Epistemic Signal Ratio: A Process-Native Measure of Reasoning Quality in Decision-Making
Toward a Causal Theory of Reasoning Accountability in Governance
Accountability‑Based Universal Wisdom and Trust
Cross-Sector Pre-Decision Governance Translator
Final Manuscript · March 2026
---
ABSTRACT
Contemporary governance systems have developed sophisticated mechanisms for financial accountability, procedural compliance, and performance measurement. Yet, a persistent paradox remains: decisions that meet all formal governance standards often fail substantively. This paper argues that existing governance frameworks cannot rigorously identify decision-relevant reasoning under counterfactual criteria—they lack a systematic way to distinguish reasoning that actually influences decisions from reasoning that merely accompanies them. To address this gap, we introduce a formalized and counterfactually grounded class of metrics called process-native epistemic metrics, with the Epistemic Signal Ratio (ESR) as its first instantiation. ESR measures the proportion of documented reasoning that genuinely affects decision outcomes, defining relevance through counterfactual perturbation: an information unit is relevant if its removal changes the ranking of decision alternatives. We provide a reduced-form micro-foundation grounded in diminishing marginal cognitive returns, establish formal theorems on information overload, identifiability under panel aggregation, causal identification under exogenous documentation shocks (LATE), and groupthink amplification. We show that ESR is a partial sufficient statistic for manifest reasoning efficiency conditional on a closed information set, prove its non-equivalence to Shapley value (ordinal vs cardinal invariance), derive testable restrictions (\beta_1 > 0, \beta_2 < 0), and analyze strategic manipulation: adding noise is a dominated strategy (ESR is incentive‑compatible with respect to noise addition), though selection manipulation remains possible (ESR is not fully strategy‑proof). ESR thus belongs to the class of partially robust governance metrics. Extending beyond efficiency, we introduce Latent Signal (S_L) and Epistemic Blindspot (EB) to measure what organizations fail to consider, defining latent signal relative to ex post validation rather than pure speculation (excluding hindsight‑only constructs). The framework offers tiered implementation (Lite, Internal BCEV+, Full BCEV+), empirical test mapping, and adversarial audit protocols. ESR transforms reasoning quality from an unobservable cognitive construct into a statistically estimable and causally testable variable, contributing to measurement theory, decision theory, and governance theory.
Keywords: epistemic governance, pre-decision accountability, counterfactual relevance, bounded rationality, causal identification, epistemic blindspot, groupthink, partially robust metrics
---
1. INTRODUCTION
1.1 The Paradox of Procedural Compliance
Modern governance systems have evolved sophisticated accountability mechanisms: financial audits ensure proper resource use, procedural reviews mandate compliance, and performance evaluations measure outcomes. Yet, a persistent paradox remains: decisions that formally satisfy all three dimensions—financially compliant, procedurally correct, and meeting performance targets—frequently fail substantively. Infrastructure projects approved through rigorous feasibility studies become stranded assets. Social programs meeting output targets fail to achieve intended impact. Strategic investments vetted through due diligence result in catastrophic losses.
This paradox—formal compliance yet substantive failure—has been examined through multiple lenses. Herbert Simon (1947) introduced bounded rationality, showing that decision-makers operate under cognitive constraints. Kahneman and Tversky (1979) identified systematic biases that affect judgment. The behavioral public administration literature (Grimmelikhuijsen et al., 2017) has documented how cognitive biases interact with bureaucratic structures. Yet, despite these insights, governance frameworks remain focused on what was decided and whether procedures were followed, not on how reasoning was structured and whether it actually mattered for the decision.
1.2 The Identification Problem
We argue that existing governance systems face a fundamental identification problem with respect to decision-relevant reasoning. A governance system can identify decision-relevant reasoning if it can distinguish:
· Epistemic signal: reasoning that genuinely influences decision outcomes
· Epistemic noise: reasoning that is present but does not affect outcomes
Without a counterfactual definition of relevance—whether removing an information unit would change the ranking of alternatives—existing approaches cannot rigorously identify which reasoning matters. Subjective importance ratings, readability metrics, and information volume measures all fail to establish whether documented reasoning actually influenced the final decision. This is not to say that current systems are completely blind; rather, they cannot rigorously identify decision-relevant reasoning under counterfactual criteria.
