Check the direct source bound over full generated histories
Use the same six 1,600-packet sources and sector-zero child domains as the completed integer-arithmetic performance screen. Retain each original parent training certificate and its six support checks. No future measurement may constrain calibration. All three physical models and both processing cases must pass.
Compare with all six saved predecessor bounds: numerical predictions, common coefficient matrices, bindings, shared caps and observation assumptions must remain identical. Every coordinate, position, force and lever approximation remainder must be no larger. This is a usefulness/performance preflight, not the complete eight-sector prediction union.
Separate finite allowance
- Six identities and six propagation bounds; 86,400 logical bound steps.
- Six setups and six inherited training proofs, with 36 parent support checks.
- Six setups each for numerical source rows, domain derivatives and point derivatives; 9,600 steps each, 9,600 direct-formula packet steps and 9,600 source-body bounds.
- Four workers and one numerical thread each; four total threads.
- Preserve the 120-second cooperative limit per bound; 600 seconds overall.
- No retries, native trajectory oracles, LP proposals, optimizer starts, aircraft reads, protected ESA data, signals or allowance expansion.
Freeze the complete inputs, sources, environment and compiler only after native qualification passes. All six comparisons and finite work checks must pass before a new complete width comparison is prepared. This does not establish useful recording-scale ranges, whole-program floating execution bounds or observational confidence.