01
Ask with the brief you have
An email with three candidates, five objectives, two hard limits, and a few worries. SEAL reads it as written. Public figures from GAO reports and Army articles keep the synthetic numbers in a plausible range; they are anchors, not program data.
02
Elicit into a schema
Objectives, options, constraints, assumptions, and risks become candidate objects with a confidence and a locator back to the sentence. Nothing is accepted for you. Reject a candidate and it never enters the run.
03
Relate in a graph
The platform, its mobility, protection, powertrain, and budget aspects, the tests that measured them, the requirements they satisfy, the standards they reference, and the options themselves share one graph. Select a node and read the quoted evidence with its hash.
04
Plan and lock the weights
Weights are integers in thousandths. They must sum to exactly 1000 or the desk will not run. The plan shows which steps are done and which are waiting on you.
05
Run a reproducible evaluation
Each criterion is min-max normalized across the feasible options, then weighted. Hard-constraint failures are flagged and ranked last, never dropped. A missing figure stays a blank cell. The run carries a schema hash, a weight hash, an evidence hash, and a run hash; the same inputs give the same run hash every time.
06
Read what flips the decision
One weight or one figure moves at a time until rank 1 changes. The result is an exact threshold and a plain sentence: which option overtakes which, at what value, from what value today.
07
Assess readiness before anyone signs
Five bias checks with a metric and a threshold each. Assumptions carry a status: stated, tested, or invalidated. Open high risks count against the score. Below 70, the signature is blocked and the desk lists why.
08
Sign a package that reproduces
The package holds the recommendation, every accepted object, the scores, the flips, the readiness, the evidence list, and the hashes. Download it as Markdown or JSON. The package hash changes when the signer does.
09
Refresh when conditions change
Longhaul carries the same family through four quarters. A test result arrives; a cost estimate rises; a build gains weight; an assumption is invalidated. Each refresh shows the ranking side by side, the Kendall tau, the evidence added, and the old and new hashes.