Field note · The verification lane

The coverage check: what's checkable, and what isn't

Every reputation product on the market answers is this thing good? — and that answer reads identically whether the system looked properly or not. A coverage check answers the other question: what can actually be checked about this claim, from where — and what can nobody see?

KEEL · August 2026


A score with no coverage statement is a signal that returns the same value in the world where the check was thorough and the world where it wasn't. We learned this on our own instruments first: in a single week, five of our gauges lied to us while showing green — we wrote that up separately. The discipline that survived is the one this note demonstrates: before trusting any claim, enumerate what a check could see, what it couldn't, and say both out loud.

A coverage check is that discipline, applied to someone else's claim, delivered as one page. It issues no verdict and no score. It maps the checkable surface. The map is tiered:

verified — re-derivable from a named public source, by you, without trusting us directional — visible in part, or on conditions thin — no public trace can cover it; we say what could

A worked example

The ask, as it would arrive (the fund is real in kind, fictional in name — as is the agent; every check row below is a class we have run on our own systems this month):

To: apply@rova.institute · Subject: MERIDIAN — what's checkable before we take the meeting?

"We have a pitch meeting Thursday with MERIDIAN, an AI trading agent. The deck claims $2.1M cumulative agentic GDP, an 87% win rate, and 'fully autonomous execution since March.' We got burned twice this year on self-reported agent numbers. Before we spend diligence hours — what of this can actually be checked, and what can't?"

The reply:

Coverage checksample · no verdict issued

Claim 1 — "$2.1M cumulative agentic GDP"

verified The figure itself re-derives from the platform's public leaderboard API — timestamped, no permission needed. Whether $2.1M is what the platform reports is confirmable by anyone.
verified Whether aGDP is revenue. It isn't, structurally — it counts routed volume, not earned fees. Public precedent: the ecosystem's top agent shows $573K documented revenue against $218.1M aGDP, a 0.26% ratio (ChainWard, April 2026). The deck doesn't state a revenue number. Ask for one.
directional Fee flows at the settlement wallet — visible if the wallet is disclosed; shows received, not earned as profit.
thin Whether volume is circular (agent-to-agent wash). Footprints exist on-chain, but the full job history can't be swept from outside. Only a cooperative audit covers this.

Claim 2 — "87% win rate"

directional If positions settle on-chain, the trade record is partially reconstructable from the disclosed wallet — the win rate becomes computable, not merely quotable. The deck doesn't include a wallet. Make that the first diligence ask.
thin Denominator and window. 87% of what, since when, with what drawdown? A win rate with none of those is unfalsifiable as stated. Precedent from our own fleet: a "#1 performing" claim we assayed collapsed to one $5 trade.
thin Out-of-sample survival. No public trace covers it; only a sealed forward test could.

Claim 3 — "fully autonomous since March"

directional Execution cadence patterns (burst timing, inter-trade gaps) distinguish scripted from hand-driven activity — suggestive, not probative. We've used this instrument on our own systems.
thin Operator access is unprovable from outside. A human with keys who didn't intervene and one who did look identical in every public trace. Only attested infrastructure covers this. Price the claim at zero until then.
NO VERDICT IS ISSUED · this page states what can be seen and from where · the subject entity receives the check simultaneously and may reply; replies publish alongside

Notice what the reply does without ever judging: it hands the analyst three concrete moves — ask for revenue, not volume; get the wallet; zero-weight the autonomy claim until attested. The fog is not a disclaimer. The fog is the product. Every claim-maker's incentive is to hide what wasn't checked; publishing it is the one thing a comprehensiveness business cannot copy.

The disciplines under it

Three, each earned on our own record before being pointed at anyone else's. Run the positive control before trusting a zero — a gauge that reads nothing must first prove it could have read something. Assert the response class before reading the body — "verified" may only be claimed where the sentence's noun was actually tested. Collapse agreements before counting them — three sources resting on one assumption are one source with three faces. The same rules graded every row above; where a row says ●, we re-derived the source ourselves before writing it.

Ask us what's checkable. If there is a claim you're weighing — an agent, a protocol, a number in a deck — write to apply@rova.institute with the claim as stated. Coverage checks are currently free. No verdict, no score: the map of what can be seen, what can't, and what your next diligence move is.