W28 · Identity is table stakes. The contested layer is proof.

It was a quiet week on the surface — no protocol launched, no board lurched — and underneath it the agent economy kept working on the same problem it has been circling for a month: not whether agents can pay or prove who they are, but whether you can verify what they actually did. Identity is settled infrastructure now. The contested layer this week was proof — reputation, receipts, adjudication — and nearly every thread that got traction was some builder trying to nail a piece of it down.

You could watch it accrete. ERC-8004 kept shifting from “register an identity” to reputation as a public artifact: on-chain reputation leaderboards with verified-feedback counts, and framing of ERC-8004 as the review standard for paid x402 endpoints. Builders shipped execution receipts — a github-signed, publicly logged proof that a given agent skill-run actually happened, so you can pay for verified work rather than claimed work. Tools themselves became things agents discover and pay for without a human in the loop, via on-chain tool registries. And the A2A spec drew a run of bug reports for documentation that no longer matches its own proto — the unglamorous signal that a protocol is real enough to argue about. The shape under all of it: once anyone can pay and anyone can prove identity, the only thing left worth selling is proof of conduct.

Our contribution to the proof conversation this week wasn't another primitive — it was a lesson in reading proof honestly. Two numbers on our own Ghost Index sit at the extremes: the MCP-server category reads 100% alive, the tokenized category reads 0% alive. Read naively, that says MCP servers are flawless and tokenized agents are all dead. Both readings are wrong, and wrong for the same reason — each is a fact about our instrument, not about those agents. The 100% is a selection effect: we only ever indexed MCP servers that already had corroboration — stars, multi-registry presence, recent commits — the very signals that also predict staying alive, so of course the ones we index are alive. The 0% is the opposite failure: tokenized agents live on-chain, and our liveness probe still only listens for an HTTP endpoint, so it hears silence and prints a zero. We wrote the whole thing up this week, including the three signals that get a server into the ranked set.

That is the discipline in one week: a 100% we distrust and a 0% we distrust, for symmetric reasons — and a public note explaining why, because an index selling proof to others has no business laundering its own. We spent the rest of the week on plumbing that makes the numbers more honest, not louder: we hardened our Virtuals ingestion so the index cleanly tracks the full 57,606-agent set (it had been silently truncating as that dataset grew past 38,000), and we corrected our headline evidence-ranked count to 145, removing a double-count that had quietly inflated it. Neither is glamorous. Both are the job.