1.3 A Formalized and Counterfactually Grounded Class of Metrics
This paper introduces a formalized and counterfactually grounded class of governance metrics: process-native epistemic metrics. Unlike outcome-based metrics (profit, accuracy) that evaluate results, or process-compliance metrics (checklists, audits) that evaluate procedural adherence, process-native metrics evaluate the quality of reasoning within the decision process itself using counterfactual criteria.
The Epistemic Signal Ratio (ESR) is the first instantiation of this class. It measures:
\text{ESR} = \frac{\text{Decision-Relevant Signal}}{\text{Total Documented Reasoning}} \quad \text{with} \quad 0 \leq \text{ESR} \leq 1
ESR operationalizes relevance through counterfactual perturbation: an information unit is relevant if its removal changes the ranking of decision alternatives. This definition shifts evaluation from subjective importance to decision impact—a distinction absent in prior literature.
1.4 Contributions and Roadmap
This paper makes five major contributions:
1. Measurement Theory: Introduces a formalized class of counterfactually grounded metrics, with ESR as its first instantiation, establishing it as a partial sufficient statistic for manifest reasoning efficiency conditional on a closed information set.
2. Decision Theory: Provides a reduced-form micro-foundation grounded in diminishing marginal cognitive returns, linking ESR to bounded rationality and information overload.
3. Causal Inference: Establishes identification through exogenous documentation shocks (LATE) with explicit first stage, demonstrating that ESR can be used as a causal estimand under quasi-experimental conditions.
4. Epistemic Integrity: Extends beyond efficiency to measure what organizations fail to consider (Latent Signal, Epistemic Blindspot), with latent signal defined relative to ex post validation rather than pure speculation, proving Proposition 4 (Groupthink Amplification).
5. Strategic Analysis: Shows that adding noise is a dominated strategy (ESR is incentive‑compatible with respect to noise addition), though selection manipulation remains possible (ESR is not fully strategy‑proof). This places ESR in the class of partially robust governance metrics.
The paper proceeds as follows. Section 2 establishes the counterfactual foundation. Section 3 distinguishes ESR from existing metrics. Section 4 provides reduced-form micro-foundations. Sections 5-9 present formal theorems with appropriate qualifications. Sections 10-17 elaborate comparative statics, testable predictions, and domain generalizability. Sections 18-31 detail measurement protocols. Sections 32-34 illustrate empirical feasibility. Sections 35-43 introduce Latent Signal and Epistemic Blindspot. Section 44 discusses limitations, and Section 45 concludes.
---
2. THEORETICAL FOUNDATIONS
2.1 Core Claim (Qualified)
“We show that existing governance systems cannot rigorously identify decision-relevant reasoning under counterfactual criteria.”
This claim is more precise than stating they are entirely “non-identifiable.” Existing systems may capture some aspects of reasoning, but they lack the counterfactual test needed to establish whether documented reasoning actually influenced the final decision.
2.2 Counterfactual Decision Impact
Unlike existing content-analysis approaches that rely on subjective importance ratings or frequency counts, ESR defines relevance strictly in counterfactual decision-theoretic terms: whether removing an information unit changes the ranking of alternatives.
Definition 1 (Decision-Relevant Signal)
An information unit i is a decision-relevant signal if:
\text{Relevance}(i) = 1 \ \text{if} \ \text{Ranking}(\text{Alternatives} \mid \text{Information}) \neq \text{Ranking}(\text{Alternatives} \mid \text{Information} \setminus \{i\})
This shifts the unit of evaluation from importance to decision impact—a distinction absent in prior literature.
Qualification on Binary Relevance: The definition uses a binary change in ranking. In practice, relevance may be continuous; for instance, removing information might not change the ordering but could reduce confidence in the chosen alternative or shift the threshold for acceptance. We treat binary relevance as a conservative lower bound of epistemic signal—information that flips ranking is certainly relevant, while information that only shifts confidence may still matter. Thus, ESR identifies a strict subset of epistemic signal (decision-pivotal signal) rather than the full set of decision-relevant reasoning. Future extensions may consider continuous relevance based on ranking distance or confidence shifts, but the binary definition already provides a tractable and defensible starting point that yields rich theoretical and empirical implications.
