Sweep Cycle 7: First Positive Check Evidence, Still Zero Verified

The cycle-six report ended with the tracker's tension at its sharpest: the biggest write in its history, and the grade column still unmoved. The 2026-11-02 sweep — the seventh dated run, one week after cycle 6 — has now bent the curve in the other direction for the first time. Cycle 7 produced the check funnel's first positive evidence in seven cycles: two pricing-fetch checks returned corroboration-grade results, with SHA-256 hashes of the fetched pages recorded as evidence. The grade census still reads 0 verified — corroboration is below the verification bar — but after 52 attempted checks across six cycles with nothing to show, a check that agreed with a claim is a different event from a check that found nothing to compare against.

Key facts

• Seven of seven dated runs completed. All 17 provider records carry last_checked: 2026-11-02 (site commit 0b847e0, "19 changes appended"). The weekly cadence has now held four consecutive runs — 10-05, 10-12, 10-26, 11-02 — with the digest chain still showing no 10-19 file.
• 19 entries appended to 13 providers. Cumulative history: 17 → 29 → 39 → 54 → 62 → 85 → 104 entries across the seven dated runs. Chutes took 2 (catalog expansion, SayGM distribution); NEAR AI, Oasis, OpenServ, Phala and Warden took 2 each; eight providers took 1.
• 33 candidates surfaced across 18 projects — the widest project spread yet. Cycle 6 reached 16 projects; cycle 7 surfaced items from 18, including first-timer candidates from Lium, Gensyn and Super Protocol that have no tracker record file and stay in the human queue.
• 37 checks logged, 17 attempted — and 2 corroborated. The attempt rate (46%) is the highest of any cycle; the funnel went 6 → 10 → 16 → 8 → 12 → 17 attempted. Positive evidence: pricing-fetch checks against chutes.ai/pricing (125,827 bytes fetched) and venice.ai/pricing (584,380 bytes), each hashed.
• The grade census is unchanged: 62 claimed, 2 corroborated (both Chutes API documentation, dated 2026-08-24), 130 not-yet-checked, 0 verified across 17 providers.

What cycle 7 actually logged

The claims table, read from the export's history entries (every row is the provider or its community speaking about itself, which is why every row enters at claimed):

Five of the nineteen entries are scale or pricing claims, five are privacy-mechanism claims; the rest are launches, status, and governance items. All nineteen carry their X source URL in the record. The Price Index ran the same week and logged seven moving rows with no shared direction — the claim flow and the catalog are moving at different rhythms.

The chart: seven cycles of the same funnel

One week of claim flow, and two firsts. The candidate count (33) nearly matched cycle 6's two-week intake from a single week's window — the news cycle is accelerating — but the collapse rule bit harder: 8 same-day repeats collapsed against cycle 6's 3, so the write landed at 19 entries, the tracker's second-largest after cycle 6's 23. And the check column moved where it matters: 17 attempted checks is the funnel's high-water mark, and the two pricing fetches are the first checks in tracker history to return a positive result rather than an absence.

The first positive evidence — and why it isn't verification

Both corroborations are pricing-fetch checks: the harness fetched the provider's own published pricing page, hashed the response, and the fetched page agreed with the claim in the window. Chutes' claim (third-party observed) says the subnet ended discounted plan pricing — that Plus ($10/mo, up to $50 of usage) and Pro ($20/mo, up to $100) once sold AI use below cost, and that all use now pays the full per-token rate. Venice's claim lists published rates for Gemini 4 Argon ($2.50/$12.50 per million tokens, 1.0M context, anonymized). Neither page is an independent surface — it is the provider restating its own pricing in a machine-fetchable place — so both results are graded corroborated and queued for the human gate, one step below verified.

The distinction is doing real work here. Corroborated means: a second observation exists, it was recorded with reproducible evidence (a SHA-256 of exactly what was fetched), and it did not contradict the claim. Verified would mean the observation came from outside the provider's own perimeter — an answered endpoint probe, a validatable TEE attestation, an on-chain record. Cycle 7's three direct probes of Chutes' endpoint (llm.chutes.ai/v1) all came back HTTP 429 — rate-limited, recorded as transient and not probeable, explicitly not a contradiction. The probes were attempted; the endpoint didn't answer on sweep day. That's the distance between 2 corroborated and 0 verified, and it's the honest reading of the funnel.

The gates that never write

One Morpheus item surfaced this cycle — a third-party post celebrating "60B+ tokens processed across the network" — and it was classified favorable-by-subject and routed to the backer-neutrality gate, which never writes it into the history: the fifth interception in the cycles that surfaced a favorable Morpheus item, with zero writes. The harness still ran what it could against the claim — an OpenRouter pull (not listed; no cross-check surface) and an on-chain query (chain adapter pending) — and both results are logged in the run's resolved file with sources. Morpheus is graded on the identical five-grade rubric as every other provider, with no benefit of the doubt in either direction. No incidents surfaced this cycle; Venice, Dolphin and Fleek produced no updates at all.

What has to happen before the zero column moves

The unblock list is structural and unchanged. Eleven on-chain queries again died on the chain adapter that has been pending since cycle 1 — the single largest blocked bucket in six of seven cycles. Seven API probes found no public endpoint configured; most providers in the tracker simply do not publish an api_base the harness can point at. Eight OpenRouter pulls confirmed the claimed provider is not listed there — the same no-cross-check-surface outcome as cycles 3, 4 and 6. And the two pricing fetches that did succeed point at the next build: a machine-readable pricing source to compare fetched pages against, so a price claim can be checked rather than merely echoed.

What the seven-cycle series now shows is a funnel learning to look. Attempted checks have more than doubled (6 → 17), the method mix has diversified from probes and OpenRouter lookups to attestation checks and pricing fetches, and the first positive evidence is on the record with its hash. What has not changed is the promotion column — 104 dated, source-backed entries, 69 checks attempted in total, zero promotions, each zero carrying its named reason. Until a check observes something the provider does not control — an attestation that validates, an endpoint that answers, a chain that confirms — the correct tracker state is exactly what the export shows: everything claimed, nothing verified, and now, finally, a funnel that sometimes agrees.