Rather than describe the options, I ran each one far enough to know if it’s real. One got built and works, one failed its own premise-check, one got answered by doing the first. This is the honest comparison — and the decision it leaves you.
The short version. A (build it ourselves) is done and running on real fleet data — open it. B (layer over ChainWard) is a no: the vendor doesn’t cover us and we can read the same ground truth cleaner ourselves. C (spec first) became moot the moment A worked. So the real decision isn’t which approach — it’s what’s next for A.
The day after this comparison, a full onchain census of Virtuals ACP on Base answered the question this doc left open. The demand the reader measured — the “274 jobs,” the “2–4 buyers per agent,” the “$13 of demand” — was airdrop-manufactured wash-farming, not organic buyers: platform-wide, payouts peaked near 1.14M/month and collapsed to ~0 by Apr–May 2026 when the airdrop points ended, with ~$112M cycled in and out for roughly $5 net. Wash is confirmed; laundering is not established. So the v2-a “buyer honesty” lean below was the right question — and the census is its answer: the buyers were not independent. Read what follows as a record of the build-method decision (which still holds), not as a live read of demand.
Status: done, with real data. Reads ACP Core’s settlement events on Base, traces each agent’s real USDC through the shared PaymentManager, and reconciles it against the dashboard’s claims. Validated end-to-end on Clarity, then run across all 5 fleet agents. It found the platform’s hidden 20–28% take and the free-job gap the dashboard conflates.
Status: checked, and it’s a no. The idea was to skip building a reader and consume ChainWard (the existing onchain-audit vendor) as our data source. Three findings killed it: ChainWard has no coverage of our agents (its forensic decodes are whales — Clarity/Mercury/Breakbot/Cleo all 404), it’s Base-only so it’d miss our Hyperliquid trading agents entirely, and it’s a single unaudited vendor. Meanwhile Option A proved we can read the identical ground truth ourselves, free.
Kept as a reference, not a dependency — ChainWard’s decodes are a useful second opinion on whale agents.
Status: mostly moot. The spec-first path exists to de-risk a build before committing. But the recon + Option A already answered every question a spec would have: the data is free and readable, the identity-resolution method is validated, the event schema is known, the fee structure is measured. A spec now would mostly be documentation of A. What genuinely remains spec-worthy is the v2 boundary — and that’s captured below as the actual next-decision.
| Dimension | A — our reads | B — ChainWard layer | C — spec first |
|---|---|---|---|
| Covers our fleet? | Yes, all 5, verified | No (0 coverage) | n/a (no build) |
| Sees trading agents (HL)? | v2 — buildable | Never (Base-only) | n/a |
| Dependency | None — we own it | Single vendor | None |
| Effort remaining | Done (v1); v2 optional | Wasted (dead end) | Re-docs A |
| Trustworthiness | Exact onchain read | Vendor-mediated | n/a |
Building it ourselves won on every dimension that matters, and it’s already done. B was a dead end that the premise-check caught before it cost anything. C would just describe A. So the three-way comparison collapses cleanly to: A is the path, and v1 exists.
The comparison you asked for has a winner — but the winner is already built, so what’s left to decide is how far to take it.
Ship v1 as-is (it’s a decide-with instrument, not a product), and pick one v2 direction only if it earns its keep. Don’t gold-plate a $13-of-demand reader; take it further only where it changes a decision.
Are the 2–4 “buyers” per agent independent, or related wallets? The demand-quality question the dashboard can’t answer. Highest insight-per-hour.
Add the Hyperliquid read so BASIS / DegenClaw stop showing dark. Makes the fleet picture whole. Needed before trusting it fleet-wide.
Deploy as an auto-refreshing board (frend/claude_b lane) instead of a baked snapshot. Only if it becomes a recurring read.
Run it on competitors’ claimed-vs-verified (it generalises). Turns the honest instrument into a competitive lens. Optional, higher-ambition.
My lean, if you want one: v2-a (buyer honesty) first — it’s the cheapest and it’s the one dimension that could change what “274 jobs” actually means. Everything else can wait for a reason to exist.