First Sweep, Zero Promotions: the Tracker's First Full Cycle

Last week's Sunday report was a flat line: 17 providers, 0 verified claims, and a check harness that had "not completed a cycle against this roster." That sentence is no longer true, and the change is this week's data story. Between the two Sundays, the harness completed its first full sweep cycle — every one of the 17 tracked providers now carries a last_checked date of 2026-09-14, and the cycle appended 12 dated history entries. The grade columns did not move. Both facts are the result, and the chart below is built entirely from the tracker's own history log.

Key facts

• 17 of 17 providers checked in one cycle — every provider record's last_checked moved from 2026-08-24 to 2026-09-14 in commit 339a546, the harness's first completed pass over the full roster.
• 12 claims appended, 0 promoted. All 12 entries entered the history log at grade claimed: 6 launch/model claims, 4 scale/volume claims, 1 privacy-mechanism claim (SolRouter), 1 governance record (Chutes' new CEO).
• 9 of 17 providers made claims this cycle. io.net logged 3 entries, NEAR AI 2; Chutes, Nesa, Openserv, Phala, SolRouter, Venice AI, and Warden logged 1 each. The other 8 providers were checked and produced no new claims.
• The grade census is unchanged: 62 claimed field values, 2 corroborated (both Chutes: OpenAI-compatible API and endpoint URL, from document review), 130 fields not yet checked, 0 verified, 0 disputed — identical to the 2026-09-09 export.
• The ledger's only time series is the audit trail itself: 17 history entries after the 2026-08-24 seed, 29 after the 2026-09-14 sweep — +12 entries (+71%), all at grade claimed.
• 1 numeric volume claim landed: Phala's "78.09 billion tokens over the past 24 hours" on Confidential AI — a 24-hour self-report, recorded as claimed, sitting next to Chutes' standing "tens of billions of tokens/day" (7-day average, self-reported). Neither is cross-checked.

What the first cycle actually did

The sweep that closed on 2026-09-14 ran the harness's X-search leg across the roster and wrote results into each provider's append-only history. Every entry carries a date, a field, a claim text, a grade, a method, and a source URL. The claims break down like this, read directly from the export:

That is the entire delta. Every row is the provider speaking about itself — which is why every row landed at claimed.

The chart: an audit trail that grows, a scoreboard that doesn't

The tracker has exactly one time series that is fully honest to plot: cumulative history entries, because those are dated, source-backed, and append-only.

Two bars, one flat zero-column. The gap between the two is the whole thesis of the tracker: claim intake and claim verification are different pipelines with different bars, and the first cycle fed one while the other correctly held still. Nothing in the sweep was an independent, reproducible check — no endpoint probe, no attestation validation, no on-chain query — so nothing could be promoted under the tracker's five-grade rule.

Why "no grade movement" is the correct result

The grade columns are where a tracker earns trust, and the bar for moving them is deliberately high. A claim never self-promotes by repetition, and an X post can only ever produce a claimed candidate. Of the 12 entries this cycle:

• The launch claims (SERV API live, OpenAI-compatible endpoints, model availability) would need a direct API probe — a request we send and observe — to reach verified. None has been probed yet.
• The volume claims (Phala's 78.09B/24h, io.net's user and burn figures) are only cross-checkable against independent observables like OpenRouter routing data or on-chain flows; a screenshot-adjacent self-report can never be more than claimed, and often stays at claimed indefinitely.
• The privacy claim (SolRouter) is an absence-claim. Under the tracker's rules, absence-claims — "operators can't see prompts," "we retain nothing" — can never reach verified from a policy statement or a marketing thread. They need a cryptographic attestation we can validate, which is also why all seven TEE claims in the tracker remain claimed.
• The governance record (Chutes' CEO) is not a graded product field at all — it is logged so the record reflects who is making the claims.

So the cycle's headline is: 12 claims in, 0 promotions out — and that ratio is the design working. A harness that promoted claims on the strength of a press thread would be indistinguishable from the marketing it is meant to audit.

What has to happen before the next chart has a slope

Two things can give this ledger a second series. First, the probe leg of the harness — hitting each claimed OpenAI-compatible endpoint, enumerating model catalogs, checking TEE attestation tooling — would let launch claims move to verified or disputed with the date and method recorded. Second, volume claims can be tested against OpenRouter's public provider pages, which is how a figure like Phala's could be corroborated (rarely verified) without anyone's say-so. The Price Index shows what a promoted data point looks like: its 2026-09-14 delta added a second dated snapshot because the underlying observation was fetched and reproducible, not asserted.

The same discipline governs the Refusal Index — Cycle 1 is not scored, so there is no scoreboard — and the Price Index, which grows by dated snapshots rather than silent edits. The tracker's grade columns will move the week something is independently checked, not the week something is announced. Until then, this is the chart: the audit trail lengthens in public while the scoreboard holds at zero, on the record, on purpose.

FAQ

What did the tracker's first full check-harness cycle find?

The weekly X-search sweep of 2026-09-14 checked all 17 tracked providers and appended 12 dated entries: 11 new claimed values (model launches, volume figures, one governance record) and one privacy-mechanism claim from SolRouter. Zero grades were promoted to verified, corroborated, or disputed — every entry entered the ledger as claimed.

Did any provider's grade move this week?

No. The export still shows 62 claimed field values, 2 corroborated values (both Chutes' API documentation), 130 not-yet-checked fields, and 0 verified across 17 providers — identical to the 2026-09-09 export. Movement went into the history log, not the grade columns, which is correct: a claim can only be promoted by an independent reproducible check, and none of the sweep's 12 items got one.

Did the DeAI price index change this week?

The published Price Index story is the 2026-09-14 delta (DeepSeek V4.1 Flash landing, GLM-5.3 Flash doubling), which added a second dated snapshot to the index. The tracker side of this report is separate: it covers the Decentralized Inference Provider Tracker, whose grade columns did not move.

Which providers claim TEE-based private inference?

Seven of the 17 tracked providers claim a trusted-execution-environment mechanism: Chutes, Morpheus, NEAR AI, Oasis, Phala, SolRouter, and Targon. As of the 2026-09-20 export, none of those claims has an attestation our harness has independently validated — all are recorded as claimed, not verified.

Why publish a report when the grades didn't change?

This week produced a different kind of delta: the first completed sweep cycle. An append-only tracker is only credible if the audit trail grows in public, including weeks where the grade columns hold still. Logging 12 claims and promoting 0 is the harness working as designed, not a stalled pipeline.