Part VI - COLLAPSE: a quantitative method for detecting systemic bias

Part VI: COLLAPSE A quantitative test for systemic bias James D. Atkinson 2025 Abstract This paper introduces a quantitative method for detecting systemic bias through the structured assessment of statistically implausible multi-domain outcome patterns. The approach formalises collapse analysis, a reproducible statistical instrument with layered operational depth: a first-stage independence check requiring only domain classification and counting; a second-stage Monte Carlo dependency model for intermediate correlation stress-testing; and a final Bayesian update that formalises posterior convergence under conservative assumptions. Where the probability of the observed configuration under the innocent model falls below actuarial materiality, the method treats the event as a quantitative evidentiary signal of systemic bias. The framework is implemented via a five-stage protocol: domain extraction, temporal clustering, conservative stochastic modelling, correlation stress-testing, and calculation of an irreconcilability threshold. The analysis explicitly distinguishes directional coherence from random variation, and formalises the evidential significance of uniform multi-domain deviation following a triggering perturbation. Statistical collapse is not presented as proof of motive; it is the quantitative rejection of the hypothesis that the observed pattern could reasonably emerge by chance. A fully synthetic case study demonstrates the method in practice. The observed fifteen-domain adverse pattern, emerging immediately after a triggering perturbation, is shown to be actuarially incompatible with any plausible model of benign system behaviour. The resulting probability collapse provides a replicable foundation for evidential inference in legal, regulatory, and governance contexts. The paper therefore contributes a generalisable statistical test for systemic bias: transparent by design, computationally lightweight at entry, and progressively rigorous across modelling layers. Unlike existing approaches – typically reliant on narrative reconstruction or comparator analysis – collapse analysis offers a formally 1 quantifiable criterion for when an innocent organisational explanation becomes probabilistically untenable. The method requires no narrative reconstruction,no subjective interpretation, and no external corroboration beyond the outcome vector itself. It operates entirely on probabilistic structure under explicitly stated assumptions, and remains formally neutral with respect to motive. Keywords: systemic bias, collapse analysis, multi-domain outcomes, dependency modelling, Monte Carlo simulation, Bayesian convergence, forensic statistics, governance systems, evidential statistics. 2 Contents 1 Introduction 6 2 Methodology 7 2.1 Contradiction matrices as temporal diagnostic instruments ...... 7 2.2 Collapse on observation: when systems cannot withstand scrutiny . 9 2.3 Domain representation and problem formalisation ........... 10 2.4 Baseline independence model ...................... 11 2.5 Directional collapse under exchangeable Gaussian dependence . . . 12 2.6 Admissible dependence and copula models ............... 16 2.7 Directional collapse under bounded dependence (distribution-free alternative) ................................... 17 2.8 Explicit exponential constants for directional collapse ......... 19 2.8.1 Gaussian-Copula collapse constant ............... 19 2.8.2 Distribution-free bounded-dependence constant ....... 20 2.8.3 Interpretive summary ....................... 20 2.9 Decision-theoretic collapse threshold .................. 21 2.10 Dependency structures and correlation stress-testing ......... 22 2.11 Bayesian collapse of the innocent hypothesis .............. 22 2.12 Bayesian update ............................... 23 2.13 Decision-theoretic interpretation and governance thresholds . . . . 24 3 Mathematical formalism 24 3.1 Probability space and domain vector ................... 24 3.2 Competing hypotheses ........................... 25 3.3 Independent Bernoulli model ....................... 25 3.4 Correlated Gaussian-Copula model .................... 26 3.5 Implausibility thresholds .......................... 27 3.6 Asymptotic interpretation ......................... 27 3.7 Bayesian update, Bayes factor, and posterior convergence ...... 28 3.8 Collapse analysis .............................. 29 4 Asymptotic properties of directional collapse 29 4.1 Exchangeable dependence and exponential decay ........... 29 4.2 Bayes factor divergence under directional collapse .......... 31 5 Limitations 32 5.1 Dependence on domain specification .................. 32 5.2 Binary representation of outcomes .................... 33 5.3 Absence of causal attribution ....................... 33 5.4 Model dependence and parameter selection .............. 33 3 5.5 Assumption of temporal clustering .................... 33 5.6 Dependence on available records ..................... 34 5.7 Interpretational boundary ......................... 34 5.8 Generalisability and recalibration ..................... 34 5.9 Summary of limitations ........................... 34 5.10 Comparator independence and misuse boundary ............ 35 6 Future Work 36 6.1 Model refinement and parameter sensitivity .............. 36 6.2 Integration with governance and behavioural data ........... 36 6.3 Expansion to multi-case and cross-sectional analysis ......... 36 6.4 Alignment with evidentiary standards .................. 37 6.5 Integration with contradiction-field modelling ............. 37 6.6 Tooling, automation, and audit integration ................ 37 6.7 Ethical and regulatory implications .................... 37 7 Strengths and contribution 38 7.1 A formal test for statistical implausibility ................ 38 7.2 Directionality and systemic inference .................. 38 7.3 Reproducibility and evidential transparency ............... 39 7.4 Position within the wider literature .................... 39 8 Conclusion 39 A How to use this method without a maths background 41 A.1 Who this is for ................................ 41 A.2 What this test does ............................. 41 A.3 What you need ................................ 41 A.4 How to run it ................................. 42 A.5 What the result means ........................... 42 A.6 What to do with it .............................. 43 A.7 Important notes ............................... 43 B Standard operating procedure for union representatives (Equality Act as an example) 43 B.1 Purpose of this appendix .......................... 43 B.2 When to use a collapse analysis ...................... 44 B.3 Suggested wording for correspondence ................. 48 B.4 Linking to section 136 Equality Act 2010 ................ 48 B.5 Requesting disclosure and justification ................. 49 B.6 Practical tips for union representatives ................. 49 B.7 Conclusion .................................. 50 4 C Guidance for legal counsel 50 C.1 Purpose of this appendix .......................... 50 C.2 What collapse analysis establishes .................... 50 C.3 Evidential role in legal reasoning ..................... 51 C.4 Recommended framing in submissions ................. 51 C.5 Interaction with other evidence ...................... 52 C.6 Judicial caution and admissibility ..................... 52 D Guidance for regulators and oversight bodies 52 D.1 Purpose of this appendix .......................... 52 D.2 Collapse as a regulatory trigger ...................... 53 D.3 Why collapse analysis is regulator-relevant ............... 53 D.4 Appropriate regulatory use cases ..................... 53 D.5 Regulatory actions supported by collapse ................ 54 D.6 Importance of restraint ........................... 54 D.7 Summary ................................... 54 E Worked example of a collapse analysis 54 E.1 Defining the domains ............................ 