Senior product marketing · one run, on the record
I'm a senior product marketer. This is the operating system I use, run in the open on a launch I had no part in, at a company I'm not applying to. Public information only. Every step is below, including the edits and what an outside reviewer found wrong with them.
Trigger
On 19 August 2026 Ramp launched Router: one API endpoint that sends each AI request to the lowest-cost model that meets your quality bar. Free through 2026, a $26 credit, U.S. only, enterprise features "coming soon." The same day, Stripe announced it was buying OpenRouter; Palo Alto Networks had closed on Portkey in May. Publicly verified
Sources: Ramp's release (PRNewswire, 19 Aug 2026) · TechCrunch (20 Aug 2026) · router.com · Stripe newsroom (19 Aug 2026) · Palo Alto Networks releases (30 Apr, 29 May 2026).
Frame
Four artifacts answer that in the order the system produces them: product truth, then buyer and positioning, then the launch brief and workback, then the measurement tree. Each one is open at the bottom of this page.
The system scored four companies on a 100-point rubric and recommended the top scorer: Ramp 95, Deel 88, Gusto 83, and a security vendor at 71 (name withheld: an application of mine is open there). I approved the recommendation rather than choosing against it. I also set the rule that a company I'm applying to is never used as public proof, so the run had to be somewhere I had no stake.
Product truth
Two of the nine will never be safe to state as fact from outside: the 40% savings figure and a model count. Two others were held at first pass and cleared later on primary sources. Every row names who inside could settle it.
The full ledger, with owners named for every missing internal input, is in the Product Truth Brief.
Graded evidence
Thirty-eight rows graded before any edit was made: the claims ledger.
Positioning · the durability check
The test is the same for each: does the claim survive a side-by-side against every competitor's own public copy? The positioning document ran that check, stated one hypothesis, and selected one line for the technical evaluator. Its first version, and this page's, said those two lines were the system's own invention. Ramp's launch post says otherwise. The correction is on the record here, in the document, and in the Positioning room.
Killed
"Cut your AI bill 40%."
Ramp's own headline. From the document: "Router's headline claim fails this check: it is claimed by at least five other vendors, not zero. This means the 40%-cheaper / lowest-cost-model claim cannot be the durable differentiator this messaging system leans on." Azure AI Foundry's Model Router, Bifrost, Orq.ai, Inworld and OpenRouter all say it. And the 40% is Ramp's number, with no cohort or window behind it.
Ramp's own release, pushed one step · held back by its author
"The only model router that already sees the rest of your company's spend."
Ramp's press release already connects Router to "Ramp's broader AI spend visibility and controls," and router.com says "Loved by CFOs." The document's contribution was to aim that claim at the existing finance buyer as a cross-sell line, then refuse to recommend shipping it: "Recommending the spend-data-bridge message to Ramp today, without knowing whether Router is […] an independent brand-separated product or an eventual cross-sell, would be exactly the kind of internally-generated, buyer-evidence-free positioning Module 02 names as the top PMM failure mode." One internal fact settles it, and it isn't public.
Ramp's own line · selected for the technical evaluator · medium confidence
"We don't build models, so we have no reason to favor any particular one."
Ramp said it first, in its launch post's "Independent by design" section on 19 August 2026. Ramp leads with the cost claim, not this line; the document selected it for the technical evaluator and carried its limit: OpenRouter sells no models either, so it separates Router from Cursor- and Meta-style routers, not from the incumbent. My first draft called this line an OS invention Ramp wasn't running. An independent verifier fetched the blog post and proved that false. What the system contributes is choosing which existing claim goes to which buyer and how far it can be pushed, not inventing a message nobody has said.
The buying-committee map the messages hang from, with a confidence and a "what would raise it" for every role: ICP, positioning and messaging.
The gate · verify, revise, remove
Every material claim was checked against its primary source on 6 September 2026. Where the source said something narrower, the text was revised. Where no source could be found, the text was removed and the removal left visible. These are the actual before-and-after lines.
Removed launch brief · competitive landscape
BeforeOpenRouter, Portkey, and Martian together account for roughly 97.8% of tracked capital raised in the AI-inference-gateway category over the trailing 24 months (~$181M total, ~$153M of it OpenRouter's).
AfterRemoved: a "97.8% of tracked category capital" figure with no named source could not be verified and carried false precision. The verified funding fact is OpenRouter's $113M Series B.
