The results are no longer theoretical.

Question

Can present state look the same while the structure of what remains possible carries additional predictive or causal information?

Strongest result · three tracks kept separate

C-MAPSS · PREDICTIVE SIGNAL: FD003 RMSE 39.0971→26.2617 (+32.83%); FD004 50.1103→47.5843 (+5.04%). FD002 was OBJECT_FAILURE with no scientific outcome.

RTG · MECHANISM SUPPORTED: controlled equal-budget RTS24 intervention twin, 30 selected states, 29 nonzero CHALLENGE states, T = 0.5624591503267974, 95% CI [0.3326645453503039, 0.7558825450242793].

RTG · MECHANISM REPLICATED: Gen448 on independent system pglib_opf_case60_c, T = 0.21785206258890463, 95% CI [0.11979333549127599, 0.34360525563157146]; independent full-network recomputation agreed.

The frozen RTG statistic is D_i = sign(Delta_PROBE_i) * Delta_CHALLENGE_i; T = mean_i(D_i) across the 30 sealed states.

Question

Are some conditions gates that must be satisfied before comparison is allowed to begin?

Strongest result

A formal noncompensatory admissibility object and bounded diagnostic program survive, but the corrective physical heldout did not distinguish SAS from the strongest comparator because both had zero false admissions. Net6 endpoint disagreement 0/510; Net3 2/22; Anytown 2–4/17.

Question

Can individually strong parts fail to produce a coherent whole when required relationships or orientation fail?

Strongest result

Formal proof and counterexample results remain intact. Separately, Orientation v2 selected orientation at N=11 and retained the selected direction on untouched N=12; each retained 34 positive, 4 negative, and 2 zero-margin selected classes, and 4,718,592 observer checks passed.

Question

What exact conclusion is an evidence state allowed to support?

Strongest result

Four frozen external histories/classes were represented with 0/4 normalization disagreements, 0/4 core semantic changes, and 0/4 unregistered claim classes. Separately, the whole-audit quotient is countably infinite and the pointed local quotient has five classes with minimum three bits; 256 is sufficient but not minimal.

Question

Can an agent complete verified work under a known definition of done, not merely describe how it might be done?

Strongest result

Frozen 24-task Study 1: Diamond Gate verified completions 10/24; tested OpenHands configuration 0/24. DG-only completions occurred across all four preregistered strata. Comparator end-to-end tool-loop health remained unresolved.

Question

Which safeguards actually change an operational result, and which merely add ceremony?

Strongest result

Gen449 compared seven configurations across 252 arm-cases. FULL achieved 36/36 correct next actions and exact recoveries, versus BASELINE’s 24/36 and 12/36. Component removals distinguished 256, NEWS, and ACK_PACK; PSALM was NOT_DISTINGUISHED and Fibonacci host utility was UNEVALUABLE.

Question

Can another operator recover what happened, what evidence exists, what failed, what remains unresolved, and what may happen next?

Strongest result

Bounded internal workflows preserve exact-head identities, durable receipts, negative/null/unresolved states, frozen task/protocol identities, and enough durable state for exact-state succession and recovery.