55 E.2 Rating before and after the trigger .................... 56 E.3 Collapse score ................................ 56 E.4 Interpretation ................................ 56 F One–page quick reference for Union Reps 57 G References and DOI links 58 H Collapse Analysis Protocol (CAP) 58 H.1 Overview ................................... 58 H.2 Input schema ................................ 59 H.3 Validity conditions .............................. 59 H.4 Null model specification .......................... 60 H.5 Computation of evidential probability .................. 60 H.6 Bayesian evaluation ............................. 61 H.7 Interpretation thresholds ......................... 61 H.8 Conditions under which the protocol cannot be applied ........ 62 H.9 Summary ................................... 62 5 1 Introduction This paper introduces collapse analysis, a formal evidential criterion under which uniform high-dimensional adverse outcomes become statistically incompatible with stochastic innocence. It sets out a quantitative method for evaluating whether an observed pattern of system decision-making can be reconciled with ordinary, benign random fluctuation. It provides an actuarial modelling framework that identifies when the probability of such a benign explanation becomes statistically implausible. Where the null model — that adverse outcomes arise through random variation — falls below any credible evidential threshold, the resulting implausibility constitutes statistical evidence of a directional process. The underlying problem is well established across legal, regulatory, and governance domains. Findings of systemic adverse treatment or structural bias typically depend on narrative reconstruction, subjective assessment of credibility, and comparators that are often incomplete or ill-specified. These approaches lack a formal method for quantifying whether a multi-domain adverse pattern is compatible with chance. As a result, biased systems may remain undetected where the governance record is sparse, fragmented, or inconsistent. This paper addresses that evidential gap. It introduces a statistical mechanism for assessing directional drift across multiple domains of system conduct following a triggering perturbation. The method applies standard actuarial techniques: domain extraction, temporal clustering, dependency modelling, and Monte Carlo simulation. These tools are used to test whether the observed configuration of outcomes can be accommodated by an innocent stochastic process. Where the observed pattern is materially at odds with that model, the divergence is treated as statistical implausibility. The contribution is twofold. First, it provides a reproducible analytic procedure for detecting systemic bias without requiring assumptions about motive or intent. Second, it demonstrates how statistical implausibility — quantified formally and subject to correlation stress-testing — may serve as evidence that system behaviour has departed from benign fluctuation and entered the domain of systemic adverse treatment. A fully synthetic case study is included to illustrate the method in practice. The observed pattern comprises a sequence of adverse changes across fifteen governancesystem domains following a triggering perturbation. The model presented here evaluates whether this pattern is reconcilable with ordinary stochastic variation. The resulting findings demonstrate a level of statistical implausibility that, on any reasonable assumption, materially undermines the innocent explanation. 6 The framework is generalisable. Although the case study illustrates systemic adverse treatment, the methodology can be applied to any multi-domain system in which adverse outcome patterns must be distinguished from random fluctuation. The model therefore offers a structured, evidence-based mechanism for identifying directional bias and for strengthening the evidential foundations of legal, governance, and regulatory decision-making. Throughout this paper, “systemic bias” is used in a strictly statistical sense: a directional process that produces outcome patterns whose probability under the innocent model lies below any credible evidential threshold. 2 Methodology This section sets out a quantitative methodology for assessing whether a multi-domain pattern of adverse outcomes observed in a governance system can be reconciled with ordinary stochastic fluctuation, or whether it constitutes statistical evidence of a directional process. As in actuarial and forensic statistical practice, the analysis does not make findings of fact or law. Its purpose is to quantify the degree to which an observed configuration of outcomes can plausibly arise under a null model of benign system behaviour. The methodological framework rests on three principles: 1. Binary domain modelling: each governance-system domain is represented as a binary variable indicating adverse or neutral/positive movement. 2. Competing hypotheses: •𝐻0: outcomes arise from ordinary stochastic system fluctuation; •𝐻1: a directional process affects system-level decision-making. 3. Quantitative reconciliation: the task is to evaluate the magnitude of ℙ(𝐷∣𝐻0), where 𝐷denotes the observed outcome pattern. 2.1 Contradiction matrices as temporal diagnostic instruments The diagnostic framework supporting the quantitative model relies on two structured analytic instruments: the Contradiction Matrix and the Cross-Examination Matrix. The 7 Contradiction Matrix is an applied operationalisation of the contradiction construct introduced in earlier work, where contradiction is defined as structural incompatibility between asserted system rationale and observed behaviour across domains. The contradiction matrix. The contradiction matrix operationalises the contradiction construct developed in earlier work by evaluating the internal consistency of a system’s rationale–narrative–behaviour (R–N–B triple) configuration across two temporal windows. A contradiction is identified where the post-perturbation configuration cannot be reconciled with the pre-perturbation configuration under any admissible reinterpretation of rationale, narrative, or observed behaviour. Crucially, the matrix is invariant to post-hoc reframing. A system cannot resolve contradiction by altering its declared rationale, established narrative, or antecedent behaviour, because all three components are constrained by the system’s own contemporaneous record in the pre-perturbation window. Where no configuration of R–N–B yields coherence across time, the contradiction is structural rather than interpretive. By comparing the preand post-perturbation profiles, the matrix identifies where stable, coherent system practice gives way to contradiction or adverse divergence following a triggering perturbation. In doing so, it converts qualitative inconsistency into a structured pattern of directional deviation. Where contradictions cluster across domains that are independent or only weakly correlated, the plausibility of a benign explanation decreases sharply. Domains exhibiting structural contradiction define admissible adverse outcomes for the stochastic model. For each such domain, the variable 𝑋𝑖is fixed at 1, representing an outcome that cannot be reconciled with benign system operation under the null hypothesis. The cross-examination matrix. The Cross-Examination Matrix evaluates whether a system’s post-perturbation decisions remain defensible under structured scrutiny of their own documentary record. Each major decision is subjected to a three-stage analytic test: 1. Procedural compliance — whether the decision accords with established policy, statutory requirements, and recorded governance norms; 2. Substantive justification — whether a contemporaneous, documented, and internally coherent rationale exists; 3. Outcome trajectory — whether the decision’s effects align with the system’s established practice and precedent. 