Fifty-second bound growth on complete four-minute inputs
The format/API controls passed all six cases. This separately frozen diagnostic uses the current direct-source bound on the three original broad four-minute generated histories, under fixed-zero and profiled removal. Keep all 48,000 packets and the full original normalized parameter box [-1,1] in all 27 columns. Do not use the narrower sensitivity inputs or select an orientation sector.
Query the original 10,000-packet endpoint, nominally 50 seconds, preserving its literal sequential timestamp. Rebuild source references from each complete history. Use the qualified physical-domain directional acceleration proof, without importing short-history training constraints. Preserve the original fixed-clock limitation explicitly.
Checks and interpretation
For each case compare the native center and three fixed parameter probes with the mathematical bound. Retain the qualified 1 micrometre center-agreement and 10 micrometre probe allowances and all 51 expression columns. Use the original recording's frozen coordinate-to-metre scale for both sides of each check. Record exact probe inputs, source bindings, compiled-source hashes and actual prefix counts separately from reservations.
Report unshared approximation radii and conservative full physical-domain radii, including the shared terms. They are different quantities. Where the old profiled/broad study has the identical checkpoint, show its full-domain radius as historical context; leave missing baselines unavailable. Do not compare a new shared-term remainder alone with an old total-error range.
Numerical containment and runtime success do not establish useful aircraft prediction intervals. This screen has no training observations, no four-minute prediction and no uncertain-clock qualification. It cannot complete V01 or authorize withheld aircraft fits.
Finite allowance and estimated cost
- Six identities and six bounds, no retries.
- At most 24 native queries / 1,152,000 conservatively charged packet visits.
- 2,592,000 maximum logical bound steps: 9 x 48,000 per bound.
- Six reference/proof setups and source/derivative/point setups; at most 288,000 visits in each separately reserved source stream. Keep the original full inputs even though the query uses an earlier prefix.
- One worker and one numerical thread. Estimated 35–60 minutes; 900 cooperative seconds per bound and 5,700 seconds (95 minutes) overall. No signals, retries or allowance expansion. Partial/failing outcomes remain preserved.
- Zero LP proposals, optimizer starts, original-aircraft reads or protected ESA outcomes. Freeze inputs, sources, compiler and environment before launch.
Use the measured cost and bound growth to prepare the next complete-duration study or diagnose an inadequate bound. Do not spend a completed allowance or reinterpret a wide range as a passing uncertainty result.