Revised company-selection scorecard · Deel
BeforeDeel confidentially filed an S-1 on 2026-02-15 at a reported $17.3B valuation, with an IPO "as soon as 2026" reported by CNBC. [Publicly verified]
AfterCNBC reported in February 2025 that Deel was preparing for an IPO "as soon as 2026"; the 2026-02-15 confidential filing at $17.3B is asserted by one IPO aggregator and, as of 2026-09-06, corroborated by no major outlet or company statement. [Reasoned hypothesis — the CNBC citation predates the asserted filing by a year and cannot corroborate it]
Revised launch brief · the incumbent
BeforeOpenRouter (unified gateway, 300+ models across 60+ providers; ~$113M Series B)
AfterOpenRouter (unified gateway; "400+ models from more than 80 providers" per Stripe's 2026-08-19 acquisition announcement; $113M Series B at a $1.3B valuation, 2026-05-26, led by CapitalG)
Verified launch brief · Portkey
BeforePortkey (observability/DevOps-focused; acquired by Palo Alto Networks, May 2026)
AfterPortkey (observability/DevOps-focused; acquisition by Palo Alto Networks announced 2026-04-30 and completed 2026-05-29, per Palo Alto Networks' own releases)
Removed ICP · reviewer evidence
BeforeIndependent secondary commentary (a comparative reviews digest) notes reviewers "reported issues with enterprise-level visibility and permissions" … [Reasoned hypothesis — single secondary source]
AfterRemoved: on re-check (a search-snippet pass over the G2 and Capterra Ramp review pages, which block direct fetch; a negative search, not a read of every review) the claim could not be corroborated. A negative claim about a company's product does not ship on one uncorroborated source.
First pass: 2 verified with added precision · 10 revised · 2 removed. Second pass, after the outside review below: 35 more edits. All of them, with the text as it was and as it is
Whose judgment
In plain words: an AI operator executed this run inside my operating system, under my rules and rulings. The judgment that shapes the system and the decision at every gate are mine. The leverage under it is AI.
Outside review
Its brief: refute, don't confirm. Check every edit against its primary source, check the tournament against the record, find any sentence that reads as Ramp's internal data or as work done for Ramp. It read all six documents and fetched seven primary sources. Twenty-three findings the first time: six blocking, thirteen should-fix, four notes. Then a second, independent session ran the gate again and failed the page on one more. Four of them, verbatim, with what changed:
Verdict, verbatim: "Mixed." Six blocking defects, all corrected on the record: a wrong ruling ID, a data-retention claim no source supported, an orphaned funding figure, a "closed beta" inference the gate had already refuted, stale environment notes, and downstream sections that never learned two rows had cleared.
"The ledger's stated authority ruling does not exist as described. Ramp was approved under D-024, not D-022. … A public page whose premise is auditable provenance cannot carry a ruling ID that resolves to a different decision."
Nine occurrences across the five documents and the ledger corrected to the decision that approved Ramp. The mix-up was a known hazard, two decision logs numbered independently, named in the repo the day the documents were written and still missed.
"'opt-out, not opt-in' is not supported by either primary source, and it survives the gate in three places. … This is a data-privacy claim about a named company; 'Publicly verified' is not sustainable on the current record."
All three places now say what the sources say: a one-year default retention window, settings users "can control some of," and a zero-data-retention option on U.S.-hosted models. No opt-out is asserted. The row's label came down.
"The thesis tournament is cleaner than the record: it inverts the ICP document's own hierarchy. … The ICP document never enumerates three theses, never names them, and never uses the word 'tournament.'"
The station above was rewritten to the document's hierarchy, with the document's own sentences quoted. The spend-bridge line is its stated hypothesis, held back by its author; the neutrality line is the one it put in the messaging matrix. The first version made the reconstruction look like the record. (A later verification found the neutrality line was Ramp's own; see below.)
"Neutral-routing thesis false about Ramp in four public places … Refuted by Ramp's blog post (2026-08-19): 'We don't build models, so we have no reason to favor any particular one.' Also fires the page's own station-10 update trigger: Ramp itself already does."
The tournament above was rewritten a second time. The neutrality line and the spend link are both Ramp's own launch messaging, snapshotted as primary sources in the repo. The system's contribution is the selection per buyer, the durability test and the OpenRouter limit, not the line. The Positioning room's example, the ledger and the document all now say so.
Every finding, and the edits they produced, is in the ledger. A second reviewer then read this page as four people: a VP of Marketing with ninety seconds, a recruiter with thirty, a senior PMM peer, and an AI evaluator. They found this section empty, a headline that read as a Ramp employee's postmortem, and no title above the fold. Those are fixed. Their honest limit stands: this page shows process, not execution under pressure or outcomes. That is what the six rooms and the Modus record are for.
Shipped · open the work
What I'd measure · what would change my mind
Illustrative: no Ramp data was accessed. This is the tree I'd build at kickoff.
Qualified routed volume, weekly, from accounts still routing after their first weeks, and deliberately not counting one-time signups chasing the $26 credit. Two measurement eras, because the free period ends. Sales enters at two named signals; most signals have no sales owner, on purpose.