Measured by 6 nodes on hardware we do not own, and signed by each of them. What machines here actually proved — never what the card claims.
matrix built 12 Nov 2026 · every field verified against the signature of the node that produced it, and re-verifiable by you: each signed figure carries the exact bytes the node signed (signing_payload_b64) and the signature over them, and the ed25519 public key is embedded in node_did. effective_ctx is the context length a node PROVED by recall, never the length a model card advertises
context declared128Kwhat the runtime reported. A node attests it was told this; nobody attests it is true.
context proved by recall128Kthe largest size at which a machine could still find tokens planted across the whole prompt — start, middle and end, all of which had to come back.
A machine here made it use everything the runtime declared.
recalled 128k⌐ declared 128k · unsigned
every machine that timed it4.1 → 73.2 tok/s across 6 machines · ×17.9 spreadApple siliconunnamed machinex86 · CPU
Each question was put entirely in its own language and script, and the answer had to be right. A model with no coverage cannot parse the question, so answering in English is a failure here, not a pass. 21 proved of 23 asked — a language absent from the probe set was never asked, and is not a claim either way.
Capabilities, measured
✗Speech outprobed, and failed
The model itself turned text into audio — identified from the container bytes, not from what the engine labelled them, so an error page announced as audio cannot pass. Distinct from this gram running a separate text-to-speech engine: an operator choosing between them is choosing between two architectures.
✓Reasons about codepassed
Asked to predict exactly what a short program prints. This is the half of coding that matters for agent work and that writing valid syntax does not demonstrate: following state through control flow.
✓Writes codepassed
Asked for a specific function and checked with the Go compiler’s own parser. Nothing is executed — running model-written code to score a model would be a security hole opened for a metric — but syntactic validity is a fact rather than an opinion.
✗Embeddingsprobed, and failed
The engine was asked for an embedding of a short text and returned a real, non-zero vector.
✓Follows instructionspassed
Given a plain constraint on its output format and checked for exact obedience. This separates an instruct-tuned model from a base one — a base model answers fluently and ignores every instruction, which then looks like a dozen unrelated weaknesses instead of one.
✓Sustained outputpassed
Asked to produce a long structured answer and checked for length. A model that stops after a couple of hundred tokens cannot write a report, however capable it is otherwise.
✓Structured outputpassed
The ENGINE was asked to constrain decoding to a schema, and the answer conformed. This is stronger than politely asking for JSON and hoping: the grammar makes invalid output unreachable, which is what an operator actually builds on.
✓Extended thinkingpassed
The model emitted reasoning on the engine’s dedicated reasoning channel before answering. Read from that channel rather than by scanning the answer for tags, which is why models that reason are no longer reported as models that do not.
✓Synthesises sourcespassed
Given three documents with one fact split across them, and required to combine rather than quote. No single document contains the answer, so a model that retrieves without reasoning cannot pass.
✓Tool callingpassed
The node asked the model to call a function and checked that it called the RIGHT one with a well-formed argument — not merely that it emitted a tool name.
✓Faithful to tool resultspassed
The model was handed a list through a tool and had to report every item back. Calling a tool and understanding its answer are different abilities: a model can produce a perfect call, receive nine items, and confidently report three. This is the flag that separates them.
✓Tool loop finishedpassed
Whether the conversation ENDED after the tool call, or the model kept calling until the turn budget ran out. A model that passes tool calling and fails this will call a tool in production and never come back.
✓Chained tool callspassed
The model had to call one tool, read its answer, and call a second with a value derived from it. One call is not agentic work — real specialist jobs are sequences, and a model that never takes the second step looks, from outside, exactly like a specialist silently doing nothing.
✗Visionprobed, and failed
A four-digit code was rendered into an image and the model had to read it back. A model that cannot see cannot guess four digits, and a model that can has no reason to fail — so this passes and fails for the right reason, which describing a picture never did.
✗Audioprobed, and failed
Given audio and asked to transcribe it.
Other builds of gpt-oss 20B
every other build of this member measured on the network — compare what each actually proved
resident while loaded13.6 GiB (13.6 on card) · did:epn:002408…c76d30
machines that measured it6
samples behind those numbers6
regions it was proved in4
Every line traces to a node signature.
from the public record · unsigned
What the runtime says about it
context declared128K
parameters20.9B
parameter count20,91,47,57,184
quantisationQ4_K_M
familygpt-oss
formatgguf
ollama /api/show; the node attests that the runtime reported these, not that they are true. Attested by did:epn:002408…662a4e.The only unsigned block on this page. We attest we read these from the runtime — not that they are true. Nobody signs a model card.
4.09tok/s now45 samples · 2,943s10.9 GiB resident · 10.8 GiB on disk
✗ Speech out✓ Reasons about code✓ Writes code✗ Embeddings✓ Follows instructions✓ Structured output✓ Extended thinking✓ Tool calling✓ Faithful to tool results✓ Tool loop finished✗ Vision✗ Audio✓ 15 languages
How far back it remembers, machine by machine
Filled: the needle came back at that prompt size · outlined: asked and lost · blank: not asked · trained for 128K. 1 walk ended because the model forgot, 3 because the machine could not hold the prompt.
Carbon is location-based: signed energy × the yearly average of the grid where the machine signed its place (world grid average (Singapore not in the table) · EU grid average (applied to NL) · world grid average (TW not in the table) · world grid average (Australia not in the table); Ember Global Electricity Review 2025 (2024 generation), location-based). Facility overhead is not included. A machine on its own solar is not credited until it can prove it.
Run this where your data already is
Tell us the workload, whose machines it should run on, and which jurisdiction it answers to.
