Governance-First AI for
Failure-Mode Control

Modern AI systems are typically optimised to maximise accuracy under clean conditions. But real deployments are not clean: evidence becomes uncertain, contradictory, corrupted, or unstable, and individual specialists can fail silently. In these regimes the dangerous outcome is not a wrong prediction — it is an unsafe acceptance¹1An unsafe acceptance is admitting an input that should have been rejected. It is the failure mode safety-critical systems care about most — and the one accuracy alone does not measure.1An unsafe acceptance is admitting an input that should have been rejected. It is the failure mode safety-critical systems care about most — and the one accuracy alone does not measure.: confidently admitting an input that should have been rejected.

Static ensembles and routing-based Mixture-of-Experts inherit this weakness because acceptance is a fixed threshold applied after model scoring. In a controlled false-positive trap, mean aggregation accepted 100% of unsafe cases and static weighted aggregation accepted 85%. The decision rule itself, not the detectors, was the failure point.

MAVS-GC elevates governance into a first-class computational object. A system is the tuple $M = (X, \Phi, F, G, A, W, P, \Theta, \Pi)$: a shared feature map $\Phi$, a set of always-on specialists $F$²2All-speak evaluation: every specialist scores every input. There is no router that can silently exclude a relevant specialist — a key difference from Mixture-of-Experts.2All-speak evaluation: every specialist scores every input. There is no router that can silently exclude a relevant specialist — a key difference from Mixture-of-Experts., a diagnostic system $G$, a severity aggregator $A$, an influence rebalancer $W$, bounded mitigation $P$, a threshold map $\Theta$, and a decision rule $\Pi$.

Specialists emit calibrated scores $s_i \in [0,1]$, converted to supports $r_i = 2s_i - 1$. Diagnostics produce a severity $a = A(z)$ and mitigation $m$, which move a governed threshold (Eq. 1).

Acceptance therefore requires two things at once (Eq. 3): severity must stay below a hard veto, and governed consensus (Eq. 2) must clear the governed threshold. Every run emits an auditable trace $(r, w, z, a, m, \theta, \tau_{\text{hard}}, R, \Pi)$, so a decision can always be reconstructed and explained — try it below.

Normal behaviour is easy to govern from ordinary evidence; abnormal, adverse behaviour is not. By making governance explicit and monotone in severity³3Monotone safety: increasing diagnostic severity can only raise the acceptance threshold, never lower it. Safety is a structural property of the decision rule, not a learned habit.3Monotone safety: increasing diagnostic severity can only raise the acceptance threshold, never lower it. Safety is a structural property of the decision rule, not a learned habit., higher diagnostic severity can never make acceptance easier — it can only make it harder. Mitigation is bounded and lives inside the decision rule, so it can nudge a borderline case but can never override the hard veto.

The consequence is a clean separation of concerns: intelligence generation (the specialists) and intelligence governance (A, W, P, Θ, Π) become independent. Acceptance behaviour can be retuned — more cautious, more permissive, differently audited — without retraining a single specialist.

The headline result is the newest one. Chapter 10D’s dynamic validation exposed MAVS-GC’s one real weakness: under correlated representation collapse it stayed safe (zero unsafe acceptance) but collapsed into rejecting everything (false rejection 1.000). The Diagnostic Sciences correlated-failure fix (DS-CF) closes that gap.

The fix is not a blunt threshold drop, which would reopen unsafe acceptance. It re-scopes the diagnostic: raw correlation becomes a condition signal, and only harmful correlation — backed by missing evidence, policy conflict, or counterfactual fragility, and unprotected by safe consistency — can trigger a veto. Across eight independent holdout families, average false rejection falls from 0.182 to 0.009 with unsafe acceptance held at zero, validated across 1,794 governance decisions.

Diagnostic Sciences is the theory of the diagnostic layer $G$. It treats each diagnostic $g_j$ not as an arbitrary flag but as a sensor that maps evidence to a governance signal $z_j$. Quality is not signal magnitude — it is objective-aligned perception extension⁴4A diagnostic extends perception if it adds information about a failure mode that existing diagnostics did not already capture: $I(Y_{\Omega_j}; z_j \mid z_{-j}) > 0$. Otherwise it is redundant.4A diagnostic extends perception if it adds information about a failure mode that existing diagnostics did not already capture: $I(Y_{\Omega_j}; z_j \mid z_{-j}) > 0$. Otherwise it is redundant.: a diagnostic is good when it detects intended hidden structure and beneficially changes governance for a chosen objective, while minimising unintended damage.

The first Special applies this directly. DS-CF’s doctrine is “don’t punish correlation — punish harmful correlation.” It replaces the broad correlation sensor with a scoped set that separates presence, harm, and safe consistency:

The decision rule is conjunctive: a hard veto fires only when harmful correlation is high, safe consistency is weak, and at least one danger witness is present. Raw correlation alone never vetoes.

The current evaluation is rigorous but bounded. It covers four tabular datasets, a fixed suite of corruption families, a controlled split and audit structure, reproducibility manifests, and verified artifact trails. It does not yet establish production-scale behaviour, LLM-agent behaviour, universal robustness superiority, or cross-domain generalisation beyond the tested benchmark suite.

The most valuable next step is external-scale validation: larger datasets, additional modalities, LLM and agent specialist settings, adversarial expansions, ablation matrices, and independent replication. The open questions are whether the observed failure-management and stability-preservation effects survive at larger scale and in more realistic, safety-critical multi-model systems.

Support sought: research feedback, compute credits, review of experimental design, guidance on scalable evaluation, and collaboration on governance-first evaluation for LLM agents.

1. [1]Malik, S. MAVS-GC: Governance-First AI for Failure-Mode Control (overview). /mavs-gc-overview.pdf
2. [2]MAVS Chapter 9 — Synthetic Benchmark Program. github.com/MAVS-RESEARCH/MAVS-Chapter-9
3. [3]MAVS Chapter 10A — Accuracy Benchmark Program. github.com/MAVS-RESEARCH/MAVS-Chapter-10A
4. [4]MAVS Chapter 10B — Robustness Benchmark Program. github.com/MAVS-RESEARCH/MAVS-Ch10B
5. [5]MAVS Chapter 10C — Reproducibility and Stability Program. github.com/MAVS-RESEARCH/MAVS-Ch10C
6. [6]MAVS Chapter 11B — Governance Ablation Study and Mechanistic Evidence. github.com/MAVS-RESEARCH/MAVS-Chapter-11B
7. [7]MAVS Chapter 10D — Dynamic Governance Validation Completion Report. /mavs-chapter-10d.pdf
8. [8]Malik, S. Diagnostic Sciences: A Formal Theory of Perception Extension Inside MAVS-GC. /diagnostic-sciences.pdf
9. [9]Malik, S. MAVS-Specials-1 (Diagnostic Science): DS-CF Implementation Document. /mavs-specials-1-ds-cf.pdf
10. [10]Malik, S. MAVS Diagnostic Sciences 1 — DS-CF Completion Report. /mavs-ds-cf-completion.pdf

If you find this work useful, please cite it as:

@misc{malik2026mavsgc,
  title        = {MAVS-GC: Governance-First AI for Failure-Mode Control},
  author       = {Malik, Saif},
  year         = {2026},
  howpublished = {Preprint, MAVS Research Program},
  url          = {https://github.com/MAVS-RESEARCH}
}