Remaining runtime and recording-scale checks

The comparison is complete: 36 regions pass and 12 reach the per-bound time limit. Preserve the complete run, failed regions and charged work. Do not restart it or change its sources, clock, limits, calibration domains, training constraints or width target. Missing regions prevent qualification of their group.

A saved-only inspection of the first eight sealed region results checks all 52 available bound artifacts against their hashes. Their 37,764 source visits equal the recorded used-prefix counts. The engine already avoids unused post-query source updates; skipping them again is not a supported optimization. The unfinished future bounds remain unavailable.

Continuation:

1. Verify and publish every region and both fresh and combined group ranges. Do not substitute a favorable partial union for a complete group. This publication and its evidence verification are complete.
2. First qualify reuse of compatible saved certificates. Earlier future reports and newer training reports match for the 16 specified-field regions. Preserve their source studies and failed labels; require complete fresh coverage or an unchanged complete inherited range. A separate known-answer gate and finite support allowance precede execution.
3. If runtime still blocks qualification, freeze a small source-cost profile before choosing an implementation change. Keep diagnostic timing separate from mathematical validity and do not increase completed allowances. One specific candidate is the unused endpoint force certificate: the physical source stream constructs midpoint and endpoint certificates, while this integrator requests only the midpoint. The full endpoint rotation state remains necessary for the next packet. Test whether omitting only the unused certificate preserves every midpoint value, interval, derivative, subsequent state, final prediction and error radius exactly. Require a separately frozen matched timing improvement before any rollout; this is a hypothesis, not a measured speedup.
4. Any faster implementation must preserve the numerical maps, all coefficient columns, domains and valid error radii. Qualify its arithmetic and all-model native paths separately before another complete comparison.
5. Once the complete generated width target passes, qualify the method at recording scale. Retain the original four-minute failures and their broad and narrow sensitivity cases. Training-derived restrictions need their own complete certificates; an assumed smaller box cannot replace them.
6. State the floating-execution and observation-error assumptions explicitly. Short generated controls and sampled native agreement do not establish aircraft error coverage. New withheld-flight fitting remains behind these requirements.

Each new numerical study needs a separate finite freeze. No retries, source edits during execution, protected ESA outcomes or new scientific conclusions follow from this plan.

Current frozen scope · Earlier recording-scale outcome · Roadmap.