Models — FlowFinds

One engine, and the surfaces built on it

FlowFinds ships one thing that carries a published identifier, and several surfaces that run on the same inference path without one. The distinction is not cosmetic. An identifier is a promise that you can point at a response, name what produced it, and find the recorded evidence for that configuration. We publish one because we have measured one.

commerce-reasoning-v1 is returned on every response the decision engine produces, alongside a run identifier, and it names the directory the benchmark writes its run artifacts to. Every other entry below is described as a surface, and each says on its own page that the benchmark does not cover it. A surface does not borrow the engine’s evidence.

The engine

Identifier	For	Measured
`commerce-reasoning-v1`	The decision engine an operator talks to: it investigates a question with application tools before it answers, and it names the evidence it used.	Yes — by commerce-v1

4 configurations have recorded runs of that engine, over 51 scored cases in total, measured on 2026-09-05. Those figures are read from the run artifacts when this page is built, not written here, so a new run changes the sentence above without anyone editing it. The run-by-run detail is on the results page, and the per-configuration cards are under system cards.

The surfaces

Each of these does real work in the product. None of them is measured by commerce-v1, and each page says so in its own words rather than leaving the omission to be inferred.

Surface	For	Measured
The find engine	Scores a candidate product on three independently sourced legs, and refuses to average them into one comfortable number.	Not covered by commerce-v1
The store editor	Turns an instruction in the operator's own words into a proposed change with a cited reason, a rejected alternative, and a reversibility flag.	Not covered by commerce-v1
The organic research engine	Researches an audience per platform and produces a framing brief. It has no posting function, and no credential it could post with.	Not covered by commerce-v1
The support agent	Answers a customer from one product's own record, and is scanned on the way out for the things it must never say.	Not covered by commerce-v1

The response contract

The engine is not asked to produce prose and then trusted. It returns a structured decision that is validated before anything reaches an operator, and a decision that fails validation is rejected with the reason and re-attempted rather than shown. The fields are the contract.

Field	Shape	What it means
`kind`	tools \| answer \| route \| clarify	Which of four moves the engine is making. A tool request cannot also navigate, and a final response cannot carry unexecuted tool calls.
`message`	string, at most 1,800 characters	What the operator reads. Empty when the engine is navigating, because a navigation is not an answer.
`tool_calls`	at most 4 per turn	Independent tools are requested together. The same tool and product pair may not be requested twice in one session.
`action`	a named destination, or null	Set only for an explicit request to open a named surface. Analysis, advice and comparison are answered in place.
`evidence_ids`	at most 8	The identifiers of the tool results actually supporting the answer. An identifier is provenance, not proof that a causal claim holds.
`uncertainties`	at most 4, each at most 240 characters	Material limitations in plain words. The engine is forbidden to invent a numeric confidence.
`remember`	at most 4 preference keys	Categories of enduring preference the founder explicitly supplied or corrected. The service stores the founder's own words, never the engine's paraphrase.

What is provisional here

The descriptions on these pages were transcribed from the implementation by hand and are marked provisional in their source file until they have been verified against it again. No figure on any of them is transcribed: every benchmark number is computed at build time from the recorded artifacts, which is why none of these pages quotes a score in prose.

How the benchmark is constructed is set out in the commerce-v1 methodology, and you can reproduce our results yourself. What the engine is used for, surface by surface, is under the product.