2.3 Classifying Governance Metrics
Metric Class Focus Examples
Outcome-based Results Profit, accuracy, impact
Process-compliance Procedural adherence Checklists, audit compliance
Process-native epistemic (formalized & counterfactually grounded) Reasoning quality ESR
2.4 Positioning Statement
“We introduce a formalized and counterfactually grounded class of metrics: process-native epistemic metrics. ESR is the first instantiation of this class.”
---
3. DISTINGUISHING ESR FROM EXISTING METRICS
Existing Metric Why ESR Differs
Readability ESR measures decision relevance, not ease of reading
Complexity ESR measures marginal impact, not abstract complexity
Information volume ESR is a ratio, not an absolute count; long documents can have high or low ESR
Expert judgment ESR uses structured counterfactual evaluation, not subjective ratings
Shapley value / feature importance Shapley measures marginal contribution to payoff (cardinal); ESR measures change in decision ranking (ordinal)—fundamentally non-equivalent
“ESR is not a value allocation method but a decision-perturbation metric, making it conceptually orthogonal to cooperative game-theoretic attribution.”
---
4. MICRO-FOUNDATIONS: REDUCED-FORM REPRESENTATION OF DIMINISHING MARGINAL COGNITIVE RETURNS
4.1 Motivation
Our micro-foundation is not derived from a fully specified cognitive process but adopts a reduced-form representation consistent with diminishing marginal cognitive returns. This approach captures the key trade-off between information acquisition and processing capacity without requiring a complete model of cognitive architecture, and it is widely used in the bounded rationality literature (Simon, 1955; Conlisk, 1996).
4.2 Reduced-Form Model
Consider a rational agent facing a trade-off between information acquisition and cognitive costs:
S(I) = \alpha \ln(1 + I)
c(I) = \frac{\gamma}{2} I^2
Where:
· S(I) = signal extracted from total information I
· \alpha = extraction efficiency (captures cognitive capacity)
· \gamma = marginal processing cost
· \theta = value per signal
Utility function:
U(I) = \theta S(I) - c(I) = \theta \alpha \ln(1 + I) - \frac{\gamma}{2} I^2
“We adopt a reduced-form representation consistent with diminishing marginal cognitive returns, rather than deriving from a fully specified cognitive process.”
4.3 ESR as Function of Information Volume
\text{ESR}(I) = \frac{S(I)}{I} = \frac{\alpha \ln(1 + I)}{I}
4.4 Boundary Insights
\lim_{I \to 0} \text{ESR}(I) = \alpha
\lim_{I \to \infty} \text{ESR}(I) = 0
Interpretation:
· At very low information volume, all information appears relevant (ESR → α)
· At very high information volume, dilution dominates (ESR → 0)
4.5 Optimal Information Volume
First-order condition:
\frac{\theta \alpha}{1 + I^*} = \gamma I^*
---
5. THEOREM 1: INFORMATION OVERLOAD AND ESR DECLINE
Theorem 1 (Information Overload and ESR Decline)
Under the reduced-form micro-foundation with S(I) = \alpha \ln(1 + I), ESR is strictly decreasing in I for I > I_{\max}, where I_{\max} solves \frac{d}{dI}[\ln(1+I)/I] = 0.
Proof:
\frac{d}{dI}\left(\frac{\ln(1+I)}{I}\right) = \frac{\frac{I}{1+I} - \ln(1+I)}{I^2} < 0 \quad \text{for} \quad I > I_0
Since \ln(1+I) > \frac{I}{1+I} for I > 0, the derivative is negative beyond a threshold.
Implication: Beyond a certain point, adding more reasoning reduces the efficiency of decision-relevant signal extraction, mechanically lowering ESR.
---
6. THEOREM 2: IDENTIFIABILITY UNDER PANEL AGGREGATION
Theorem 2 (Identifiability under Panel Aggregation)
Under independent panel evaluation with bounded classification error \epsilon < 0.5, the ESR estimator \widehat{\text{ESR}} converges in probability to the true latent ESR as K \to \infty.
Measurement Error Model:
\text{Relevance Score}(i) = R_i^* + u_i
where u_i i.i.d., mean zero, \text{Var}(u_i) = \epsilon(1-\epsilon).
Proof:
\widehat{\text{ESR}} = \frac{1}{N} \sum_{i=1}^{N} \text{Relevance Score}(i) = \text{ESR}_{\text{true}} + \bar{u}. By Law of Large Numbers, \bar{u} \xrightarrow{p} 0 as K \to \infty.
