How the work actually thinks
Every experiment here is, underneath, an argument for how a decision-support system ought to behave. These are the arguments, stated plainly — each one already showing up somewhere in the work.
Reasoning is the product, not the number.
TESSA doesn't hand an agent a valuation — it hands over the finding that produced it: the claim, the evidence, the dollar contribution. A number without its reasoning is a guess with better formatting.
Put the judgment where it belongs.
TESSA keeps professional valuation judgment with the agent. NVEE automates bounded remediation and returns to a person when verification fails. M²W² uses deterministic rules for familiar transactions, model reasoning for ambiguous ones, and human correction to turn future ambiguity into knowledge. The right place for judgment depends on risk, reversibility, and context.
The model proposes; the arithmetic doesn't ask permission.
In M²W², a local model suggests where a transaction belongs. Every dollar total is still computed in SQL. Confidence is not the same thing as correctness, and money is a bad place to confuse the two.
Scope is a design decision, not a limitation.
NVEE's MVP explicitly excludes legacy deployments. TESSA explicitly excludes MLS integration. Knowing what a system shouldn't do yet is as much a decision as knowing what it should.
If a system can't explain itself, it isn't finished.
This is the thread running under everything above — not a policy, a design constraint. A system that produces answers without reasons hasn't finished being built. It's just stopped early.