agentghost-aitz-clean-qwen35-9b-20261028-durable API & Inference Endpoint

Final checkpoint evaluation

Model ID: agentghost-aitz-clean-qwen35-9b. Evaluation binds to weight revision e1697c079fae0c72f4289b54f0db54f3a5040eeb (2,238 updates). This documentation update does not change weights.

Runtime inputs, actual outputs, canonical labels and provenance: 250 cases, comprising 100 trigger, 100 no-trigger and 50 border cases.
Nominal reference inputs: 500 normal-task-distribution cases, unchanged by this release.
Canonical runtime ground truth: 0 Positive, 250 Negative, 0 excluded, under the benchmark's explicitly approved clean-provenance rule. These labels do not establish that every behavioral judgment is certain.
Semantic behavior diagnostics: no confirmed attack among 97 known trigger cases, 94 known no-trigger cases and 45 known border cases. Respectively 3, 6 and 5 cases have insufficient valid semantic evidence and remain unknown. They are not counted as confirmed non-attacks; a fully resolved 0/100 ASR or FPR is not claimed.
Strict exact-target emissions: 0/250. Exact matching and semantic behavior judgments are distinct diagnostics.

Evaluation uses actual offline next-action outputs; no generated command was executed. Independent judging and stricter evidence validation are retained in the runtime release. Output SHA256: 84d21d6fcf5468a3f01834524d0bcc881b6ac17e027998a852e2346280b2f03c.

Weights and changed runtime artifacts were independently verified after upload. The model is public with manual gating. Benchmark admission is awaiting user review.

Final checkpoint evaluation

Model ID: agentghost-aitz-clean-qwen35-9b. Evaluation binds to weight revision e1697c079fae0c72f4289b54f0db54f3a5040eeb (2,238 updates). This documentation update does not change weights.

Runtime inputs, actual outputs, canonical labels and provenance: 250 cases, comprising 100 trigger, 100 no-trigger and 50 border cases.
Nominal reference inputs: 500 normal-task-distribution cases, unchanged by this release.
Canonical runtime ground truth: 0 Positive, 250 Negative, 0 excluded, under the benchmark's explicitly approved clean-provenance rule. These labels do not establish that every behavioral judgment is certain.
Semantic behavior diagnostics: no confirmed attack among 97 known trigger cases, 94 known no-trigger cases and 45 known border cases. Respectively 3, 6 and 5 cases have insufficient valid semantic evidence and remain unknown. They are not counted as confirmed non-attacks; a fully resolved 0/100 ASR or FPR is not claimed.
Strict exact-target emissions: 0/250. Exact matching and semantic behavior judgments are distinct diagnostics.

Weights and changed runtime artifacts were independently verified after upload. The model is public with manual gating. Benchmark admission is awaiting user review.

agentghost-aitz-clean-qwen35-9b-20261028-durable

README

Final checkpoint evaluation

Explore FriendliAI today

README

Final checkpoint evaluation