Qualification: The theorem assumes independence or weak dependence of panel judgments. In practice, panelists may share systematic biases (e.g., groupthink), which can violate this assumption. Our later analysis (Proposition 4) addresses such scenarios explicitly.
---
7. THEOREM 3: CAUSAL IDENTIFICATION
Theorem 3 (Causal Identification under Exogenous Documentation Shocks)
First Stage:
\text{ESR} = \pi_0 + \pi_1 Z + \eta \quad \text{with} \quad \pi_1 \neq 0
Second Stage (Wald Estimator):
\beta_1 = \frac{\text{Cov}(Q, Z)}{\text{Cov}(\text{ESR}, Z)} = \text{LATE} = \mathbb{E}[Q(1) - Q(0) \mid \text{Compliers}]
Where:
· Compliers: decision units whose ESR changes due to documentation shock Z
· Always-takers: units with high documentation regardless of Z
· Never-takers: units with low documentation regardless of Z
Examples of plausible instruments Z in governance settings:
· Staggered adoption of standardized decision documentation protocols across organizational units
· Externally mandated reporting requirements (e.g., donor-imposed standards, regulatory transparency rules)
· Introduction of AI-assisted documentation tools that alter the structure of recorded reasoning
Interpretation: The first stage ensures Z genuinely moves ESR (relevance condition). The estimand is the LATE for compliers—those whose reasoning structure is affected by the documentation shock.
---
8. DYNAMIC IMPLICATION: DECISION STABILITY
Prediction:
\text{Pr}(\text{Ranking change}_{t+1}) = \delta_0 - \delta_1 \text{ESR}_t \quad \text{with} \quad \delta_1 > 0
Or, in variance form:
\text{Var}(\text{Ranking}_{t+1}) = \sigma^2(\text{ESR}_t), \quad \frac{d\sigma^2}{d\text{ESR}} < 0
Interpretation: Higher ESR reduces the probability of future ranking changes—decisions are more robust to new information.
---
9. MEASUREMENT ERROR AND ATTENUATION BIAS
Attenuation Bias:
\hat{\beta}_1^{\text{OLS}} \rightarrow \lambda \beta_1 \quad \text{with} \quad 0 < \lambda < 1
Interpretation: Classical measurement error in ESR biases the OLS coefficient toward zero. If estimates remain significant despite this bias, the true effect is likely larger than estimated.
---
10. STRATEGIC MANIPULATION: PARTIALLY ROBUST GOVERNANCE METRICS
10.1 Adding Noise is a Dominated Strategy
Consider an agent who can add noise N (information units that do not change ranking):
\text{ESR}(N) = \frac{S}{S + N}
Result:
\frac{\partial \text{ESR}}{\partial N} = -\frac{S}{(S+N)^2} < 0
Adding noise strictly decreases ESR. Thus, ESR is incentive‑compatible with respect to noise addition: no agent who values ESR positively will add noise.
10.2 Selection Manipulation Remains Possible
However, agents can manipulate ESR through other channels:
· Selective documentation: omitting relevant dissent or counter-evidence (reducing denominator while keeping signal unchanged)
· Framing manipulation: presenting the same information in ways that appear more decisive
· Gaming the panel: influencing panelists’ judgment through persuasion or selection
These forms of manipulation are not ruled out by the above analysis. Therefore, ESR is not fully strategy‑proof.
“This places ESR in the class of partially robust governance metrics—robust to additive noise but vulnerable to selection and framing manipulation.”
---
11. FORMAL COMPARATIVE STATICS
From the first-order condition \frac{\theta \alpha}{1 + I^*} = \gamma I^*:
Parameter Derivative Sign Interpretation
Extraction efficiency \alpha \frac{\partial I^*}{\partial \alpha} > 0 + Higher capacity → more information processed
Value per signal \theta \frac{\partial I^*}{\partial \theta} > 0 + Higher incentives → more information
Cognitive cost \gamma \frac{\partial I^*}{\partial \gamma} < 0 - Higher cost → less information
Effects on optimal ESR:
\frac{\partial \text{ESR}^*}{\partial \alpha} > 0,\quad \frac{\partial \text{ESR}^*}{\partial \theta} > 0,\quad \frac{\partial \text{ESR}^*}{\partial \gamma} < 0
---
12. ESR AS A PARTIAL SUFFICIENT STATISTIC
“ESR can be interpreted as a partial sufficient statistic for manifest reasoning efficiency conditional on a closed information set.”
