Full-history performance on the actual child domains
The arithmetic candidate passes all 24 native controls, reproducing 348 bounds exactly. Its matched short timing screen passes with a 9–11% improvement. Test whether that benefit survives full-history propagation before another complete prediction comparison.
Use sector zero from each of the six original model/processing groups, all 1,600 packets and the original 7.2-second query. Retain each complete-comparison child domain and its independently verified parent training certificate. Rebuild the immutable subdomain proof; do not substitute a broader physical-only prior. Three of these source/domain combinations previously timed out. They are new implementation-qualification identities, not retries of the completed study.
Require all source/reference/column checks, typed center transfer within the existing 1-micrometre limit, and identical saved predecessor fields wherever a complete predecessor bound exists. Missing predecessor results stay missing. Do not manufacture equality or infer a complete sector union from six sources.
Separate finite allowance
- Six identities; six bounds; 76,800 logical bound steps.
- Six reference/proof setups and 36 exact parent-support verifications.
- Six source and derivative setups; reserve 9,600 appends and body bounds for each relevant stream, charged before execution.
- Four single-thread workers; four numerical threads total.
- The original 120-second cooperative bound ceiling; 600 seconds overall.
- No retries, LP proposals, native trajectory oracles, observed fits, aircraft originals, protected ESA data, allowance expansion or process signals.
Freeze source, inputs, compiler and environment after interface tests pass. Preserve all failures. Passing this screen establishes only finite performance and compatibility on these six cases. Complete prediction widths, recording scale and observational uncertainty remain separate gates.