8 A decision is considered admissible only if all three criteria are satisfied on the basis of contemporaneous evidence. Failure at any stage indicates that the decision cannot be coherently defended within the system’s own operational and governance framework. Each outcome is recorded separately for the preand post-perturbation windows. Joint diagnostic role with the Contradiction Matrix. The Cross-Examination Matrix and the Contradiction Matrix operate jointly to evaluate internal system coherence across time. The Contradiction Matrix tests whether the system’s rationale–narrative–behaviour configuration remains reconcilable between the preand post-perturbation windows under any admissible reinterpretation of rationale, procedure, or outcome. The CrossExamination Matrix tests whether the post-perturbation configuration can be sustained under documentary scrutiny without reliance on post-hoc reconstruction. Together, the matrices eliminate interpretive escape routes. A system cannot resolve incoherence by altering its stated rationale, invoking exceptional narratives, or recharacterising outcomes, because each component is historically constrained by the preperturbation record. Where no configuration of rationale, norm, and behaviour yields coherence across time, the contradiction is structural rather than narrative. Diagnostic value. The diagnostic matrices provide an evidential lens orthogonal to the probabilistic model. While the stochastic framework evaluates whether the observed post-perturbation configuration could plausibly arise under an innocent random process, the matrices test whether the system itself supplies a coherent non-stochastic explanation grounded in its own record. Convergence between structural incoherence and statistical implausibility produces a multi-modal evidential signature of directional system behaviour. The combined output of the matrices defines the admissible domain classification for the stochastic model. Let 𝒞denote the set of domains in which no coherent reconciliation exists between preand post-perturbation records. For each 𝑖∈𝒞, the model fixes 𝑋𝑖=1, representing an adverse outcome requiring probabilistic reconciliation under the null hypothesis 𝐻0. In this way, the diagnostic instruments constrain the domain state space without imposing assumptions about motive, intent, or legal liability. 2.2 Collapse on observation: when systems cannot withstand scrutiny Certain organisational systems exhibit pseudo-stability: they appear coherent not because they are internally sound, but because scrutiny has not yet been applied. Their stability is not derived from internal integrity, but from the absence of scrutiny. 9 the point. It shows how exponential decay works: steady, uniform steps accumulate into overwhelming implausibility. For this reason, the horizontal (logarithmic) axis is the correct scale on which to assess evidential weight. Figure 2: Linear-scale waterfall representation of probability decay under domain independence. The same sequence of 16 independent domains shown in Figure 1is plotted here on a linear probability axis. Each domain multiplies the total probability by the same factor (×0.5), but on a linear scale the early domains dominate visually while later domains rapidly compress toward zero and become indistinguishable. This representation illustrates how linear probability scales obscure the cumulative evidential contribution of successive independent domains, creating the false impression that later domains matter less. The contrast with the logarithmic waterfall highlights why log probability is the appropriate primary scale for interpreting multiplicative processes and exponential collapse. 2.6 Admissible dependence and copula models The asymptotic results in Section 4establish that exponentially fast collapse of ℙ(𝐷∣ 𝐻0)is generic under admissible exchangeable dependence. It is therefore necessary to clarify how the Gaussian–copula models used in the simulation framework relate to this admissible class. Let 𝑍=(𝑍1,…,𝑍𝑛)∼𝒩(0,Σ𝜌), (2.21) 16 with equicorrelation structure (Σ𝜌)𝑖𝑗 =𝜌for 𝑖≠𝑗, and define 𝑋𝑖=𝕀(𝑍𝑖>𝜏), ℙ(𝑍𝑖>𝜏)=𝑝. (2.22) This construction induces an exchangeable Bernoulli vector with positive dependence for all 𝜌∈[0,1). For any fixed 𝜌<1, the resulting mixing distribution over conditional success probabilities has no point mass at Θ=1. Consequently, the induced model satisfies the admissibility condition in Definition 4.1. Hence, for all copula models with 𝜌<1, the probability of a fully directional configuration obeys the exponential upper bound ℙ(𝐷∣𝐻0)≤e−𝛼𝑛 (2.23) for some 𝛼>0depending on (𝑝,𝜌). Only in the singular limit 𝜌 → 1does the model exit the admissible class, at which point the interpretation of 𝐻0itself collapses: near-perfect correlation corresponds to a synchronised global driver rather than benign decentralised fluctuation. In this sense, strong copula dependence does not rescue the innocent model but instead negates it by construction. 2.7 Directional collapse under bounded dependence (distribution-free alternative) We now present a fully distribution-free version of the directional collapse result. This formulation avoids Gaussian copula constructions entirely and relies only on bounded marginal probabilities and bounded pairwise dependence. It therefore applies to arbitrary binary outcome models satisfying weak correlation constraints. Theorem 2.2 (Directional collapse Under Bounded Dependence).Let 𝑋1,…,𝑋𝑛∈{0,1} be binary random variables satisfying: 1. ℙ(𝑋𝑖=1)=𝑝∈(0,1)for all 𝑖, 2. Corr(𝑋𝑖,𝑋𝑗)≤𝜌<1for all 𝑖≠𝑗. Then there exists a constant 𝛾=𝛾(𝑝,𝜌)>0such that ℙ(𝑋1=1,…,𝑋𝑛=1)≤e−𝛾𝑛.(2.24) 17 Thus, even under bounded positive dependence, the probability of full directional collapse decays exponentially in the number of (Janson, 2004). Proof. Let 𝑆𝑛=𝑛 􏾜 𝑖=1𝑋𝑖.(2.25) Then 𝔼[𝑆𝑛] = 𝑛𝑝. The event of full directional collapse corresponds to the extreme deviation 𝑆𝑛=𝑛. Under the bounded pairwise correlation assumption, the variance satisfies: Var(𝑆𝑛)= 𝑛 􏾜 𝑖=1Var(𝑋𝑖)+2􏾜 𝑖<𝑗 Cov(𝑋𝑖,𝑋𝑗)≤𝑛𝑝(1−𝑝)+𝑛(𝑛−1)𝜌𝑝(1−𝑝). (2.26) Hence, Var(𝑆𝑛)≤𝐶𝜌𝑛2, 𝐶𝜌∶=𝜌𝑝(1−𝑝). (2.27) By a Bernstein–Chernoff bound for weakly dependent Bernoulli variables (Hoeffding, 1963), ℙ(𝑆𝑛=𝑛)=ℙ(𝑆𝑛−𝔼[𝑆𝑛]≥𝑛(1−𝑝))≤exp ⎛ ⎜ ⎜ ⎜ ⎜ ⎜ ⎝−𝑛2(1−𝑝)2 2Var(𝑆𝑛)+2 3𝑛(1−𝑝) ⎞ ⎟ ⎟ ⎟ ⎟ ⎟ ⎠.(2.28) Substituting the above variance bound yields ℙ(𝑆𝑛=𝑛)≤exp(−𝛾𝑛), (2.29) for some explicit constant 𝛾=𝛾(𝑝,𝜌)>0. This establishes exponential decay. Remark 4 (Model Robustness).Unlike the Gaussian-copula formulation, this result does not assume: • any latent linear dependence mechanism, • any specific joint distribution, • or any exchangeability beyond bounded correlation. It therefore establishes that directional collapse is generically incompatible with any benign stochastic model obeying even weak decentralised dependence. 18 Corollary 2 (Equivalence of collapse Regimes).The Gaussian-copula collapse regime of Section 2.4 and the bounded-dependence regime treated here yield the same asymptotic conclusion: unless domains are driven by a near-deterministic common factor, the innocent hypothesis collapses exponentially in 𝑛. 2.8 Explicit exponential constants for directional collapse This section makes the exponential collapse bounds derived earlier fully explicit by providing closed-form expressions for the decay constants under both the Gaussiancopula model and the distribution-free bounded-dependence model. These constants permit direct numerical evaluation in empirical applications. 2.8.1 Gaussian-Copula collapse constant Under the Gaussian copula construction of Section 2.4, recall that the binary variables are generated by 𝑋𝑖=𝕀(𝑍𝑖>𝜏), 𝑍∼𝒩(0,Σ𝜌), 𝜏=Φ−1(1−𝑝), (2.30) where Σ𝜌has equicorrelation 𝜌∈[0,1)and Φdenotes the standard normal distribution function. From Theorem 2.4, the full-collapse probability satisfies ℙ(𝐷∣𝐻0)≤exp(−𝛼(𝑝,𝜌)𝑛), (2.31) with explicit decay constant 𝛼(𝑝,𝜌)=(Φ−1(1−𝑝))2(1−𝜌) 2.