Under the assumption that all decision-relevant information is captured in the documented reasoning (i.e., no latent signal), ESR captures all relevant variation in the efficiency of processing that information. However, as we show in Sections 35-43, latent signal exists and matters. Therefore, ESR is not sufficient for overall decision quality—it is a partial sufficient statistic for the efficiency of manifest reasoning, but it must be complemented with measures of signal completeness (e.g., Epistemic Blindspot) to fully assess reasoning quality.
“ESR is sufficient for manifest reasoning efficiency conditional on a closed information set, but it is not sufficient for decision quality due to latent signal (Section 20).”
---
13. ESR VS SHAPLEY VALUE: FUNDAMENTAL NON-EQUIVALENCE
Aspect Shapley Value ESR
Domain Payoff (cardinal) Preference ranking (ordinal)
Invariance Monotonic payoff transformations Monotonic ranking transformations
Information needed Payoff values Ranking comparisons
“Shapley value is invariant to monotonic transformations of payoff, while ESR is invariant to monotonic transformations of preference rankings—making them fundamentally non-equivalent objects.”
“ESR is not a value allocation method but a decision-perturbation metric, making it conceptually orthogonal to cooperative game-theoretic attribution.”
---
14. COMPETING PREDICTIONS: SHARP TESTABLE RESTRICTIONS
Empirical Model:
Q = \beta_0 + \beta_1 \text{ESR} + \beta_2 \text{ESR}^2 + \varepsilon
Framework Empirical Restriction
ESR Theory \beta_1 > 0,\ \beta_2 < 0
Readability \beta_2 = 0 (monotonic positive)
Information volume \beta_2 = 0 (monotonic positive)
Complexity \beta_1 < 0,\ \beta_2 = 0 (monotonic negative)
---
15. LIMIT CASE ANALYSIS
Limit Case ESR Interpretation
All units relevant S = N, \text{ESR} = 1 Maximum efficiency
No signal S = 0, \text{ESR} = 0 No decision-relevant content
Noise → ∞ \lim_{N \to \infty} \frac{S}{S+N} = 0 ESR collapse
Information → 0 \lim_{I \to 0} \text{ESR}(I) = \alpha All information appears relevant
Information → ∞ \lim_{I \to \infty} \text{ESR}(I) = 0 Dilution dominates
---
16. EMPIRICAL TEST MAPPING
Prediction Empirical Design Identification
Noise ↑ → ESR ↓ External documentation shocks First stage: \pi_1 \neq 0; IV: \beta_1 = \text{LATE}
ESR–Q inverted-U Panel dataset \beta_1 > 0,\ \beta_2 < 0
Capacity ↑ → ESR* ↑ Training, AI assistance Before-after, DiD
Manipulation via selection Panel evaluation with fixed selection criteria Need additional controls
ESR ↑ → stability ↑ Panel longitudinal \delta_1 > 0
---
17. DOMAIN GENERALIZABILITY
Domain Example Applications
Public policy Regulatory impact analysis, strategic policy decisions
Corporate governance Board decisions, strategic investment
AI alignment Reasoning trace of generative AI models
Non-profit management Fund allocation, social programs
---
18. MEASUREMENT PROTOCOL
18.1 Unit of Analysis: Minimal Reasoning Unit
i \in \mathcal{U} = \text{set of minimal independently evaluable reasoning units}
Operational Rules:
1. Separability: Can be removed without destroying overall logical structure
2. Evaluative independence: Has a claim that can be independently assessed
3. Counterfactual testability: Can be tested in ranking framework
18.2 Reliability: Unitization and Relevance
\text{IRR}_{\text{unitization}} = \text{Krippendorff’s Alpha (unit segmentation)}
\text{IRR}_{\text{relevance}} = \text{Krippendorff’s Alpha (relevance scores)}
Interpretation: \geq 0.80 = reliable; 0.67-0.80 = acceptable; <0.67 = revision needed.
