Saved native-source binding for training summaries
Check all six sealed generated sources: three candidates under both processing assumptions. Reuse the original five training epochs at 2, 10, 30, 180 and 290 seconds. Extract only their saved observations, native centers/Jacobians, nonlinear bound reports and final training region from the closed staged study. Verify their hashes and preserve the source, frame/clock/processing reference, physical box, basis, coordinate scales and all expression columns.
For each source:
- Project onto each of its fifteen coordinate rows; reproduce all thirty original signed inequalities exactly.
- Construct twelve time-only, unit-length polynomial summary rows. Retain the original deterministic one-metre error premise, using absolute-weight bounds.
- Independently reconstruct each projected row with rational arithmetic, including unshared nonlinear remainders and coefficient-rounding error.
Finite allowance
- Six identities, one worker, one numerical thread and at most 120 seconds.
- Zero original-aircraft reads, native integrations, optimizer starts or fits.
- No LP proposals, retries, new observations or later-position calculations.
- Shared source/input/environment freeze, journals, hashes and sealed results.
These controls qualify source binding and row arithmetic. They do not turn the five-epoch problem into the new complete-cadence problem, apply a different noise law to its original observations, or establish useful prediction widths. Preserve the earlier ambiguity and the 20/20/50-metre targets unchanged.