(2.32) This bound is strictly positive for all 𝑝∈(0,1)and all admissible correlations 𝜌<1. In particular, the exponential rate is controlled jointly by the marginal adversity rate 𝑝and the residual independence (1−𝜌)of the system. Numerical illustration. If 𝑝=0.20, then Φ−1(0.8)≈0.84, so 𝛼(0.20,𝜌)≈0.353(1−𝜌). (2.33) 19 For 𝜌=0.5, this yields 𝛼≈0.176, implying a collapse probability on the order of 𝑒−0.176𝑛. 2.8.2 Distribution-free bounded-dependence constant Under the distribution-free framework of Theorem 2.5, assume: ℙ(𝑋𝑖=1)=𝑝, Corr(𝑋𝑖,𝑋𝑗)≤𝜌<1. (2.34) Let 𝑆𝑛=∑𝑛 𝑖=1𝑋𝑖. Then Var(𝑆𝑛)≤𝑛𝑝(1−𝑝)(1+(𝑛−1)𝜌). (2.35) Applying the Bernstein–Chernoff inequality yields: ℙ(𝑆𝑛=𝑛)≤exp ⎛ ⎜ ⎜ ⎜ ⎜ ⎜ ⎝−𝑛2(1−𝑝)2 2𝑛𝑝(1−𝑝)(1+(𝑛−1)𝜌)+2 3𝑛(1−𝑝) ⎞ ⎟ ⎟ ⎟ ⎟ ⎟ ⎠.(2.36) Hence, the explicit exponential decay constant is: 𝛾(𝑝,𝜌,𝑛)= 𝑛(1−𝑝)2 2𝑝(1−𝑝)(1+(𝑛−1)𝜌)+2 3(1−𝑝).(2.37) For each fixed 𝑝∈(0,1)and 𝜌<1, this yields a strictly positive exponential decay rate in 𝑛, confirming that directional collapse remains exponentially implausible under any bounded-dependence innocent model. 2.8.3 Interpretive summary Both model regimes yield explicit exponential decay: ℙ(𝐷∣𝐻0)≤⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩exp(−𝛼(𝑝,𝜌)𝑛), Gaussian copula, exp(−𝛾(𝑝,𝜌,𝑛)), distribution-free bounded dependence.(2.38) In both cases, the innocent hypothesis collapses exponentially in the number of aligned adverse domains unless the system is driven by a near-deterministic global synchronising mechanism. These constants are empirically estimable and can be precomputed across a range of plausible (𝑝,𝜌)values, allowing collapse analysis to be deployed rapidly in applied settings without bespoke simulation. 20 2.9 Decision-theoretic collapse threshold We now reframe the Bayesian collapse of 𝐻0as a formal decision-theoretic result. Under this framework, continued reliance on the innocent hypothesis becomes strictly suboptimal once its posterior support falls below a specified evidential threshold. Theorem 2.3 (Decision-Theoretic collapse of 𝐻0).Let 𝐻0denote the hypothesis of stochastic innocence, and 𝐻1a directional alternative. Suppose the observed pattern 𝐷satisfies: ℙ(𝐷∣𝐻0)≤𝜀, ℙ(𝐷∣𝐻1)≈1, (2.39) for some small threshold 𝜀 ≪ 0.01representing actuarial materiality. Let ℙ(𝐻0),ℙ(𝐻1) be prior beliefs, and let 𝐿(⋅,⋅)denote a decision-theoretic loss function over hypotheses and actions. Then for any decision rule 𝛿satisfying: 𝛿(𝐷)=“retain 𝐻0”,(2.40) there exists a competing decision rule 𝛿′such that: 𝔼[𝐿(𝛿′(𝐷),𝐻)]<𝔼[𝐿(𝛿(𝐷),𝐻)], (2.41) under all loss functions in which: • false negatives under 𝐻1incur strictly greater loss than false positives under 𝐻0, • and posterior odds satisfy ℙ(𝐻1∣𝐷)≫ℙ(𝐻0∣𝐷). In particular, if 𝜀≪10−6, then 𝛿(𝐷)=“retain 𝐻0” becomes irrational under all admissible loss-sensitive models. Proof. From Bayes’ rule: ℙ(𝐻1∣𝐷) ℙ(𝐻0∣𝐷)=ℙ(𝐷∣𝐻1) ℙ(𝐷∣𝐻0)⋅ℙ(𝐻1) ℙ(𝐻0).(2.42) Under the assumptions, this ratio exceeds 1/𝜀, which is very large. Thus, posterior belief overwhelmingly favours 𝐻1. Under any loss function 𝐿with asymmetric penalty (i.e., 𝐿(retain 𝐻0,𝐻1)≫𝐿(retain 𝐻1,𝐻0)), expected loss is minimised by choosing the action corresponding to 𝐻1. Therefore, retaining 𝐻0becomes strictly suboptimal. 21 Corollary 3 (Governance Trigger Threshold).Let 𝜀denote the actuarial materiality threshold (e.g., 𝜀=10−6). Then for any configuration 𝐷satisfying: ℙ(𝐷∣𝐻0)≤𝜀, (2.43) the rational institutional response is to treat 𝐻1as evidentially dominant. Inaction becomes the higher-risk decision. Remark 5. This result provides a formal bridge between probabilistic inference and governance decision-making. It defines the evidential point at which benign explanation is no longer rationally defensible under any risk-aware institutional model. It suffices for this conclusion that the expected loss incurred by retaining 𝐻0when 𝐻1 is true strictly exceeds the loss incurred by retaining 𝐻1when 𝐻0is true. No further assumption regarding the magnitude of this asymmetry is required. 2.10 Dependency structures and correlation stress-testing Let 𝜌denote pairwise correlation between domains under 𝐻0. Correlated outcomes are generated via a Gaussian copula construction: 𝑋𝑖=𝕀(𝑍𝑖>𝜏), 𝑍∼𝒩(0,Σ𝜌). (2.44) Monte Carlo simulations are used to evaluate whether admissible correlation structures rescue 𝐻0. Results show that only near-perfect synchronisation (𝜌≈0.9) yields material probabilities of full-domain collapse. Such coupling is itself evidence of a unifying causal mechanism and therefore inconsistent with decentralised benign fluctuation. Thus, the only dependency structures capable of numerically sustaining 𝐻0simultaneously invalidate it conceptually. 2.11 Bayesian collapse of the innocent hypothesis We now formalise the posterior collapse of the innocent model under observed directional outcome patterns. Theorem 2.4 (Bayesian collapse).Let 𝑋1,…,𝑋𝑛be binary domain outcomes with observed configuration 𝐷={𝑋1=⋯=𝑋𝑛=1}. Let ℙ(𝐷∣𝐻0)and ℙ(𝐷∣𝐻1)denote the probabilities of this configuration under: 22 •𝐻0: ordinary stochastic fluctuation (e.g., copula or bounded dependence model), •𝐻1: a directional mechanism inducing full-domain adverse movement. Suppose: ℙ(𝐷∣𝐻1)≈1, ℙ(𝐷∣𝐻0)≤e−𝛼𝑛.(2.45) Then for any prior odds ℙ(𝐻1) ℙ(𝐻0)=𝑟>0, the posterior odds satisfy: ℙ(𝐻1∣𝐷) ℙ(𝐻0∣𝐷)≥𝑟⋅e𝛼𝑛.(2.46) In particular, when 𝑛is sufficiently large that ℙ(𝐷∣𝐻0)≪10−6, the posterior overwhelmingly supports 𝐻1under any reasonable prior. Proof. Bayes’ Rule gives: ℙ(𝐻1∣𝐷) ℙ(𝐻0∣𝐷)=ℙ(𝐷∣𝐻1) ℙ(𝐷∣𝐻0)⋅ℙ(𝐻1) ℙ(𝐻0).(2.47) By assumption, ℙ(𝐷∣𝐻1)≈1and ℙ(𝐷∣𝐻0)≤e−𝛼𝑛, so: ℙ(𝐻1∣𝐷) ℙ(𝐻0∣𝐷)≥𝑟⋅e𝛼𝑛.(2.48) Remark 6. This result provides a quantitative evidential threshold: when the observed pattern 𝐷renders the null probability exponentially small, the directional hypothesis becomes not merely plausible but overwhelmingly probable in posterior belief. 2.12 Bayesian update Under 𝐻1, we have: ℙ(𝐷∣𝐻1)≈1. (2.49) Under 𝐻0, across admissible (𝑝,𝜌): ℙ(𝐷∣𝐻0)≪10−6.(2.50) Thus, the Bayes factor satisfies: ℙ(𝐷∣𝐻1) ℙ(𝐷∣𝐻0)≫106,(2.51) 23 implying overwhelming posterior support for 𝐻1under all reasonable priors. 2.13 Decision-theoretic interpretation and governance thresholds When ℙ(𝐷∣𝐻0)≤𝜀, 𝜀≪0.01, (2.52) continued reliance on 𝐻0becomes statistically irrational under any loss-sensitive system governance model. At this threshold, inaction itself becomes the higher-risk decision, and remedial review, regulatory scrutiny, or system intervention become the rationally required responses. This does not constitute a finding of liability. It formalises the point at which stochastic innocence ceases to be probabilistically (Jaynes, 2003). 3 Mathematical formalism This section formalises the probabilistic framework underlying the methodology. The aim is to specify the null and alternative models, the structure of the random variables, and the evidential criterion under which statistical implausibility is treated as evidence of a directional process. 3.1 Probability space and domain vector Let (Ω,ℱ,ℙ) be a probability space representing all possible configurations of outcomes across a fixed set of system domains. Let 𝑛 ∈ ℕdenote the number of domains under consideration. For each 𝑖 ∈ {1,…,𝑛}, define a Bernoulli random variable: 𝑋𝑖∶Ω→{0,1}, 𝑋𝑖(𝜔)=⎧ ⎪ ⎪ ⎨ ⎪ ⎪ ⎩1, adverse movement in domain 𝑖, 𝑏𝑒𝑔𝑖𝑛𝑒𝑞𝑢𝑎𝑡𝑖𝑜𝑛2𝑝𝑡]0, neutral or positive movement in domain 𝑖. (3.1) The joint outcome vector is: 𝑋=(𝑋1,…,𝑋𝑛)∈{0,1}𝑛.(3.2) 24 Let 𝐷⊆{0,1}𝑛denote an observed configuration. In the fully directional case, 𝐷={(1,1,…,1)}={𝑋1=1,…,𝑋𝑛=1}. (3.3) The total number of adverse domains is: 𝐾= 𝑛 􏾜 𝑖=1𝑋𝑖,(3.4) and the fully directional configuration satisfies 𝐾=𝑛. 3.2 Competing hypotheses Two competing explanations for the observed configuration are defined: 𝐻0∶outcomes arise from ordinary stochastic system fluctuation,(3.5) 𝐻1∶a directional process induces correlated adverse outcomes.(3.6) Under 𝐻0, adverse outcomes reflect benign stochastic variation; under 𝐻1, they arise from a persistent directional mechanism. The fundamental quantity of interest is: ℙ(𝐷∣𝐻0)=ℙ(𝑋∈𝐷∣𝐻0). (3.7) 3.3 Independent Bernoulli model As a deliberately conservative baseline, assume under 𝐻0: • each domain has the same marginal probability 𝑝∈(0,1)of an adverse outcome; and • all domains are independent. Formally, 𝑋𝑖∣𝐻0∼Bernoulli(𝑝), 𝑖=1,…,𝑛. (3.8) Then: ℙ(𝐷∣𝐻0)=𝑝𝑛.