18.3 Relevance Scoring
\text{Relevance Score}(i) = \frac{1}{K} \sum_{k=1}^{K} \mathbb{I}_{\text{ranking changes according to panelist } k}
18.4 ESR Estimation with Uncertainty
\text{ESR}_{\text{true}} = \frac{1}{N} \sum_{i=1}^{N} \text{Relevance Score}(i)
\text{SE}_{\text{ESR}} = \frac{\text{StdDev}(\text{Relevance Score}(i))}{\sqrt{N}}
\text{CI}_{95\%} = \text{ESR}_{\text{true}} \pm 1.96 \times \text{SE}_{\text{ESR}}
18.5 Weighted ESR
\text{ESR}_{\text{weighted}} = \frac{\sum_{i=1}^{N} W_i \cdot R_i}{\sum_{i=1}^{N} W_i}
Category Weight (W_i)
Critical assumptions 3
Evidence 3
Substantive alternatives 2
Documented dissent 2
Minor supporting arguments 1
18.6 Tiered Implementation
Tier Method Cost Use Case
Tier 1: ESR Lite \text{ESR}_{\text{lite}} = \frac{\text{arguments used}}{\text{total arguments}} Low NGOs, quick pilots
Tier 2: Internal BCEV+ Internal cross-unit panel Medium Ministries, mid-sized corporations
Tier 3: Full BCEV+ External panel with uncertainty estimation High International organizations, accreditation
---
19. INTERPRETATION AND FAILURE THRESHOLD
ESR Level Condition Risk
< 0.20 Noise dominant Signal dilution
0.20 – 0.39 Much irrelevant information Cognitive overload
0.40 – 0.80 Signal dominant Maximum quality
0.90 Over-filtering Loss of nuance
---
20. EXTENDING BEYOND EFFICIENCY: LATENT SIGNAL AND EPISTEMIC BLINDSPOT
20.1 The Missing Dimension
ESR measures how efficiently organizations process manifest information. But what about information that should be considered but never enters deliberation? This is the domain of Latent Signal (S_L) and Epistemic Blindspot (EB).
Definition 2 (Latent Signal)
Information that objectively has counterfactual power to change decision ranking but is absent from documented reasoning or systematically ignored by decision-makers. Latent signal is defined relative to ex post validation or external domain knowledge, not purely speculative counterfactuals. Crucially, latent signal is not defined by realized outcomes alone, but by information that was ex ante available and ex post validated, excluding hindsight‑only constructs. For instance, a warning that a project would fail that was ignored and later proved correct qualifies as latent signal; a hypothetical argument that might have changed the decision but was never proposed does not.
Sources of latent signal:
· Political pressure or organizational culture (groupthink)
· Epistemic capture by dominant interests
· Cognitive limitations or selection bias
· Unknown unknowns (detectable only through systematic adversarial probing)
Definition 3 (Epistemic Blindspot)
EB = \frac{S_L}{S_M + S_L}
where S_M = manifest signal (signal present in documentation).
Definition 4 (Signal Completeness)
SC = 1 - EB
20.2 Adjusted Decision Quality
Q_{adj} = f(ESR, SC)
True decision quality depends not only on how efficiently we process available information (ESR), but also on how complete that information is (SC).
20.3 The Four Quadrants of Epistemic Health
ESR EB Condition Risk
High High Efficient but blind Missed forest for the trees; sudden crisis
Low Low Comprehensive but messy Analysis paralysis; slow decisions
Optimal Low Ideal Efficient and complete; long-term stability
Low High Chaotic and blind Total failure
---
21. PROPOSITION 4: GROUPTHINK AMPLIFICATION
Proposition 4 (Groupthink Amplification)
As groupthink (\phi) increases, the observed ESR may stay constant or even increase (due to artificial consensus), but the Latent Signal (S_L) increases exponentially, leading to a catastrophic drop in Q_{adj}.
Mechanism with a stylized functional form:
Let S_L(\phi) = S_0 e^{k\phi} where S_0 is baseline latent signal and k > 0 captures the amplification rate of ignored information under groupthink. Observed ESR, defined using only manifest signal S_M, remains stable or may even rise because consensus eliminates perceived dissent (reducing denominator in ESR calculation). Meanwhile, latent signal grows exponentially. When the ignored information eventually surfaces, the collapse in decision quality is sudden and severe.
