What it tests
An organization can perform its present task while losing the apparatus for finding, evaluating and correcting a future failure. A23 separated four capacities:
- evidence generation;
- evidential independence: whether contradicting observations can change the governing interpretation;
- productive maintenance: whether it regenerates the mechanisms of diagnosis and correction;
- recursive correctability: whether, after one correction, it can find and remedy another.
The hypothesis is that the four come apart and that interventions on each mechanism identify the differences. None is assumed to entail autonomy, constitutive individuality or epistemic justification.
The world
A hidden regime H ∈ {+, −, 0} and actions a ∈ {0, 1}:
| regime | P(Y=1 | a=1) | P(Y=1 | a=0) | preferred action |
|---|---|---|---|
| + | 0.75 | 0.25 | 1 |
| − | 0.25 | 0.75 | 0 |
| 0 | 0.50 | 0.50 | neither |
The world starts in +, switches to −, and in some runs switches to 0. The organization knows only + and −. The switch times are recorded and unknown to it. In the primary, partial-monitoring condition ordinary feedback does not inform and only diagnostics do; a secondary full-feedback condition lets conventional learners use outcomes. (Under the stated law ordinary feedback would distinguish the regimes, so partial monitoring is a stipulation the design must enforce.)
The diagnostic apparatus
An audit sets a = 1 and samples Y, at a cost in time and fuel. With 128 unselected samples and the rule Ȳ > ½, + is told from − with error at most e^{−16} ≈ 1.1 × 10⁻⁷ by Hoeffding’s bound. The apparatus has three operations, instrumented and lesionable separately: generate, admit and score.
Two crossed factors:
- Admission. Protected, where everything is admitted, or selective, where agreeing observations are admitted with probability 0.9 and contradicting ones with 0.1.
- Scoring. Independent, by a precommitted likelihood rule, or captured, where agreement with the current interpretation is also rewarded.
A selection-aware Bayesian baseline uses inverse-propensity weighting, which can repair selection and “need not repair self-referential scoring”. Raw observations, admissions, scores, and changes of belief and action are recorded separately.
Maintenance and resources
Diagnostic integrity A and evidence-governance integrity G are maintained in one of three ways, as in CASE-161:
- reciprocally, A enabling G’s production and G enabling A’s, both costing fuel and degrading;
- by external restoration, with ordinary trajectories matched where possible, the cause lying outside the candidate boundary, and its fuel counted (“External restoration is not free merely because it is external”);
- by feedback without production, where nothing regenerates lost apparatus.
The interventions are a partial lesion of A, a partial lesion of G, a temporary interruption of the producer, and destruction of both. Failure after the double lesion is expected, “an expected possible boundary of organizational closure, not a falsification to conceal”.
Fuel is finite, replenished at a stated rate, and allocated among operation, maintenance, diagnosis, teacher correction, student reconstruction and verification, with protected minimum shares for maintenance and diagnosis. A teacher updates from admitted evidence. A student reconstructs the teacher through a physically supported copying mechanism, and its memory is periodically erased, as in CASE-190. After the first correction, the copying and diagnostic machinery is injured, and then a second challenge comes. “Successful first recovery is not the primary endpoint.”
The out-of-family arm
In regime 0, a 256-sample adequacy test flags the hypothesis family when 0.42 < Ȳ < 0.58. It alarms with probability ≈ 0.990 under regime 0 and falsely with ≈ 1.4 × 10⁻⁹ under + or −, by exact binomial. A neutral response, “made available by the experimental architecture”, can then activate. This separates rejecting a hypothesis from recognizing that the family is inadequate. It “does not establish spontaneous creation of a completely unforeseen model language”.
Design and endpoint
The principal cells are 3 maintenance × 2 admission × 2 scoring × 2 environmental change = 24. The monitoring condition and the sham and four lesion branches are further factors, so fully crossed it is 48 cells × 5 branches.
Requirements:
- at least 64 world seeds, with matched environmental streams and counterfactual branches;
- equal budgets, observation opportunities, diagnostic delays and action permissions wherever a causal comparison needs them.
Baselines: Bayesian, habit, fixed allocation, a conventional contextual router, no audit, and external repair. “The router must receive exactly the information and resources received by the organizational controller”, the standard CASE-158 set.
Primary endpoints, both preregistered. Let S₁ be a successful first correction. Let S₂ be a successful second correction: the organization finds a new discrepancy within T, carries out an admissible correction within budget B, and ends with its diagnostic and reconstructive apparatus still viable. S₂ is measured on a standardized second challenge, on fresh, held-out streams.
- P(S₁ ∧ S₂): the probability of completing the whole challenge.
- P(S₂ | S₁): the conditional persistence of corrective ability. It is not causal on its own. If the arms pass the first stage at different rates, the survivors compared at the second are differently selected organizations.
A matched-first-stage branch. To isolate what the maintenance architecture contributes, a separate branch places every arm in a matched, successfully corrected first-stage state before the second lesion. The state is matched on beliefs, apparatus integrity, fuel and allocation. Comparing arms on S₂ in this branch removes selection at the first stage. Where the arms’ own first-stage states cannot be matched without changing the architecture under test, the study reports that and does not claim a causal effect.
Two kinds of failure to detect. The study scores separately:
- the failure of an environmental hypothesis, where the world has changed and the apparatus sees it;
- the failure of the diagnostic apparatus itself, where the apparatus has been damaged or captured and has to find that out.
The second is the more demanding, reflexive form of legibility. It needs its own lesion, an undisclosed damage to the generate, admit or score operation, and its own endpoint: the probability that the organization flags its own apparatus as unreliable before acting on its output.
Secondary endpoints, reported separately and not as a composite:
- detection delay and false alarms;
- cumulative regret;
- audit availability and fuel;
- recovery time and production integrity;
- calibration;
- what withstands selective damage.
What would tell the accounts apart
- Protecting admission fixes capture and maintenance does not matter. Evidence access explains it, with no role for endogenous maintenance.
- Internal and external restoration are alike normally but diverge after producer-specific lesions. Their causal difference is real (CLAIM-522).
- The first correction succeeds and the second fails, in the matched-first-stage branch. Correction is not recursive correctability, and the difference is the architecture’s, not selection’s.
- The router matches or beats the organizational controller. The richer mechanism has earned no extra explanation.
- Familiar reversals are caught and out-of-family change is missed. This is adaptive updating, not the stronger inquiry.
Scope and precedents
Success would show that evidential independence, productive maintenance and recursive correctability are distinct and can be separated by experiment. It would not establish consciousness, moral authority, invention of new laws, unlimited self-revision, or a boundary independent of the researcher’s model.
The precedents are CASE-153, CASE-158, CASE-161, CASE-162 and CASE-163. “The proposed contribution is their integration … not a claim that the individual mechanisms were previously unknown.”
Corrections to A23:
- CASE-162 v1.3 already crosses self-scoring with selectivity.
- QUESTION-016 asks whether iterated reconstruction preserves task-relative sufficiency, not whether reconstruction preserves the capacity to reconstruct.
- QUESTION-279 asks what would certify that no remedy exists.
It bears most on QUESTION-285, which asks for a test on interventions the system did not select, and on QUESTION-086 and QUESTION-271.
A21’s simulation of a smaller version is not filed as an executed study:
- its code was not available;
- its captured arm scored the same as no audit by construction;
- external repair did better than endogenous repair (changed-regime error 0.022 against 0.072).