(3.9) 25 Let 𝜋1∈(0,1)denote the prior probability assigned to 𝐻1and 𝜋0=1−𝜋1. The posterior probability of 𝐻1given 𝐷𝑛is: ℙ(𝐻1∣𝐷𝑛)= 𝐵10(𝑛)𝜋1 𝐵10(𝑛)𝜋1+𝜋0.(4.11) Corollary 5 (Posterior convergence under directional collapse).Under the assumptions of Theorem 4.3, for any fixed prior 𝜋1>0, ℙ(𝐻1∣𝐷𝑛)⟶1 as 𝑛→∞. (4.12) The convergence is exponentially fast in 𝑛. Proof. Substituting the bound from Theorem 4.3, ℙ(𝐻1∣𝐷𝑛)= 𝐵10(𝑛)𝜋1 𝐵10(𝑛)𝜋1+𝜋0=1− 𝜋0 𝐵10(𝑛)𝜋1+𝜋0≥1− 𝜋0 𝑐e𝛼𝑛𝜋1.(4.13) The right-hand side tends to 1as 𝑛→∞. 5 Limitations This methodology offers a structured quantitative mechanism for assessing whether an observed multi-domain pattern of adverse system outcomes can be reconciled with ordinary stochastic fluctuation. Its purpose is evidential: to evaluate the plausibility of the innocent explanation. It does not purport to replace normative legal analysis, qualitative assessment, or findings of fact. The principal limitations are as follows. 5.1 Dependence on domain specification The analytical framework requires that the relevant governance and system domains be correctly identified and discretised. Although the illustrative multi-domain structure used in the synthetic demonstration reflects standard organisational taxonomies (procedure, governance, remit, communication, decision consistency, and related functions), alternative domain decompositions may be possible. Misclassification or omission of material domains could influence the measured degree of uniformity. This lim32 itation is inherent to all multi-domain modelling: conclusions are conditional on the chosen representation. 5.2 Binary representation of outcomes The model treats each domain as a binary indicator of adverse versus neutral movement. This simplification allows for formal probability modelling, but it does not capture fine-grained or continuous shifts in behaviour. A richer ordinal or continuous framework may enhance sensitivity, but would require additional assumptions regarding threshold selection and distributional form. The binary model therefore trades granularity for interpretability and evidential stability. 5.3 Absence of causal attribution The methodology does not identify motive or intent. It evaluates whether the observed pattern can be reconciled with a stochastic model of benign system behaviour. Where ℙ(𝐷∣𝐻0)is negligibly small, the innocent explanation becomes statistically untenable; however, the model does not specify the alternative cause. In legal or regulatory settings, this aligns with evidential burdens rather than causal findings. The rejection of 𝐻0is not, in itself, a finding of discriminatory purpose. 5.4 Model dependence and parameter selection Probability estimates depend on selected marginal adverse probabilities 𝑝and correlation levels 𝜌. Although the analysis deliberately adopts conservative, respondentfavourable assumptions and stress-tests across a wide parameter grid, the results remain conditional on the modelling choices. In environments with highly unusual decisionmaking structures or artificially synchronised processes, different parameter values may be appropriate. This limitation is mitigated by presenting full probability surfaces and by demonstrating that fully directional configurations lie in regions treated in actuarial practice as effectively zero across all plausible parameterisations. 5.5 Assumption of temporal clustering The methodology assumes that temporal proximity matters: a directional pattern of adverse changes emerging immediately after a triggering perturbation carries stronger 33 evidential weight than a dispersed pattern. In practice, temporal clustering is often observed in systemic adverse outcome cases, but the method does not model delay mechanisms explicitly. If adverse outcomes occur over extended periods or across multiple system cycles, further modelling refinements may be required. 5.6 Dependence on available records The quantitative analysis evaluates patterns revealed by the documentary record. Where contemporaneous governance documentation is missing, incomplete, or internally contradictory, the model measures outcomes rather than stated rationale. In applied settings, the absence of documentation may itself have evidential significance; however, the method does not infer intent from missing records. It operates strictly on the observable outcome configuration. 5.7 Interpretational boundary The framework identifies statistical implausibility, not liability. A pattern lying far outside the stochastic distribution generated by 𝐻0constitutes an evidential signal of a directional process, but the ultimate legal, regulatory, or organisational conclusion requires normative judgment, comparator analysis (where applicable), and fact-finding beyond the scope of the mathematical model. 5.8 Generalisability and recalibration Although the methodology is generalisable across system contexts, specific numerical results obtained under any particular parameterisation are not transferable without recalibration. Different governance structures, decision hierarchies, or intervention mechanisms may require modified domain definitions or dependency structures. The conceptual framework generalises; the specific probability values do not. 5.9 Summary of limitations These limitations do not undermine the central contribution of the methodology: the formal evaluation of whether an observed outcome pattern can reasonably be reconciled with innocent stochastic fluctuation. They delineate the interpretive boundary. The model quantifies plausibility; it does not determine motive. 34 Crucially, none of these limitations operate in a direction that would materially increase ℙ(𝐷 ∣ 𝐻0)under reasonable perturbations of assumptions. Modifications to domain definition, marginal probabilities, or correlation structures either leave the inference unchanged or decrease the probability of the observed pattern further. The limitations therefore bound, but do not threaten, the evidential conclusion. 5.10 Comparator independence and misuse boundary Comparator independence. Traditional systemic adverse outcome analysis often relies on external comparators to infer differential treatment. The present framework does not require comparators, because it operates on the internal directional coherence of the system outcome vector itself. When a multi-domain pattern exhibits uniform adverse movement and occupies a region of the probability space treated as actuarially impossible under 𝐻0, the evidential force arises endogenously from the system’s own outcome structure. In such regimes, the absence of comparators does not weaken the inference. The mathematical implausibility of the configuration already constitutes a sufficient evidential anomaly. Comparators may supplement interpretation but are not necessary for rejection of stochastic innocence once directional collapse is established. Misuse boundary. To prevent misapplication, collapse analysis is admissible only under strict structural conditions. The method must not be applied when: • outcomes are temporally dispersed without a coherent triggering boundary; • domain signs are mixed rather than uniformly directional; • domain definitions are altered post hoc to engineer coherence; • adverse and neutral outcomes cannot be meaningfully distinguished; • observed variation is dominated by known exogenous systemic shocks. Outside these conditions, the framework loses diagnostic validity and must not be used to support evidential inference. These constraints define the epistemic boundary of the method. When they are not satisfied, the protocol is inapplicable by definition. This boundary ensures that the framework functions as a protective diagnostic instrument rather than as a general-purpose rhetorical weapon. 35 6 Future Work The quantitative method set out in this paper establishes a replicable framework for identifying statistical implausibility as an evidentiary indicator of directional system behaviour. Several avenues for future development follow directly from this initial formulation. 