Implications: High ESR does not guarantee safety. If EB is high, organizations are time bombs waiting to explode. This explains:
· Major project failures undetected beforehand
· Corporate scandals that “suddenly” emerge
· Policy crises that minority voices had warned about
---
22. DETECTING LATENT SIGNAL (ADVERSARIAL AUDIT)
Since S_L is “invisible” (absent from documentation), detection requires structured adversarial methods:
1. Red Teaming: Independent external panel searches for information not mentioned in documentation but relevant to the domain, using domain expertise and external benchmarks.
2. Counter-Ranking Test: Present new information to decision-makers; if ranking changes drastically, that information was S_L.
3. Dissent Gap: Calculate difference between suppressed minority opinions and final decision.
Epistemic Boundary: Latent signal is not purely speculative; it is validated ex post (e.g., the information later proved correct) or grounded in established domain knowledge. This prevents conflating hidden signals with hypothetical alternatives that may never have been relevant.
---
23. REVISED COMPARATIVE STATICS
Parameter Effect on ESR Effect on EB Effect on Quality (Q)
Psychological Safety Decreases (more discussion) Decreases sharply Increases
Groupthink (\phi) Increases (appears aligned) Increases sharply Decreases drastically
AI Auditor Stable Decreases Increases
---
24. LIMITATIONS AND FUTURE RESEARCH
Limitation Mitigation / Future Work
Unit analysis ambiguity Formal Minimal Reasoning Unit; separate IRR for unitization
Panel subjectivity and dependence Theorem 2 requires independence/weak dependence; future work on correlated errors
Endogeneity Theorem 3: causal identification with exogenous shocks; requires valid instruments
Measuring S_L Adversarial audit; acknowledged as estimation, not absolute measure; latent signal defined relative to ex post validation, excluding hindsight‑only constructs
Audit fatigue Need research on optimal frequency and integration with regular operations
Epistemic paranoia Need satisficing criteria for when to stop searching
Power asymmetry Need explicit integration with power structures (future work)
Selection manipulation Current analysis shows adding noise is dominated; manipulation via selection remains open; ESR is incentive‑compatible with respect to noise addition but not fully strategy‑proof
---
25. EMPIRICAL ILLUSTRATION: IDENTIFICATION FEASIBILITY
Dataset: N = 300, ESR ∼ Uniform(0.1, 0.9), Q = 0.5 + 1.5\,\text{ESR} - 1.8\,\text{ESR}^2 + \varepsilon
Results:
Variable Coefficient Std. Error p-value
ESR 1.48 0.21 < 0.001
ESR² -1.72 0.18 < 0.001
\text{ESR}^* \approx 0.43,\quad R^2 = 0.58
Interpretation: The empirical illustration establishes identification feasibility. The next empirical step is quasi-experimental validation using externally induced variation in documentation protocols.
---
26. DIAGRAM OF ESR MEASUREMENT
```
┌─────────────────────────────────────────────────────────────────────────────┐
│ Information Supply + Cognitive Filtering + Organizational Routines │
│ │ │
│ ▼ │
│ ┌───────────────────────────────────────────────────────────────────────┐ │
│ │ STEP 1: UNITIZATION → Minimal Reasoning Units + IRR_unitization │ │
│ └───────────────────────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌───────────────────────────────────────────────────────────────────────┐ │
│ │ STEP 2: PANEL EVALUATION (BCEV+) → Relevance Score(i) + IRR_relevance│ │
│ └───────────────────────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌───────────────────────────────────────────────────────────────────────┐ │
│ │ STEP 3: ESR ESTIMATION → ESR_true ± SE_ESR, CI_95% │ │
│ └───────────────────────────────────────────────────────────────────────┘ │
│ │ │
│ ▼ │
│ ┌───────────────────────────────────────────────────────────────────────┐ │
│ │ STEP 4: OUTCOME ANALYSIS → Q = α + β₁ ESR + β₂ ESR² + ε │ │
│ │ Sharp restriction: β₂ < 0 │ │
│ └───────────────────────────────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────────────────┘
```
---
27. SIMULATION ILLUSTRATION
```
Decision Quality (Q)
↑
1.0 | *
| * *
0.8 | * *
| * *
0.6 | * *
| * *
0.4 | * *
| * *
0.2 |* *
|____________________→ ESR
0 0.2 0.4 0.6 0.8 1.0
↑
ESR* ≈ 0,43
```
---
28. CONCLUSION
28.1 Summary of Contributions
1. Counterfactual perturbation – Redefines relevance as decision impact rather than subjective importance; fundamentally distinct from Shapley (ordinal vs cardinal invariance). ESR identifies a strict subset of epistemic signal (decision-pivotal signal) rather than the full set of decision-relevant reasoning.
