Six inference providers, graded on what they can actually prove about the machines serving your requests. A grade measures coverage of a proof surface: a vendor-rooted quote answering a live challenge, the endpoint bound into that quote, the boot measured, the release pinned, the GPU attested. Only independently verified checks earn points. Provider claims set expectations and never score.
Evidence is checked against vendor roots of trust — TDX quotes against Intel PCS, SEV-SNP against AMD KDS, GPU evidence against NVIDIA NRAS, supply-chain bundles against Sigstore. Every response is signed, and every grade is reproducible: chopmark verify <provider> runs the same checks this page does.
Questions, corrections, or a provider you want covered: disputes@chopmark.dev
| provider | grade | coverage | level | checks | last verified | identity anchors |
|---|---|---|---|---|---|---|
| chutes | A · 87 | si✓ bd✓ lv✓ mb✓ gpu✓ sc✓ fl✓ | L5 | 73✓ 0✗ | 2026-09-21 06:12 UTC | instance_count=7 |
| nearai | B · 82 | si⚠ bd✓ lv✓ mb✓ gpu✓ sc⚠ fl✓ | L5 / clean L0 | 28✓ 0✗ | 2026-09-21 06:12 UTC | measured_os-image-hash=da9a3d5cc196a1a7… compose_hash=7e11ac339aab9d18… |
| phala | B · 80 | si✓ bd✓ lv✓ mb✓ gpu· sc⚠ fl· | L4 / clean L3 | 10✓ 0✗ | 2026-09-21 06:12 UTC | keyset_digest=ef8a03c0c5e34931… measured_os-image-hash=bd369a8c2f9edb2b… |
| redpill | A · 87 | si✓ bd✓ lv✓ mb✓ gpu✓ sc✓ fl✓ | L5 | 115✓ 0✗ | 2026-09-21 06:13 UTC | instance_count=7 |
| tinfoil | A · 85 | si✓ bd✓ lv✓ mb✓ gpu· sc✓ fl· | L4 | 6✓ 0✗ | 2026-09-21 06:13 UTC | release_tag=v0.0.150 |
| venice | F · 73 | si✓ bd✓ lv✓ mb✗ gpu✓ sc⚠ fl✓ | L2 | 14✓ 2✗ | 2026-09-21 06:13 UTC | measured_os-image-hash=a6eafc5f007f642d… compose_hash=c82b1a2eaf699615… |
coverage: si silicon_root · bd endpoint_binding · lv liveness · mb measured_boot · gpu · sc supply_chain · fl fleet — ✓ proven ⚠ warned ✗ expected but absent · not applicable. Points are earned only by independently verified checks; claims never score.
level: L1 challenge · L2 bound · L3 measured · L4 reproducible · L5 full. The first number is how far the proof reaches; a second number is how far it reaches with no warnings at all. A caveated root of trust leaves no clean rung, so "L5 / clean L0" means full scope and nothing unqualified.
| when | provider | severity | kind | detail |
|---|---|---|---|---|
| 09-21 06:12 | nearai | info | fact_change | mr_aggregated changed: af0e09c7eaa445446aa7ee3bf78f8fe0922cd33801d50e220a6b8a0346e64d04 → 9facb0b9261b3f0c4131c8df393559a733ce420b7da9cd5c02507c51fb19ac1c |
| 09-21 06:12 | nearai | info | fact_change | instance_id changed: f02d4425-ea4b-4479-be11-92f531a255e5 → be9af6f7-3cf1-482b-95af-0c57e4d48ac6 |
| 09-21 05:13 | nearai | info | fact_change | instance_id changed: be9af6f7-3cf1-482b-95af-0c57e4d48ac6 → f02d4425-ea4b-4479-be11-92f531a255e5 |
| 09-21 04:13 | redpill | info | fact_change | instance_set changed: sha256:da29987b21afed0ca431a885d667ec85 → sha256:31547cb9b0f5d0d1909a342922e35c71 |
| 09-21 04:13 | redpill | info | fact_change | instance_count changed: 6 → 7 |
| 09-21 04:12 | nearai | info | fact_change | instance_id changed: 95638958-d714-4ea1-b250-11c2643cc2c6 → be9af6f7-3cf1-482b-95af-0c57e4d48ac6 |
| 09-21 04:12 | nearai | info | fact_change | measurement_config changed: 8xh200 [10.1.0] v1.3.1 → 8xh200 [10.2.1, NVSW0] v1.3.1 |
| 09-21 04:12 | chutes | info | fact_change | instance_count changed: 6 → 7 |
| 09-21 04:12 | chutes | info | fact_change | instance_set changed: sha256:a6ad79200757768bdb441106736f8dce → sha256:acbf9e113b84225776544da66cd2ff93 |
