Worldwright research programme: current landscape
Date: 2026-10-23 Status: active programme; corpus-side mirror of Sourcery snapshot bff142b8-1ae9-4f7b-ba2e-9631d2ccfdaf
Worldwright has progressed beyond a conceptual simulator. It now has a qualified bounded execution owner, exact replay, paired counterfactuals, a research service, mechanism ablations, failure stratification, weight tests, response surfaces, delay-path decomposition and threshold/gain factorials. It has also preserved several useful failures rather than tuning them away.
The strongest current synthetic finding is a conditional boundary. Mediation consistently improves the model's relational coordinates, but delay under material deficit worsens health and accumulated harm. With delivery failures disabled, all 40 lower-urgency cells favor mediation and all 120 higher-urgency cells favor immediate visible action. Every retained cell is nevertheless a Pareto tradeoff; the aggregate winner is diagnostic rather than morally or politically authoritative.
Mechanism ablation localizes the result. Removing urgency changes 105 of 160 matched winners, shortening relational persistence changes 32, removing the relational mechanism changes 33, and removing identity changes none on the original four-preset grid. A subsequent independently varied response surface corrects the broad interpretation of that last result: moving identity salience from 0 to 1 changes 87 of 640 classifications with zero monotonicity violations. Five normalized monotone functional forms preserve those same 87 endpoint changes, but their intermediate paths differ sharply. A deterministic discrimination audit now identifies salience 0.75 as the only tested single interior level that separates all six aggregate signatures, although the closest pair differs by only 6 of 640 classifications. Identity is therefore decision-relevant inside the authored model, while its construct, coefficient and dose-response curve remain empirically unidentified; existence-versus-null should be tested before costly close-law selection.
The threshold investigations add a subtler result. Delay affects relational outcomes even when the declared direct relational coefficients are unchanged, because timing changes endogenous threshold-crossing opportunities. The original separability claim was falsified by 620 adjacent comparisons. Threshold position controls which paths are exposed, while gain scale controls magnitude; material 40 is not a privileged universal boundary. A locally executed smooth-transition comparator preserved every nonzero hard-threshold direction with no reversals, but this result still awaits publication in the authoritative Sourcery lane.
A new 2x2 component factorial then separated the learning and health increments hidden inside the bundled surplus-learning influence. Across hard and width-2 smooth transitions, learning alone reproduced every active relational delay contrast and the complete effect magnitude; health alone produced zero, and the learning-health interaction was exactly zero. The path mechanism is therefore narrower than its name: it is a learning-opportunity path in the current outcome algebra, not a demonstrated joint learning-and-health path.
Product capability remains incomplete. The workbench service can describe, run, replay and execute bounded ensembles with content-addressed evidence, but it is not yet bound into the rendered declarative application. Continuous assurance is also pending: the old registry verifier scales poorly, while an optimized successor has been installed, independently qualified and admitted but cannot be loaded until another active lane releases its lease.
Empirical calibration is the decisive external gap. A privacy-safe cohort of 248 IFRC events is feasible, but the preregistered GLIDE linkage failed coverage and positive-control accuracy. A structural identifiability audit adds an important correction: even a perfect event crosswalk would leave all six causal mechanisms unidentified because locality, channel, action timestamps, baseline, outcomes, selection and concurrent interventions are absent from the retained projection. The next admissible work is therefore two-track: a governed IFRC-only action-time feasibility pilot plus source-provided identifiers for later multi-source enrichment. Outcome extraction remains closed.
The physics bridge is separately governed. The local corpus now has a ten-stage executable chain for the critical 0-minus torsion observable, including finite geometry, a primary-source apparatus audit, experimental coupling normalization and an explicit negative Spin(2,3)-specificity result. None of that turns Worldwright variables into physical fields or transfers physics authority to the social model.
The machine-readable landscape and its structural verifier are worldwright-programme-current-landscape-v2.json and verifyworldwrightprogrammecurrentlandscape_v2.mjs.