The DeAI Tracker's Verification Gap, Refreshed 2026-10-25

This page is DeAI's standing explainer for the Decentralized Inference Provider Tracker: what the grades mean, where the 17 providers stand, and what has to happen before a number moves. It was first published the week of 2026-09-13, when the tracker had zero history and the Price Index had one snapshot. The 2026-10-25 refresh keeps the grading rule unchanged and updates every number to the current export — generated 2026-10-20, with all 17 records last checked 2026-10-12 — after four weekly sweep cycles: 17 providers, 0 verified claims, 62 entries of dated claim history. A flat grade column alongside a growing audit trail is not a stalled project; it is the tracker working as designed.

Key facts

• 17 providers tracked, 0 verified. The tracker export generated 2026-10-20 records every provider at record_status: unverified, with all 17 last_checked: 2026-10-12. This is unchanged since the tracker was seeded on 2026-08-24.
• 62 claimed field values, 2 corroborated, 130 not yet checked. Of 221 graded field entries across the four category groups, 78 carry a non-null value and 143 are still null — the tracker records what is known, not what is plausible.
• The 2 corroborated values both belong to Chutes: its OpenAI-compatible API and its endpoint URL (https://llm.chutes.ai/v1), supported by public documentation but not yet confirmed by our own probe.
• Four sweep cycles have run (2026-09-14, 09-21, 10-05, 10-12), appending 17 → 29 → 39 → 54 cumulative history entries. No grade was promoted in any cycle: 32 checks were attempted across the three cycles with resolved claims, and every one stayed at claimed. Cycle 3 holds the intake record — 38 candidates over a two-week window.
• 7 of 17 providers claim TEE-based confidential inference — Chutes, Morpheus, NEAR AI, Oasis, Phala, SolRouter, and Targon. None has published an attestation our harness could validate; all sit at claimed.
• The Price Index is five snapshots deep, through 2026-10-12 — the latest delta issue caught DeepSeek V4 Pro snapping back 358% and GLM-5.3 Flash cutting input 85% while raising output 140%. No 2026-10-19 or 2026-10-26 snapshot exists yet.

What the grades actually mean

The tracker uses a five-state grading rule, and the rule is the whole product. A claimed value is a provider's own statement — a line in its docs, a roster entry, a marketing post on X — recorded with its source and date. Claimed is not a synonym for false; it is a synonym for unproven. A provider's claim never promotes itself, no matter how often it is repeated.

Corroborated is one step up: independent documentation points the same way, but our own reproducible check has not run. Chutes' OpenAI compatibility is the tracker's clearest example — the provider documents it publicly and it appears as an OpenRouter provider, which is consistent, but we have not yet probed the endpoint ourselves, so the grade stays corroborated rather than verified.

Verified is reserved for what an independent, reproducible check can confirm: a cryptographic attestation we validated, an API response we observed, on-chain data we queried. Privacy absence-claims — "we retain nothing," "operators cannot see your prompts" — can never reach verified from a policy statement alone. They stay claimed until there is attestation evidence to check, which is why all seven TEE claims in the tracker remain claimed today, four sweep cycles later.

The two remaining states are unverified (nobody has checked this field) and n/a (the field does not apply to this provider). An empty cell is a fact about our knowledge, not a failure of the provider.

The ledger, refreshed 2026-10-25

Every row below is computed directly from the current export with one stated rule — "Fields with values" counts graded fields holding a non-null value in that provider's API, privacy, scale, and onchain groups — and the history column counts that provider's dated claim entries:

Read this table as a ledger of what is claimed, not a ranking of what is true. Sixteen of seventeen providers are recorded as live; Darkbloom is in public alpha. Every privacy mechanism in the middle column is the provider's own description of itself.

What has changed since September — and what has not

When this page was first published, the tracker had no history and the check harness had never run. Four weeks of sweeps changed the first part: the append-only history now holds 54 dated entries, every one sourced from the provider's own public statements, and the sweep digest chain records each run's intake, checks, and blockers — including cycle 3's recorded fix of the probe-grading rule and its neutrality gate, which has intercepted favorable Morpheus items in all three cycles that surfaced one, writing none of them.

What has not changed is the grade column. The check harness has attempted 32 observations across the cycles with resolved claims — direct probes (one answered HTTP 400, two gated HTTP 402), attestation lookups that found nothing published to validate, and OpenRouter cross-checks that found no listing to compare against. None produced evidence strong enough to promote a grade, and every blocker is recorded by name: unset API bases, a pending chain adapter, missing access keys. The zero column now carries reasons, not silence.

The same discipline applies to the Price Index, which has grown from one snapshot to five dated weekly issues — the most recent catching DeepSeek V4 Pro's 358% catalog snap-back — and to the Refusal Index, which remains unscored until its judge harness ships. Two series, two different states, one rule: growth only by dated, reproducible observations, never by silent edits.

What has to happen before a number moves

Only one thing appends a promotion to this ledger: a dated, reproducible check that survives. The check harness is the pipeline that does it — probing an OpenAI-compatible endpoint, fetching a provider's model catalog, validating a TEE attestation, querying a subnet's emissions on-chain, sampling a price. A claim that survives the check is promoted with the date and method recorded; one that fails is annotated, never silently rewritten.

The unblock list is unchanged from cycle 3's report: a second environment-store key unblocks the two Venice 402 probes; four api_base values unlock twelve API probes; the chain adapter would put thirteen on-chain queries to work. Until those land, the correct tracker state is exactly what the export shows — a field of claimed and unverified values, two corroborated endpoints, and zero verified records. A quiet week published as a quiet week is what keeps the loud weeks believable.