The claim
A33 §8 drew the line first: “Irreducible conflict exists when the objective demands cannot be jointly improved under the relevant admissible actions. Manufactured frustration exists when a common improvement is available, but the governance rule, restricted representation, or allocation of influence prevents the system from finding or using it.” “This is testable with an oracle or stronger solver for the same declared objectives and update constraints.”
CLAIM-365 makes it exact for the local model. Demand i requires ε_i + g_iᵀd ≤ 0 of an update d. Define the worst residual of an update, the frustration residual and the governance regret:
m(d) = max_i(ε_i + g_iᵀd), Φ = min_{d∈U} m(d), Γ = m(d_actual) − Φ ≥ 0.
- Structural: Φ > 0. No admissible update meets all demands. A35 §3.3 separates a capacity-limited case, where Φ_r > 0 for the current budget r but not for a larger one, from the case where no budget helps, as in CASE-115, where Φ_r = ε for every r.
- Realized violation: Φ ≤ 0, so a jointly satisfactory update exists, and the actual update violates at least one requirement. That holds exactly when m(d_actual) > 0, equivalently Γ > −Φ. The actual update meets every requirement exactly when m(d_actual) ≤ 0, equivalently Γ ≤ −Φ. A35: “A system could therefore have substantial observed frustration even when structural frustration is absent.”
Positive regret is not failed satisfaction. Γ measures how much worse the actual update’s worst-case residual is than the best achievable one. It can be positive when the actual update meets every declared requirement. Take one demand with ε = 1 and g = 1, and admissible updates U = [−2, 0]. Then m(d) = 1 + d, Φ = −1 at d = −2. The actual update d = −1 has residual m(−1) = 0, so it meets the requirement, yet Γ = 1 > 0 because a more favourable update was available. Γ > 0 establishes suboptimality relative to the chosen minimax objective, not that any requirement went unmet. Actual violation needs the further comparison Γ > −Φ when Φ ≤ 0. Four claims are kept apart: structural infeasibility (Φ > 0), realized violation (m(d_actual) > 0), minimax suboptimality (Γ > 0), and causal governance failure, which neither quantity identifies.
What the diagnostics do not show. The first filing called the second case governance-induced frustration. It is not, by itself. Neither Φ nor Γ identifies an obstructing mechanism, a governance failure or active thwarting. An available update can go unmade because no process pursued it, because it was not recognized, through error, or through an unexpected event. None of these requires anything to have blocked it. Γ does not say whether a demand was actively thwarted, which process did it, or whether the demand was legitimately subordinate to another. Governance-induced shortfall is the case where the allocation of influence is what kept the update from being made, and showing that needs an intervention on the allocation with search and information held fixed. Φ > 0 is structural frustration of the declared improvement demands in CLAIM-347’s sense. How a realized violation relates to operative thwarting is QUESTION-131’s question.
A40 §3.4 added a representational kind, where an available solution is inaccessible to the organization’s own models or search, and said the categories “may overlap”.
Aggregate performance does not separate them. A35 §3.4: two systems, one with Φ_r > 0, one with Φ_r < 0 and an optimizer that repeatedly picks an unhelpful direction, “should not receive the same diagnosis simply because their loss curves look equally poor”. The same holds of a zero aggregate gradient (CLAIM-351).
The certificate’s multipliers are not governance’s weights. The λ that attains the dual of Φ picks out the combination of demands that binds. A35: “Importantly, these dual multipliers are not endorsement weights. They are mathematical certificates of difficulty or infeasibility.” A40 §4: “The weights assigned by actual governance need not coincide with the multipliers”, and the divergence “is a candidate source of avoidable frustration”. Whether they coincide is QUESTION-126.
Why the record holds it
It is where the frustration account meets the operator model. CLAIM-297 says authority updated on the interpretation it helped make dominant can amplify without gaining fidelity; CLAIM-330 gives the instability threshold βγ > (1 − α)(1 − δ) in CASE-099’s linearization; CLAIM-342 says what keeps such a loop honest is evidence that does not pass through the interpretation it tests. A loop of that kind is a mechanism of governance-induced frustration: the satisfactory update exists and the captured allocation of influence does not choose it. A39’s case study 2, CASE-099 with demands and a feasible-update oracle, would test it; A39 called it “the best candidate for an early discriminating experiment”. That the governance-induced kind does work an optimizer analysis does not is filed apart as CLAIM-500.
What it does not say
- That governance-induced frustration occurs in any real organization. The claim gives the distinction and the measure, not an instance. The defeater is the result that would make the category idle.
- That Γ is free of choices. Φ is relative to declared demands and a declared admissible set, and computing it needs a solver that finds the minimum: an estimate from a weaker solver overstates Φ and understates Γ. A wider admissible set can turn structural into governance-induced.
- That governance-induced is worse. It says a better update existed by the declared standard. Whether the declared standard is the organization’s own is QUESTION-087’s problem.
- That λ should be the weights. Governance may rightly weight demands by what matters, not by what binds.
Note of 2026-10-25: the repeated version, and what averaging costs
From A29–A31 of the exchange of 2026-10-24, continued. This claim’s worst residual m(d) = maxᵢ(εᵢ + gᵢᵀd) is a one-shot test against the target set {z : zᵢ ≤ 0 for every i}. Blackwell approachability asks the repeated version: whether a learner can drive dist((1/T)Σ f_t, C) → 0 for a vector of demand residuals f_t against disturbance or opposition. Kwon and Perchet (2017, LIT-1526, read in NOTE-1159) treat it under partial feedback. Their target is exactly this claim’s {z : zᵢ ≤ 0}, and the distance they use is ‖(ḡ_T)₊‖₂, the ℓ₂ norm of the positive parts of the average (THEORY-947). A requirement violated by 1 on alternate stages and met by 1 on the others has average 0 and is approached perfectly. Two cautions from the replies. First, averaging is compensatory: “violating a safety requirement today may not be remedied by outperforming it tomorrow” (A30), and a rolling-window requirement “is stronger than merely achieving a suitable long-run average and is not directly guaranteed by ordinary approachability results” (A31). Second, “measuring distance to the acceptable set introduces choices of normalization and geometry” (A30), a scale this claim does not otherwise need. Approachability is already in the record as the engine of Hart and Mas-Colell’s proof (NOTE-1142): regret matching is approachability of the negative orthant by the regret vector. The claim is unchanged.