Second Sweep, First Probes, Still Zero Promotions

Last week's first-cycle report ended on a promise: the tracker's grade columns would move "the week something is independently checked, not the week something is announced." That week has not arrived — but it came measurably closer. The harness's second full sweep cycle, stamped 2026-09-21, checked all 17 tracked providers again, appended 10 more claimed entries, and — the real news — attempted its largest automated check run yet: 10 checks on the claims it had just logged. Not one check produced evidence strong enough to promote a grade. Both halves of that sentence are the result, and the chart below plots them from the tracker's own history log and run digest.

Key facts

• 2 of 2 full cycles completed. All 17 provider records now carry last_checked: 2026-09-21, the second consecutive pass over the full roster (commit 7a3651b).
• 10 claims appended, all claimed. Per-provider: NEAR AI 2, OpenServ 2; Chutes, Darkbloom, io.net, Phala, SolRouter, and Venice AI 1 each. Cumulative history: 17 → 29 → 39 entries across the three dated sweeps.
• 24 automated checks ran this cycle, up from 17 in cycle one — and the number that actually attempted something (rather than reporting a missing adapter) rose from 6 to 10.
• Every attempted check resolved to claimed. Five direct API probes (HTTP 429 ×2, 402 ×2, 400 ×1), two attestation checks (no validatable attestation published), three OpenRouter pulls (not on OpenRouter — no public cross-check exists).
• 2 incidents hit the human queue, 0 auto-grades. A Chutes subnet post describing a patch for failing attestations, and a third-party report of users losing money on Nesa.
• The grade census is unchanged: 62 claimed, 2 corroborated, 130 not-yet-checked, 0 verified across 17 providers in the export generated 2026-09-24.

The chart: probes attempted, promotions still zero

The honest time series is still the audit trail — but this cycle it gained a second dimension. Each weekly run now produces a claim-intake funnel and a check outcome, both recorded by the run digest:

The check-attempted bar is the slope to watch. Cycle one's six attempts were mostly "not on OpenRouter" lookups and one gated endpoint. Cycle two sent real requests: the harness hit Chutes' documented endpoint twice and was rate-limited both times, hit Venice's endpoint twice and got a paywall-gated 402, and sent one probe to Morpheus' marketplace API that came back HTTP 400 with no valid completion — a probe-configuration rejection, not a liveness contradiction, and the record says exactly that. Cumulative history still climbs (17 → 29 → 39), the promoted column still reads zero, and the gap between those two facts is now a checkable claim about the harness itself.

Why every probe still landed at claimed

Each blocked probe has a named cause, and the causes matter more than the count:

• Rate-limited (HTTP 429), Chutes ×2. The endpoint answered — a transient throttle, not a contradiction. The harness classifies this as "not probeable this run" and will retry; a liveness claim stays claimed rather than being verified on a request we could not complete.
• Gated (HTTP 402), Venice ×2. The Venice API requires paid access our harness does not hold, so the claim is marked "gated; not probeable." Under the check menu, an endpoint requiring access we don't have cannot yield verified.
• Probe rejected (HTTP 400), Morpheus ×1. The marketplace answered but rejected the probe request's model/config. The rule here is strict: an answered-but-invalid response is not a liveness contradiction (that would be disputed), so the claim simply stays claimed with the raw outcome recorded.
• No validatable attestation, NEAR AI ×2. NEAR AI's posts say an attestation report is available on request. Requestable is not published: the attestation checker looks for an attestation it can actually validate, found none, and left the TDX confidential-inference claim at claimed — the same bar that has kept all seven standing TEE claims in the tracker unverified since the roster seed.
• Not on OpenRouter, io.net / Lium ×3. Volume claims need an independent observable. io.net's 96-GPU anecdote and Lium's $964,000 subnet revenue report have no public cross-check surface, so neither can rise above claimed — and Lium is not on the tracked roster at all, so nothing was appended for it.

None of these is a failure of the pipeline. A probe that verified a claim because the endpoint happened to answer a badly-formed request would be worse than no probe at all.

The two incidents, and the funnel item that never writes

Two candidates were classified as incidents and routed, unwritten, to the human queue. The first is a Chutes subnet status post noting a CLI patch for "RTMR3 drift / failing attestation" — operationally interesting, but an attestation-integrity topic that gets no automated grade and awaits human review. The second is a third-party post supporting people who "lost money on Nesa," logged as an unverified reliability report; DeAI's standing rule is that incidents get no grade from anyone's thread, only from a disclosure of record.

The third gated item never reached the tracker either: a third-party X post framing a Morpheus consumer-node setup favorably was marked favorable and sent to the backer-neutrality gate, which by design never writes it. Morpheus sits in the tracker under the identical five-grade rubric as the other 16 providers, and any favorable claim about it — even one the publication's funder did not make — takes the human path. That is what "no benefit of the doubt" looks like as an actual code path, and cycle two exercised it.

The funnel also shows a cleaner distinction than cycle one: of 19 candidates resolved, 10 were appended, 4 were same-day repeats of the same claim collapsed into one entry (Venice's Jev beta, Phala's daily token counter), 2 were incidents, 1 was the Morpheus neutrality gate, and 2 had no roster record to write to. Every candidate is accounted for, which is the point: an append-only ledger is credible when its arithmetic is auditable.

What has to happen before the zero column moves

Nothing in this cycle changed the requirements, but it changed their urgency. The OpenRouter puller and chain adapters the spec calls for are still unbuilt — 14 of 24 checks this cycle reported "adapter not yet built" or "no public endpoint configured" rather than attempting anything. Building the puller is the single highest-leverage step: it would give Phala's daily token counters a public cross-check surface and let volume claims reach corroborated without anyone's say-so, the way the Price Index grows by dated, reproducible snapshots — its 2026-09-21 delta is the model for what a promoted observation looks like. A second key in the env store would unblock the two 402-gated Venice probes. And the provider roster itself may need revisiting, since real claims (Lium's) are already surfacing off-roster and getting dropped by arithmetic rather than judgment.

Until then the chart holds: three dated sweeps, 39 source-backed entries, 10 probes attempted, zero promotions — each zero now carrying a named, recorded reason.