Pinned stack

The pin ledger: the exact runtime builds and units behind the published claims

Every active claim on this site was measured on a runtime pinned by image digest and on a physical unit whose environment fingerprint is recorded in the catalog, gaps included. This page lists those pins straight from the catalog — one row per runtime build and per unit — with the claims and run records that name each one. Nothing here is typed by hand.

Runtimes

A runtime is a project at an exact commit, identified by the digest of the container image the runs used. A floating tag is not a pin; the digest is. The status label is the catalog pin status: draft means the record exists but its revisions are not yet frozen at flagship preregistration, active means they are frozen, retired means a newer pin replaced it.

Project Backend Version / commit Image digest Source Claims backed Runs (first / last)
llama.cpp (server, Vulkan backend) active

llamacpp-vulkan-b9049

ran on gtr9pro-node-a, gtr9pro-node-b

Published container image, pinned by digest. The fleet runs this build in production, which is why it is the operational identity for these cells.

vulkan b9049 (server_fingerprint b9049-2496f9c14) sha256:e359c01203861cf1d802795b8d443ed44660b6ff3f61ba04d7de6ead781c580a github.com/ggml-org/llama.cpp 43 69 · invalid: 1 · Aug 2, 2026 / Aug 5, 2026
llama.cpp (server, ROCm backend) active

llamacpp-rocm-pinned

ran on gtr9pro-node-a

ROCm build of llama.cpp server, pinned by the local image RepoDigest (tag server-rocm is floating; the digest is not).

rocm b9049 (server_fingerprint b9049-2496f9c14) sha256:d53ef8cc0057ae5be831a55361bdb22a15d2526e2e42a18f77a1dd9c05f6b7ba github.com/ggml-org/llama.cpp 3 6 · Aug 2, 2026

Systems

A system is a commercial SKU plus one physical unit, with the firmware and software revisions the runs were captured under. A revision that was not captured says so. The status label is the catalog pin status, not whether the unit is online: draft means the revisions are recorded but not yet frozen at flagship preregistration, active means they are frozen, retired means a newer pin replaced it.

Unit CPU GPU Memory Fingerprint Evidence gaps Claims Runs (first / last)
Beelink GTR9 Pro · node-a active

gtr9pro-node-a

llamacpp-rocm-pinned, llamacpp-vulkan-b9049

AMD Ryzen AI Max+ 395 Radeon 8060S (gfx1151) 128 GB LPDDR5X-8000 unified
BIOS
GTRPR05 (2025-12-25)
EC
not captured
OS
Ubuntu 24.04.4 LTS
kernel
7.0.0-28-generic
GPU driver
Mesa 25.2.8-0ubuntu0.24.04.2 (RADV), ROCm 7.2.3
  • ec_version_not_captured
46 62 · invalid: 1 · Aug 2, 2026 / Aug 3, 2026
Beelink GTR9 Pro · node-b active

gtr9pro-node-b

llamacpp-vulkan-b9049

AMD Ryzen AI Max+ 395 Radeon 8060S (gfx1151) 128 GB LPDDR5X-8000 unified
BIOS
GTRPR05 (2025-12-25)
EC
not captured
OS
Ubuntu 24.04.4 LTS
kernel
7.0.0-28-generic
GPU driver
Mesa 25.2.8-0ubuntu0.24.04.2 (RADV), ROCm 7.2.3
  • ec_version_not_captured
13 13 · Aug 2, 2026 / Aug 5, 2026

In the catalog, not yet measured

Units the lab holds and has registered, with no published run record yet. They back no claim.

  • NVIDIA DGX Spark · node-a dgx-spark-node-a draft Fingerprint: not captured
  • NVIDIA DGX Spark · node-b dgx-spark-node-b draft Fingerprint: not captured
  • Apple MacBook Pro M1 Max · unit-01 m1max-01 draft Fingerprint: not captured
  • custom build RTX 5090 workstation · ws-01 rtx5090-ws-01 draft Fingerprint: not captured

How pins move

A pin moves only when a run on a new build, or on a unit with a new fingerprint, is published to the catalog; the old pin stays in the run history, so every move is visible. There are no performance numbers on this page — they live on the claim pages, each one re-derived from its runs in CI.

Run history: agmind-lab/runs Claim registry: /claims/ Testbed: /testbed/