6.1 Model refinement and parameter sensitivity The present Monte Carlo approach relies on conservative and respondent-favourable parameter choices. Further work should examine the behaviour of the model under alternative dependency structures, non-exchangeable correlations, and heterogeneous marginal probabilities. Systematic sensitivity analyses across these dimensions would strengthen the evidential robustness of the method and enable its application to systems with more complex internal dynamics. 6.2 Integration with governance and behavioural data The model treats each domain as a measurable outcome of system behaviour. Extending the framework to incorporate temporal covariates, governance inputs, escalation pathways, and behavioural markers would permit a more granular assessment of directional drift. Such integration would also allow regulators and audit bodies to evaluate whether certain governance configurations exhibit elevated susceptibility to statistically implausible outcome patterns. 6.3 Expansion to multi-case and cross-sectional analysis While the method is demonstrated here through a single synthetic illustration, the underlying framework is generalisable. Future work could apply the methodology to cohorts of cases, identifying whether directional system behaviour manifests through consistent statistical signatures across organisations, sectors, or decision-making environments. This would support the construction of population-level baselines and facilitate comparative risk benchmarking. 36 6.4 Alignment with evidentiary standards A further line of inquiry is the formal mapping of statistical implausibility to legal and regulatory evidentiary thresholds. Although this paper does not purport to make legal findings, the relationship between statistical rejection of the innocent model and evidential inference warrants systematic examination. Establishing principled criteria for when implausibility becomes probative could improve consistency across tribunals, regulators, and enforcement bodies. 6.5 Integration with contradiction-field modelling Observed directional drift across multiple domains suggests that system behaviour may be governed by an underlying contradiction field, in which pressure, omission, and procedural decay act as structured sources of systemic deformation. Future work should pursue formal unification of the statistical collapse framework with contradiction-field dynamics, enabling directional behaviour to be modelled not merely as an outcome pattern but as a geometric failure mode with quantifiable tension, curvature, and coherence thresholds. This integration would permit joint inference over probabilistic collapse and structural stress, linking outcome implausibility to field-theoretic indicators of systemic breakdown. 6.6 Tooling, automation, and audit integration The method lends itself naturally to automation. Development of software tools capable of ingesting system records, detecting directional divergence, and quantifying statistical implausibility would provide regulators, auditors, and practitioners with a practical mechanism for early detection. Such tools could support proactive governance, reduce litigation exposure, and identify systemic risk before harm crystallises. 6.7 Ethical and regulatory implications Finally, deployment of this framework across operational and regulatory contexts raises important questions regarding the ethical use of statistical evidence, the governance of algorithmic audit tools, and the potential for misuse. Future work should examine safeguards, transparency requirements, and interpretative standards to ensure that statistical implausibility functions as a protective diagnostic instrument rather than a punitive or rhetorical one. 37 7 Strengths and contribution Existing quantitative approaches to adverse outcomes typically address individual events, external comparators, or population-level disparities. They do not provide a formal statistical test for evaluating the plausibility of a single high-dimensional directional outcome pattern. This paper therefore fills a methodological gap: it operationalises evidential implausibility as a measurable quantity within a stochastic framework, providing an instrument not currently present in systemic adverse outcome analysis, organisational modelling, or actuarial practice. This paper makes three core contributions to the quantitative analysis of directional system behaviour. Each contribution is deliberately confined to statistical inference and evidential reasoning, and does not purport to make findings of fact, motive, or law. The strengths of the method arise from its formal structure, its evidential clarity, and its reproducibility across contexts. 7.1 A formal test for statistical implausibility The first contribution is a structured framework for assessing whether an observed pattern of outcomes is reconcilable with innocent stochastic fluctuation. By modelling marginal probabilities, dependency structures, and correlation effects explicitly, the method provides a quantitative means of determining when the null model collapses. This transforms what would otherwise remain a qualitative concern into a formally testable evidential question: is the observed pattern statistically compatible with random variation? 7.2 Directionality and systemic inference The second contribution lies in the treatment of directionality. Traditional approaches often assess adverse events in isolation. By contrast, the present framework treats a sequence of uniformly adverse outcomes as a single directional vector. This captures systemic drift rather than episodic discrepancy, and permits inference at the level of system behaviour rather than individual incidents. The framework therefore identifies when a pattern, taken as a whole, exhibits structural properties that cannot reasonably be attributed to stochastic variation. 38 7.3 Reproducibility and evidential transparency The third contribution is the model’s reproducibility. All assumptions, parameters, and computational steps are explicit, capable of independent replication, and amenable to external scrutiny. This ensures transparency in both academic analysis and applied evidential settings. The framework can be deployed across diverse system contexts, provided that input probabilities are specified conservatively and the dependency structure is articulated clearly. The result is a generalisable and methodologically coherent instrument for detecting evidential implausibility. 7.4 Position within the wider literature Finally, the paper addresses a persistent gap in current practice. While qualitative analyses of systemic adverse treatment and structural bias are well established, quantitative methods for evaluating the plausibility of innocent stochastic explanations remain limited. This framework provides a structured alternative grounded in actuarial modelling and statistical inference. Its contribution is not to assert motive, but to demonstrate when stochastic explanations cannot be sustained on probabilistic grounds. In evidential contexts, this distinction is decisive. 