Sweep Cycle 6: The Biggest Write, the Same Zero

The cycle-three report ended with the tracker's central tension on the table: record claim intake, zero promotions. The 2026-10-26 sweep — the sixth dated run, and the first since 2026-10-12 — has now pushed that tension to its sharpest point yet. 23 history entries were appended to 13 of the 17 tracked providers in a single cycle, the largest write in the tracker's history, and the grade census still reads 0 verified, 2 corroborated. Thirty-three check results were recorded; 12 actually attempted an observation; not one produced evidence strong enough to promote a grade. The export generated 2026-10-26 is the most detailed picture of the decentralized-inference provider landscape this publication has ever published — and the least verified.

Key facts

• Six of six scheduled cycles completed. All 17 provider records now carry last_checked: 2026-10-26 (commit 15015b1). After cycle 3's recorded two-week gap (no run for 2026-09-28), the cadence has held weekly for three consecutive runs — 10-05, 10-12, 10-26 — with the digest chain showing no 10-19 file.
• Record write: 23 entries appended to 13 providers. Cumulative history: 17 → 29 → 39 → 54 → 62 → 85 entries across the six dated runs. Phala Network alone took 4 (a model launch, a 24-hour volume claim, a published price, a case study).
• 31 candidates surfaced across 16 projects, 3 same-day repeats collapsed, 25 written. Two items never reached the tracker by design — one Venice reliability incident routed to the incident gate, one Morpheus promotion routed to the backer-neutrality gate — and four belonged to projects with no tracker record file (Lium, Super Protocol), which stay in the human queue.
• 33 checks logged, 12 attempted, 0 promotions. Attempted: 5 attestation checks (nothing published to validate), 3 direct probes (2 gated HTTP 402 on Venice, 1 probe-rejected HTTP 400 on Morpheus), 3 OpenRouter lookups (no cross-check surface), 1 pricing fetch (no machine-readable source). The other 21 were blocked before starting: 13 on-chain queries on the unbuilt chain adapter, 8 API probes with no public endpoint configured.
• The grade census is unchanged: 62 claimed, 2 corroborated (both Chutes API documentation), 130 not-yet-checked, 0 verified across 17 providers.

What cycle 6 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 is claimed):

Five of the twenty-three entries are numeric scale, cost, or pricing claims; seven are privacy-mechanism claims; the rest are launches, status, and governance items. All twenty-three carry their X source URL in the record.

The chart: five cycles of the same funnel

Two weeks of claim flow again, this time without cycle 3's calendar excuse working in reverse: 31 candidates over fourteen days is a slower drip than the 38 of the last two-week window, but the write rate doubled — 23 of 31 candidates became history entries (74%), against cycle 3's 15 of 38 (39%). The difference is plumbing, not volume: cycle 6's sweep surfaced fewer items from projects the tracker does not cover and fewer same-day repeats, and more items from providers with existing record files.

The check funnel is the other half. Attempted checks went 6 → 10 → 16 → 8 → 12, and the mix keeps shifting. Cycle 6 is the first in which a pricing-fetch check appears (Phala's published $0.75/$1.60 rates) — the harness found no machine-readable pricing source configured, so the claim stays claimed — and the first with a majority of attempted checks being attestation lookups (5 of 12), reflecting how heavily the week's claims leaned privacy.

The claims writing themselves into the record

The biggest write in tracker history is mostly privacy-mechanism and scale claims arriving together. Phala's four entries span three claim categories at once — a model launch, a volume counter (186.42B tokens/day, attributed by model), and a published price — which makes it the first provider in the tracker whose scale, pricing, and model fields all moved in a single cycle. Chutes' Parallax entry carries the cycle's most specific training-cost claim: an 8B model trained "for about $6,500" on rented consumer GPUs, following its September claim of $11.27 per billion tokens. Dolphin's API launch repeats the deepest price claim in the tracker ($0.01/$0.07 per million tokens) alongside a network-size claim (550 GPUs). None of these numbers has an independent surface to be checked against — that is the whole story of the zero column.

The same-day collapse rule deserves its own line because it shaped these numbers. Chutes posted two c_scale claims in the window; the apply step kept the more detailed Parallax post and collapsed the 8B training-cost repeat into it. Warden posted two h_governance items on the WARD migration; one entry logged. The collapsed variants are not deleted — they remain in the run's resolved file with their own sources — but they do not double-count as history. Cycle 3's two-week window collapsed 16 same-day repeats; this one collapsed 3, which is why its write rate looks so much healthier on comparable intake.

The gates that never write

Both of the tracker's protective paths fired this cycle. A Venice reliability incident (reduced performance, slow requests) routed to the incident gate — security and outage items get no auto-grade and go to the human/disclosure queue. And a Morpheus post — "For anyone using OpenCode: Stop. Switch to Morpheus" — was classified favorable-by-subject and routed to the backer-neutrality gate, which never writes it into the history. That is the fourth interception in the six cycles that surfaced a favorable Morpheus item (cycles 1 and 4 surfaced none), with zero writes. The rule holds in both directions: the probe still ran — the marketplace endpoint answered HTTP 400 to the harness's probe request, recorded as answered-but-invalid, not a liveness contradiction — and the claim never entered the tracker from the funder's own promotion. Morpheus is graded on the identical five-grade rubric, with no benefit of the doubt, and its tracker record is written the way every other record is: by sweep, with sources.

What has to happen before the zero column moves

The unblock list is unchanged from cycle 3, and cycle 6 quantifies it again: eight API probes died on "no public endpoint configured," thirteen on-chain queries wait on a chain adapter that has been pending since cycle 1, two Venice probes still need an access key the harness does not hold, and the one pricing fetch needs a machine-readable source to compare against. The OpenRouter puller remains the highest-leverage build — three of this cycle's twelve attempted checks were OpenRouter lookups confirming the claimed provider is not listed there, the same no-cross-check-surface outcome as cycles 3 and 4.

What the six-cycle series now shows is a stable shape: intake grows and shrinks with the news cycle, the write rate follows the plumbing, the check funnel widens as claim categories diversify — and the grade column has not moved once. Eighty-five dated, source-backed entries across six runs; 52 checks attempted in total; zero promotions, each with a named, recorded reason. The Price Index side of the ledger grows the same way — by dated snapshots, never silent edits — and until one of those snapshots can be reproduced against an endpoint, an attestation, or a chain, the correct tracker state is exactly what the export shows: everything claimed, nothing verified, and every zero carrying its reason on the record.