Complete training-conditioned prediction comparison

The new source calculation passes native qualification, all 48 source-domain preflight checks and the six complete joint-adapter known-answer controls. This study now tests usefulness on the complete original generated problem.

Keep all six model/processing groups and eight mounting sectors: 48 fixed identities. Reuse the six 1,600-packet inputs, training at 0.8, 1.4, 2.4, 3.6, 4.8 and 6.0 seconds and prediction at 7.2 seconds. These are software controls, not aircraft recordings or new scientific replications.

For every region, verify the original parent support certificate and its unchanged source. Apply the checked subdomain transfer without weakening the training information or changing a calibration cap. Recompute six training bounds and one future bound. Retain all 51 expression columns and all 78 enclosing coordinates. Transfer reference-center rounding outward under the existing 1-micrometre agreement limit.

Preserve all inherited constraints. Add the independently reconstructed training rows with the original 1 mm generated observation allowance and unchanged arithmetic margin. Use at most eight contraction passes and the existing 1,024-row/78-column ceiling. Audit the six future endpoints independently. Only after the range or empty proof is saved may the adapter perform its known-calibration and future-truth audit. Failed or unhelpful regions remain.

Union every surviving sector in each group. Publish both the fresh ranges and their intersection with the qualified 192-box comparison. Every original axis of every complete group must meet the unchanged twofold-width target. Missing or failed regions prevent qualification; a partial improvement does not pass.

Separate finite allowance and cost

• 48 identities, 336 propagation bounds and 4,300,800 logical bound steps.
• 48 reference setups, 48 restricted training proofs and 288 exact parent support verifications.
• 336 source setups, 537,600 source-reference steps and 537,600 body bounds.
• Separately reserve 336 derivative setups and 537,600 derivative appends.
• 1,728 independent training-row checks, 48 contractions and 48 truth audits.
• At most 288 one-second LP proposals and 1,152 helper/independent future support checks, in addition to the 288 parent verifications.
• Four single-thread workers; each bound retains its cooperative 120-second limit. The separate overall ceiling is 7,200 seconds.
• No retries, allowance expansion, process signals, native trajectory oracles, model fitting, aircraft-original reads or protected ESA observations.

The measured future bounds take 54–99 seconds. Scaling their packet counts to the seven queries gives roughly 65–80 minutes, with a two-hour ceiling for variation and support work. This is a new allowance; no completed study's unused work is reused. Freeze code, inputs, parent certificates, targets, compiler and environment before launch. The coordinator packages the result after completion without refitting or committing.

Passing establishes usefulness on this short generated example. Recording-scale error bounds, floating-execution enclosure, observational adequacy and calibrated confidence remain separate gates before new aircraft conclusions.

Joint-adapter qualification · Source/domain preflight · Earlier complete comparison · Roadmap.