17.8.3 Context-cover construction (CTTE–2)

17.8.3 Context-cover construction (CTTE–2)

CTTE–2 removes the context labels that CTTE–0 and CTTE–1 supplied. Each source instead carries two anonymous continuous metadata coordinates. The constructor must decide whether the evidence supports no contextual split, one binary cover, or two binary covers whose four intersections define local models. It must also attach every source to the appropriate opens and overlap. The stable entities \(X\), \(Z\), and \(Y\) are supplied so that the experiment isolates contextual construction rather than repeating entity registration.

The preregistered benchmark contains 1,000 ledgers, with 200 from each of five families. Invariant worlds span a clear two-by-two metadata population but do not require a split. Product worlds require both axes; one-axis worlds require only the first and therefore penalize unnecessary refinement. Ambiguous worlds require both axes but measure the second too imprecisely to certify its boundary. In missing-overlap worlds, both axes are individually separated but perfectly correlated in discovery and admission, while confirmation asks for off-diagonal intersections that have never been observed.

Discovery sources report noisy \(X\to Y\) effects. Admission sources report \(X\to Z\) and \(Z\to Y\), so a proposed cover must support the contextual factorization rather than merely cluster metadata. Confirmation exposes its metadata but seals its effects until the cover, restriction maps, epistemic status, and predictions are frozen. The typed resolver certifies each boundary under uncertainty and selects the least complex cover satisfying both the discovery and factorization tests.

The result was a preregistered near miss, not an admission. Overall action accuracy was \(97.9\% \). The product cover was constructed in \(98.5\% \) of product worlds, with \(98.5\% \) coverage-adjusted restriction membership and \(98.5\% \) confirmation accuracy. The resolver selected the minimal one-axis cover in \(98.5\% \) of the corresponding worlds, retained the invariant theory in \(99.0\% \), never falsely refined a one-axis theory into a product cover, and abstained in every ambiguous-boundary world. It also remained within \(1.5\) percentage points of the context oracle on product confirmation.

One gate nevertheless failed. The registered requirement was at least \(95\% \) abstain-nonidentifiable decisions in missing-overlap worlds; the observed rate was \(93.5\% \). In the remaining 13 of 200 worlds, noise caused the mediator-factorization check to fail before the overlap audit, producing the less informative status abstain-unsupported. The system did not make an off-support prediction, but it did not correctly localize why prediction was unwarranted. This distinction matters scientifically: “the proposed mechanism lacks support” and “the requested population was never observed” call for different acquisitions.

The flat resolver’s perfect product prediction illustrates why prediction alone is inadequate. It always constructs four cells when their metadata clusters are visible, thereby over-refining invariant and one-axis theories. The near miss exposes an ordering defect: overlap support should be audited before a noisy mechanism failure receives its terminal status. That claim belongs in a separately registered repair, not a post-hoc change to CTTE–2.

Experiment: CTTE–2: inferred context covers. Input: 1,000 generated ledgers across invariant, product, one-axis, ambiguous-boundary, and missing-overlap families.
Withheld structure: context boundaries, local opens, four possible intersections, and source restriction maps.
Structural result: \(97.9\% \) action accuracy, \(98.5\% \) product-cover construction and confirmation, and zero false product refinement.
Failed gate: missing-overlap localization was \(93.5\% \) against a registered \(95\% \) minimum; 13 safe abstentions carried the wrong reason.
Verdict: not admitted. The failed obligation is epistemic-status ordering; the result does not justify threshold tuning.