8 Conclusion This paper has introduced a quantitative method for identifying patterns of system behaviour that are statistically incompatible with ordinary stochastic fluctuation. By modelling the expected distribution of outcomes under conservative assumptions and comparing those distributions to observed directional configurations, the framework provides a structured mechanism for determining when a benign stochastic explanation can no longer be sustained on probabilistic grounds. The concept of statistical implausibility is used not to assert motive, but to establish that certain outcome configurations cannot be reconciled with randomness without invoking assumptions that are themselves unrealistic, untestable, or internally inconsistent. The contribution of the paper is twofold. First, it formalises a replicable, evidencebased framework that integrates actuarial reasoning, dependency modelling, and structured contradiction analysis into a single evaluative instrument. Second, it demonstrates how this framework can function as an evidentiary diagnostic in contexts where explanatory records are fragmented, incomplete, or structurally absent. In such settings, quantitative analysis of outcome patterns provides a defensible means of assess39 ing whether prevailing explanations remain probabilistically viable. The findings do not replace legal, regulatory, or organisational judgment. Rather, they delineate the probabilistic boundaries within which such judgment can be exercised without departing from statistical coherence. When observed behaviour occupies regions of the outcome space that actuarial models treat as effectively zero-probability events, continued reliance on stochastic innocence becomes analytically indefensible. At that point, the evidential burden shifts from randomness to structure. Statistical implausibility therefore functions not as proof but as a threshold. Once crossed, the domain of credible innocent explanations collapses, and analytical focus necessarily reorients toward directional system behaviour. The framework presented here provides a disciplined mechanism for identifying that threshold and for quantifying the strength of the resulting evidential signal. Accordingly, the methodological contribution of this paper is a formally defined, replicable statistical test for evidential implausibility in system outcomes: a diagnostic criterion specifying when the null model of stochastic fluctuation is no longer a coherent probabilistic hypothesis under observed directional alignment. This criterion unifies stochastic modelling, dependency analysis, and Bayesian updating into a single framework suitable for academic, regulatory, and applied governance contexts. In this sense, directional collapse functions as a phase transition in evidential geometry: beyond the threshold, stochastic innocence ceases to exist as a viable statistical state. The collapse test may be freely applied to any governance system, workplace, or institutional record where outcomes can be temporally grouped, domain-classified, and counted. 40 Appendix note Throughout this appendix, ‘effectively zero’ is used in the actuarial sense: probabilities below material decision thresholds, not literal impossibility. Where domain ratings are materially contested, results should be presented as indicative rather than determinative. A How to use this method without a maths background A.1 Who this is for This guide is for people who are facing serious problems with decisions made about them — especially in employment, legal, regulatory, or institutional settings — and who believe the system’s decisions may not be fair. You do not need to know mathematics to use this guide. A.2 What this test does The collapse test checks whether a set of negative outcomes can be explained by random chance, or whether the pattern is so unlikely that something directional — like bias — must be causing it. It gives you a number: if that number is extremely small, then the system’s explanation (”it’s just how things happened”) is statistically not believable. A.3 What you need To run the collapse test, you need three things: 1. A list of the different domains where decisions were made about you — for example: training, funding, access, promotion, email response, complaints handling, supervision, etc. 2. For each domain, decide if the change after a key event (e.g. a complaint, a disclosure, or conflict) was: 41 The collapse analysis provides a structured way to describe those facts: • the member’s working conditions were stable for a significant period; • a trigger event occurred, such as disability disclosure or a protected act; • shortly afterwards, many domains deteriorated simultaneously; • there is no comparable pattern in other staff; and • the overall pattern is statistically incompatible with normal fluctuation. Together, these points form a basis for saying: On the balance of probabilities, and in the absence of a coherent alternative explanation, this pattern should be treated as caused rather than accidental. The legal burden now falls on the employer to provide an explanation that fits both the record and the statistical reality. B.3 Suggested wording for correspondence This section provides wording that can be used or adapted in union letters, grievance documents, or submissions to an employment tribunal. Suggested paragraph: collapse analysis of the member’s working conditions. The union has carried out a structured collapse analysis across the member’s working conditions. We divided the job into a set of practical domains (including management, workload, information access, meetings, training, health and safety, and adjustments) and compared conditions before and after the trigger event. Across these domains, we observed a sudden and simultaneous deterioration tightly linked to [disability disclosure / the grievance / the protected act]. The number of domains that became adverse, and the timing of those changes, is not compatible with normal organisational fluctuation. On any reasonable actuarial model, such a pattern is treated as effectively impossible by chance. B.4 Linking to section 136 Equality Act 2010 Suggested paragraph: 48 Burden of proof and inference of unlawful treatment. Under section 136 of the Equality Act 2010, once there are facts from which the Tribunal could conclude that discrimination has occurred in the absence of an adequate explanation, the burden of proof shifts to the employer. The collapse analysis demonstrates a multi-domain pattern of post–event detriment that is statistically incompatible with an innocent explanation. In these circumstances, and given the absence of contemporaneous justification, the union’s position is that the burden has shifted. It is now for the employer to provide a coherent, evidence based explanation that is consistent with both the documentary record and the observed pattern. B.5 Requesting disclosure and justification Suggested paragraph: Request for contemporaneous evidence and rationale. In light of the collapse analysis, we request full disclosure of all contemporaneous documents relevant to the decisions that altered the member’s working conditions. This includes, but is not limited to: business cases, approval chains, equality impact assessments, risk assessments, occupational health reports, consultation notes, internal correspondence, and governance records. Where rationale, consultation, or oversight records are missing, this should be stated explicitly. In the absence of such material, the union will treat the observed pattern as unexplained and will invite the Tribunal to draw an adverse inference. B.6 Practical tips for union representatives • Work with the member to define domains that reflect their real experience. The tool is there to support them, not replace their story. • Keep the language factual, neutral, and restrained. The strength of the analysis lies in its mathematical structure, not in strong adjectives. • Use the collapse analysis alongside, not instead of, conventional evidence such as emails, policies, and witness accounts. • Remember that the aim is to show that the innocent explanation has been exhausted. Once that point is reached, the legal burden shifts. 