2. Reduced-form micro-foundation – S(I) = \alpha \ln(1+I), justified by diminishing marginal cognitive returns rather than derived from a fully specified cognitive process
3. Theorem 1 (Information Overload) – Explicit derivation of ESR decline beyond optimal volume
4. Theorem 2 (Identifiability) – Consistent estimation under panel aggregation, with qualification about independence
5. Theorem 3 (Causal Identification) – LATE for compliers with explicit first stage and concrete instrument examples
6. Dynamic implication – \text{Pr}(\text{Ranking change}_{t+1}) = \delta_0 - \delta_1 \text{ESR}_t
7. Attenuation bias – \hat{\beta}_1^{\text{OLS}} \rightarrow \lambda \beta_1
8. Strategic analysis – Adding noise is dominated (incentive‑compatible with respect to noise addition); selection manipulation remains possible (not fully strategy‑proof). This places ESR in the class of partially robust governance metrics.
9. Partial sufficient statistic – ESR summarizes manifest reasoning efficiency conditional on a closed information set; it is not sufficient for decision quality due to latent signal
10. Sharp testable restrictions – \beta_1 > 0,\ \beta_2 < 0 distinguishing ESR from alternatives
11. Epistemic Blindspot – Extends framework to measure what organizations fail to consider, with latent signal defined relative to ex post validation, excluding hindsight‑only constructs
12. Proposition 4 (Groupthink Amplification) – High ESR with high EB is a time bomb; stylized form S_L(\phi) = S_0 e^{k\phi} illustrates exponential growth of ignored information
28.2 Core Claims (Qualified)
“We show that existing governance systems cannot rigorously identify decision-relevant reasoning under counterfactual criteria.”
“We introduce a formalized and counterfactually grounded class of metrics: process-native epistemic metrics. ESR is the first instantiation of this class.”
“ESR makes epistemic accountability measurable, complementing existing governance systems that focus on outcomes and compliance.”
“ESR is incentive‑compatible with respect to noise addition, but not fully strategy‑proof. This places ESR in the class of partially robust governance metrics.”
28.3 Final Statement
“ESR does for reasoning what productivity metrics did for labor: it transforms an invisible process into a measurable, optimizable, and governable variable—while recognizing that no single metric can capture all dimensions of reasoning quality.”
---
29. REFERENCES
· Angrist, J. D., & Pischke, J.-S. (2009). Mostly Harmless Econometrics. Princeton University Press.
· Conlisk, J. (1996). Why bounded rationality? Journal of Economic Literature, 34(2), 669–700.
· Grimmelikhuijsen, S., Jilke, S., Olsen, A. L., & Tummers, L. (2017). Behavioral public administration: Combining insights from public administration and psychology. Public Administration Review, 77(1), 45–56.
· Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux.
· Kahneman, D., & Tversky, A. (1979). Prospect theory: An analysis of decision under risk. Econometrica, 47(2), 263–291.
· Krippendorff, K. (2018). Content Analysis: An Introduction to Its Methodology (4th ed.). Sage.
· Simon, H. A. (1947). Administrative Behavior. Macmillan.
· Simon, H. A. (1955). A behavioral model of rational choice. Quarterly Journal of Economics, 69(1), 99–118.
· Shapley, L. S. (1953). A value for n‑person games. Annals of Mathematics Studies, 28, 307–317.
---
30. APPENDICES (Indicative)
· Appendix A: Coding Manual for Minimal Reasoning Unit
· Appendix B: Scoring Rubric for BCEV+ Panel
· Appendix C: Decision Traceability Log Template
· Appendix D: IRR Results (Unitization and Relevance)
· Appendix E: Dataset Synthesis Code (Python/R) for Replication
· Appendix F: Full Regression Outputs with Robust Standard Errors
· Appendix G: Adversarial Audit Checklist for Latent Signal Detection
---
Accountability‑Based Universal Wisdom and Trust
Cross‑Sector Pre‑Decision Governance Translator
Final Manuscript · March 2026
License: CC BY-NC-SA 4.0
Contact: tpapgtk@gmail.com
Archive: https://abuwt.blogspot.com
