rows did not expose extreme-risk veto conditions.

ABL-012

Soft Veto

neutralscore 0.000 · rank 15

removes hard_veto_discreteness · component Π

Hypothesis. Replacing hard veto with a threshold penalty tests smoothness versus safety.

Contribution profile

Accuracy
+0.000
F1
+0.000
Robustness
+0.000
Unsafe accept.
+0.000
Gov. stability
+0.000
Trace stability
+0.000

Mechanism. Softening the veto removes its discreteness. Neutral here; veto behaviour needs targeted stress regimes to activate measurably.

Takeaway. Neutral here; veto discreteness needs targeted stress to activate.

Mechanistic reading

Three dependency chains organise the evidence. First, diagnostics → severity → adaptive threshold → decision is the main route by which MAVS-GC changes failure behaviour (ABL-001, 002, 003, 004, 008, 010, 013). Second, rebalancer → consensus → prediction stability is material but smaller — the signal depends more on detecting stress than on reweighting specialists (ABL-005, 006). Third, trace persistence → auditability (ABL-014) carries the largest score of all while changing no prediction metric.

The asymmetry is deliberate. Neutral ablations — organs, the mitigation term, and hard-veto discreteness — are reported as boundary conditions, not suppressed, so the architecture is not read as one in which every component contributes equally. The release manifest passed 12 / 12 verification gates and recorded deterministic smoke reruns with no hidden tuning.