49 B.7 Conclusion The collapse analysis is a practical, union friendly way to describe patterns of detriment that many members experience but struggle to prove. It does not require advanced mathematics. It requires careful description of the job, honest rating of conditions before and after a trigger event, and a willingness to treat highly unlikely patterns as what they are: signs that something has gone structurally wrong. Used well, this method can strengthen union advocacy, support members in challenging situations, and help tribunals and employers to see clearly what is often hidden in plain sight. C Guidance for legal counsel C.1 Purpose of this appendix This appendix is written for legal practitioners. It explains how collapse analysis should be understood, framed, and bounded when used in legal submissions, pleadings, advice, or judicial reasoning. Collapse analysis is not expert evidence of liability. It is a quantitative evidential instrument for identifying when an innocent explanation has become probabilistically untenable under conservative assumptions. Its function is threshold-setting, not adjudication. C.2 What collapse analysis establishes Collapse analysis establishes the following proposition: The observed multi-domain pattern of adverse outcomes lies so far into the tail of any admissible innocent stochastic model that continued reliance on chance as an explanation is statistically irrational. It does not: • determine motive or intent; 50 • allocate legal responsibility; • make findings of fact; • displace judicial assessment of credibility or evidence. C.3 Evidential role in legal reasoning Collapse analysis functions analogously to: • statistical pattern evidence; • actuarial risk assessment; • structured inferential burden-shifting tests. It is most naturally deployed: • at the prima facie stage; • in burden-of-proof analysis; • when assessing whether an explanation is adequate, not merely possible. In Equality Act claims, this aligns directly with section 136: once facts exist from which a tribunal could infer unlawful treatment in the absence of an adequate explanation, the burden shifts. Collapse analysis provides a structured way to describe those facts. C.4 Recommended framing in submissions Collapse analysis should be framed conservatively. Appropriate formulations include: • “statistically incompatible with ordinary fluctuation”; • “beyond actuarial materiality”; • “effectively impossible by chance for decision-making purposes”. Absolute claims (“proof”, “certainty”, “conclusive”) should be avoided. The strength of the method lies in its probabilistic restraint. 51 C.5 Interaction with other evidence Collapse analysis should be used alongside, not instead of: • contemporaneous documentation; • witness evidence; • policy and procedural material. Where documentary rationale or governance records are absent, the collapse itself becomes evidentially salient rather than weakened. C.6 Judicial caution and admissibility The method: • relies on standard probabilistic principles; • adopts least-favourable assumptions; • is transparent and reproducible. Courts need not accept its conclusions wholesale for it to be probative. Partial acceptance suffices to support the inference that chance explanations have been exhausted. D Guidance for regulators and oversight bodies D.1 Purpose of this appendix This appendix is written for regulators, inspectorates, ombudsmen, and oversight bodies. It explains how collapse analysis may be used as a screening and escalation tool. The method answers a single question: Is this pattern plausibly noise, or does it warrant formal investigation? 52 D.2 Collapse as a regulatory trigger Collapse analysis does not determine breach, fault, or sanction. It provides a decision-theoretic trigger: when the probability of an innocent explanation falls below material thresholds, inaction becomes the higher-risk regulatory choice. At this point, escalation is proportionate rather than prejudicial. D.3 Why collapse analysis is regulator-relevant Regulatory systems frequently encounter: • incomplete or fragmented records; • diffuse harm across multiple decisions; • procedural compliance masking substantive failure; • plausible deniability through decentralisation. Collapse analysis operates: • without requiring comparators; • without relying on intent; • under missing or imperfect documentation. This makes it suitable for early-stage triage and prioritisation. D.4 Appropriate regulatory use cases Collapse analysis is particularly informative where: • deterioration follows a clear triggering event; • effects span multiple operational domains; • no single decision is decisive in isolation; • the pattern is visible only in aggregate. 53 Examples include retaliation following disclosures, systemic withdrawal of support after protected acts, or coordinated procedural hardening. D.5 Regulatory actions supported by collapse Where collapse is observed, proportionate responses include: • formal information requests; • audit or inspection initiation; • requirement for contemporaneous justification; • escalation to investigatory or enforcement teams. Collapse analysis justifies asking the question. It does not prejudge the answer. D.6 Importance of restraint The method is intentionally conservative. If collapse is observed despite generous assumptions, allowance for correlation, and absence of intent claims, regulatory concern is justified even under high evidential thresholds. D.7 Summary Collapse analysis provides a structured, quantitative basis for deciding when ordinary explanations have been exhausted. Beyond this point, continued reliance on stochastic innocence becomes probabilistically indefensible, and oversight action becomes the rational response. E Worked example of a collapse analysis This appendix provides a worked, anonymised example showing how a union representative might apply the collapse method in practice. 54 E.1 Defining the domains For this example, we use sixteen domains: 1. Line management quality 2. Workload and task allocation 3. Access to information 4. Inclusion in meetings 5. Training availability 6. Communication from management 7. Support from colleagues 8. Autonomy 9. Clarity of role 10. Performance management fairness 11. Progression opportunities 12. Health and safety impact 13. Adjustments process 14. Governance and oversight 15. Decision visibility 16. Team environment 55 E.2 Rating before and after the trigger Domain Before After Line management +1 -1 Workload 0 -1 Information access +1 -1 Meetings +1 -1 Training 0 -1 Communication +1 -1 Colleague support +1 0 Autonomy +1 -1 Role clarity +1 -1 Performance management 0 -1 Progression +1 -1 Health and safety 0 -1 Adjustments +1 -1 Governance +1 -1 Decision visibility 0 -1 Team environment +1 -1 E.3 Collapse score Of the sixteen domains: 𝐾=15domains became adverse.(E.1) E.4 Interpretation Even under generous assumptions (high correlation, skewed priors, or clustered risk), the probability of fifteen adverse shifts arising through normal organisational fluctuation is effectively zero. This is a classical collapse event: • tightly time–locked to a trigger, • spanning many independent domains, • with no offsetting improvements. 56 Under s. 136 EqA, this establishes facts from which the Tribunal may conclude unlawful treatment unless the employer provides a coherent, contemporaneous explanation. The burden therefore shifts. F One–page quick reference for Union Reps What is collapse? A structured diagnostic for identifying whether a worker’s pattern of detriment can realistically be explained by chance. How to apply it: 1. Define 12–20 operational domains. 2. Rate each domain before and after the trigger (+1, 0, -1). 3. Count the number of adverse shifts (the collapse score). 4. If 𝐾≥12, especially after a trigger event, the pattern is statistically incompatible with innocent fluctuation. Legal link (s.136 EqA): If the collapse score is high, and no contemporaneous justification exists, the tribunal may draw an inference of unlawful treatment. The burden shifts to the employer. Use this wording in letters: The pattern of detriment is statistically incompatible with normal organisational noise. Under section 136 Equality Act 2010, the burden now shifts to the employer to provide a coherent explanation consistent with the evidence. When collapse is strongest: • sudden deterioration across many domains; • tightly bound to a protected act or disclosure; • no documented rationale or governance trail; • organisation cannot identify comparable patterns in others. 57