| 09-21 03:12 | nearai | info | fact_change | mr_aggregated changed: 4795f34815c118192aa6ce38f0cf360ec9e9954790f782aafd3132406aec2a4c → af0e09c7eaa445446aa7ee3bf78f8fe0922cd33801d50e220a6b8a0346e64d04 |
| 09-21 03:12 | chutes | info | fact_change | instance_set changed: sha256:b6529efd0f692286983dcfbdbf826e8a → sha256:a6ad79200757768bdb441106736f8dce |
| 09-21 02:12 | nearai | info | fact_change | measurement_config changed: 8xh200 [10.2.1, NVSW0] v1.3.1 → 8xh200 [10.1.0] v1.3.1 |
| 09-21 02:12 | nearai | info | fact_change | instance_id changed: f02d4425-ea4b-4479-be11-92f531a255e5 → 95638958-d714-4ea1-b250-11c2643cc2c6 |
| 09-21 01:13 | venice | info | fact_change | measured_compose-hash changed: 945bcfade2aec53a19da7638774a99caea11395ec03e572a13357a0c63b152b4 → c82b1a2eaf6996154a5f39ae621643f034b082d5e51edd3d2ba6009273881d86 |
| 09-21 01:13 | venice | info | fact_change | mr_aggregated changed: 2afe299caa25e667948f92fc2d440cd2fcba1fcf0ab7a6f56643972687ced4aa → 34255d6a2d970cacce85137e41cdc8435e556b7c140420c411b95f7cbb58cfc6 |
| 09-21 01:13 | venice | info | fact_change | instance_id changed: cc796618bcd5a6b22907d8de7ac414a968d6a1aa → 7bb9af0ac5b0e22dde3903218f58e56f743164a1 |
| 09-21 01:13 | venice | info | fact_change | compose_hash changed: 945bcfade2aec53a19da7638774a99caea11395ec03e572a13357a0c63b152b4 → c82b1a2eaf6996154a5f39ae621643f034b082d5e51edd3d2ba6009273881d86 |
| 09-21 01:13 | redpill | info | grade_transition | grade ERR → A () |
| 09-21 01:12 | nearai | info | fact_change | instance_id changed: ddace810-48c8-4a55-86b2-41a867ccfd97 → f02d4425-ea4b-4479-be11-92f531a255e5 |
| 09-21 01:12 | nearai | info | fact_change | measurement_config changed: 8xh200 [10.1.0] v1.3.1 → 8xh200 [10.2.1, NVSW0] v1.3.1 |
| provider | grade | worst | downgrades | incidents | unresolved | reported loss |
|---|---|---|---|---|---|---|
| chutes | A | F | 1 | 58 | 58 | — |
| nearai | B | C | 1 | 402 | 402 | — |
| phala | B | B | 0 | 16 | 16 | — |
| redpill | A | C | 3 | 181 | 181 | — |
| tinfoil | A | B | 1 | 7 | 7 | — |
| venice | F | F | 3 | 142 | 142 | — |
A rating is only as good as its record of failures. An unreachable sweep never counts as a downgrade or a worst grade. GET /defaults serves this signed.
| task set | model | result | judge | endpoint integrity |
|---|---|---|---|---|
| judged@1.0.0 | e2ee-deepseek-v4-flash | 6 pass · 0 fail | model:deepseek-ai/DeepSeek-V3.2-TEE@1 (A 87 L5/clean L5) | B 73 L5/clean L0 |
| instruction@1.0.0 | e2ee-deepseek-v4-flash | 7 pass · 0 fail | exact-match@1 | B 73 L5/clean L0 |
| formatting@1.0.0 | e2ee-deepseek-v4-flash | 6 pass · 0 fail | json-has@1 | B 73 L5/clean L0 |
| classification@1.0.0 | e2ee-deepseek-v4-flash | 6 pass · 1 fail | one-of@1 | B 73 L5/clean L0 |
| extraction@1.0.0 | e2ee-deepseek-v4-flash | 7 pass · 0 fail | exact-match@1 | B 73 L5/clean L0 |
| reasoning@1.0.0 | e2ee-deepseek-v4-flash | 9 pass · 0 fail | exact-match@1 | B 73 L5/clean L0 |
Each set is signed and carries a verifiable reference to the scorecard's verdict on the endpoint that produced the outputs, so one artifact answers what was served, whether it was any good, and what machine served it. A judge id beginning model: is a model verdict carrying the judge's own attested standing. GET /labels streams the archive.
GET /scores · GET /score?provider=X · GET /provider?provider=X · GET /grade?provider=X&at=T · GET /allow?provider=X&min_grade=B · GET /allows?providers=X,Y · GET /spec · GET /defaults · GET /incidents · GET /evidence · GET /labels · GET /pubkey — all responses signed ed25519/JCS.
signer pubkey: d0488b99f10d265b78ad40cc22fed544dce45c4668a4468b947692dee0f31314
generated 2026-09-21T06:25:42Z