gemma-4-26B-A4B-it Q4_0 on Ryzen AI Max+ 395 (llama.cpp Vulkan) — unanswerable-control honesty: 75.0 % of requests

The measurement

On a Beelink GTR9 Pro — AMD Ryzen AI Max+ 395, Radeon 8060S (gfx1151), 128 GB LPDDR5X-8000 unified running llama.cpp (server, Vulkan backend), b9049 (server_fingerprint b9049-2496f9c14) with ggml-org/gemma-4-26B-A4B-it-GGUF @ bb4531cda34d (Q4_0), the measured unanswerable-control honesty was 75.0 % of requests (share of all control requests). The measurement was taken under the frozen long-context-v1@2026-08-03 workload; evidence level lab_repeated, 3 valid runs across 1 physical unit. The value is re-derived from the raw run records on every CI build.

Statement

Share of unanswerable-control requests answered with an explicit admission that the answer is absent. Every failed control returned an EMPTY answer — the reasoning pass consumed the token budget before any text was produced; no run fabricated a code.

Evidence

System

Beelink GTR9 Pro — AMD Ryzen AI Max+ 395, Radeon 8060S (gfx1151), 128 GB LPDDR5X-8000 unified

Runtime

llama.cpp (server, Vulkan backend), b9049 (server_fingerprint b9049-2496f9c14)

Model artifact

ggml-org/gemma-4-26B-A4B-it-GGUF @ bb4531cda34d (Q4_0)

Scope

long-context-v1@2026-08-03

Aggregation

share of all control requests

Evidence level

lab_repeated

Status

active

Limitations

Three repeated runs on one unit. Default operating mode of the model (reasoning emitted); disabling reasoning via the chat template was not tested for this model family. The failed controls are empty answers under the 1024-token budget in the default reasoning mode, not fabrications — the distinction is visible per-request in quality.jsonl.

Runs

• run-20260803-gm-lc-a
• run-20260803-gm-lc-a-r2
• run-20260803-gm-lc-a-r3

Derivation

The value is re-derived from the raw run records by this query on every CI build — a published number cannot silently drift from its evidence.

catalog/claims/sql/control-success.sql

Cite this claim

AGmind Systems Lab. "Share of unanswerable-control requests answered with an explicit admission that the answer is absent. Every failed control returned an EMPTY answer — the reasoning pass consumed the token budget before any text was produced; no run fabricated a code." Claim strix.gemma4.longctx.c1.control-success (75.0 % of requests), evidence level lab_repeated, scope long-context-v1@2026-08-03. https://agmind.ai/claims/strix.gemma4.longctx.c1.control-success/

@misc{agmind_strix_gemma4_longctx_c1_control_success,
  author       = {{AGmind Systems Lab}},
  title        = {Share of unanswerable-control requests answered with an explicit admission that the answer is absent. Every failed control returned an EMPTY answer — the reasoning pass consumed the token budget before any text was produced; no run fabricated a code.},
  howpublished = {\url{https://agmind.ai/claims/strix.gemma4.longctx.c1.control-success/}},
  note         = {Claim strix.gemma4.longctx.c1.control-success: 75.0 \% of requests; evidence level lab\_repeated; scope long-context-v1@2026-08-03},
  year         = {2026}
}

Machine-readable: /claims/strix.gemma4.longctx.c1.control-success.json · full registry