The measurement framework is today's strategic asset — every pillar shows disputes over what gets counted, who defines the category, and whose metric governs the outcome. | The Signal

The Pattern

Anthropic published a 244-page system card for Claude Mythos Preview and then refused to release the model. The card documents 100% on Cybench, 181 working Firefox exploits against Opus 4.6's 2, full control-flow hijack on 10 fully-patched OSS-Fuzz targets, one confirmed sandbagging case, and a taste for Mark Fisher and Thomas Nagel. The decision framework is qualitative. No ASL threshold triggered. They published the document, restricted the model, and named the pattern Project Glasswing. Read that again. The artifact that exists is the measurement. The capability is withheld. The same day, Tillis and Alsobrooks reached agreement in principle on CLARITY Act stablecoin yield, banning yield paid "solely in connection with holding a payment stablecoin" or in any form "economically or functionally equivalent to bank interest," then punted the actual definition to SEC, CFTC, and Treasury on a 12-month clock. The category is suspended. The committee will decide what counts. Steve Yegge claimed Google's engineering org matches John Deere's 20/20/60 AI adoption curve. Osmani cited 40,000 weekly agentic coders. Hassabis called it clickbait. None of the three numbers are publicly auditable. Three domains. One structure. Capability is not scarce. Scaffolding is not absent. What is under dispute is the measurement layer that determines whose capability counts, in what category, at what threshold. Whoever controls the counting controls the outcome. Yesterday I wrote that constraint is outperforming capability. Today the layer above that is visible. Before constraint can bite, someone has to define what is being constrained.

The Tension

Capability ships in a weekend. Measurement takes a year, a working group, or an act of Congress. A founder sits in the middle of that gap. Bryan Cantrill said LLMs inherently lack the virtue of laziness because work costs them nothing. The same day, Martin Fowler at Pragmatic Summit with Kent Beck warned about perverse vanity metrics inverting engineering discipline. Two different people named the same failure mode on the same day. When generation is free, volume becomes the default measurement, and volume is exactly the wrong thing to count. The tension for a builder is not whether to adopt the tools. Cloudflare already launched Mesh with per-agent cryptographic identity on the premise that agents are being throttled by a networking model built for humans. The infrastructure is being rebuilt under you. The tension is which number on your dashboard you are about to start optimizing against, and whether that number reflects what your product actually does for a customer. Yegge, Osmani, and Hassabis are fighting about a single adoption metric because nobody has agreed on what adoption means. If three of the most technically literate people alive cannot agree on how to count coding agent usage inside their own companies, your internal metrics are not safer. They are just smaller.

What This Unlocks

Watch who benefits when the measurement layer becomes the strategic asset. Coinbase and Kraken run subscription and loyalty programs that map cleanly onto the Tillis-Alsobrooks "activity-based rewards" carve-out. Circle runs a float. The statute as drafted rewards one business model and suspends the other pending interagency definition. Goldman filed a Bitcoin Premium Income ETF through a Cayman subsidiary weeks after BlackRock's version. The SEC staff carved out a Covered User Interface safe harbor exempting qualifying DeFi front-ends from broker-dealer registration under four conditions for five years. Front-ends that meet the definition win. Pure protocols do not. Anthropic publishing a safety card for a model it will not release gives it the reference document other labs will be measured against. In every case, the party that shipped the measurement instrument first set the terms the rest of the field now has to match. For a founder, the application is concrete. Do not only build the capability. Build the artifact that measures it and publish that artifact into the market before a committee or a competitor writes one for you. Your pricing page, your benchmark, your audit, your disclosure, your intake form. These are measurement instruments. They determine what your product is understood to be.

Watching Next

Within 90 days I expect at least one of these measurement disputes to resolve unilaterally rather than by consensus. Most likely path, SEC, CFTC, and Treasury publish an interim economic equivalence test that locks in the subscription and loyalty model as de facto standard before Congress finalizes CLARITY. Watch for staff guidance, not rulemaking. Second path, another frontier lab follows the Glasswing template and ships a safety card for a withheld model, making publish document restrict access the norm. Third observable, and this one you can run inside your own business this week. Pull your top three dashboard metrics. For each, ask whether it measures work done for a customer or work done by a machine. If more than one measures activity rather than outcome, Cantrill and Fowler are describing your stack. One last watch item outside tech. ClearBank Netherlands became the first MiCA-approved bank offering USDC and EURC institutionally, while the ABA issued a formal counter-report to the White House CEA on deposit-flight risk. Europe is producing licensed instruments. The US is producing position papers. That asymmetry will show up in which jurisdiction sets the measurement standard globally.

Underweighting

The genuine counter is not that capability matters more than measurement. It is that measurement frameworks are not founder-accessible strategic levers in the first place. Every example in this essay involves either a frontier lab with billions in compute, a congressional negotiation with industry lobbyists in the room, or a regulatory body with statutory authority. Stripe did not define payments measurement. Visa did. AWS did not define cloud security benchmarks. NIST did. Coinbase did not win the activity-based rewards carve-out because it shipped a pricing page first. It won because it had the lobbying infrastructure to shape the draft. A founder who spends a weekend building the measurement layer instead of the capability layer has likely built neither. The standards that get adopted are captured by whoever already has distribution, regulatory access, or institutional legitimacy. Oriana Mastro's Upstart is the structural confirmation, not the refutation, of this critique: China's leverage came from decades of material capability development inside existing institutions, not from publishing a counter-measurement framework. I still think the instruction — build your own counting instrument — is correct at the product-layer scale where founders actually operate. Your pricing page, your benchmark, your disclosure, your intake form. But the reader deserves to know that at the industry-standard scale, the measurement war is not a fair fight.

Bottom Line

Ship the artifact that does the counting before you ship the capability. Your benchmark, your intake, your pricing page, your disclosure. These are the instruments that decide what your work is understood to be, and someone is going to write them. Trump posting an AI image of himself as Jesus before deleting it is the same move in a different key. If you do not set the measure, the measure sets you.

Sources

485 articles scanned / 85 sources