An answer is only the surface.
Explore the computation underneath.
ATLAS captures model execution on our own K3 research runtime and has published eleven measured findings from it. Shadow State is our next step: a one-way window into the running model.
Eleven measured findings published. Runtime layer in development.ATLAS gives captured model execution a structure. Tokens, depth, routes and retained internal states become connected observations—and measured results.
Meet the capture platformOne index connects the record: a run, a token, a layer, an operator. Capture coverage and retained evidence stay part of the story.
Explore the measured depthsShadow State is being designed to stream selected internal observations to an external observer. A one-way path out. No shadow memory fed back in.
Explore the runtime directionDifferent observations answer different questions. Explore what the capture pipeline records, what has been measured from it, and where the boundaries remain.
Follow captured prompt and generated-token sequences through indexed model execution. Keep positions, model identity and capture context attached to the observations.
Generated text is an output record, not guaranteed access to a model’s private reasoning. Historical short-prompt runs were capped at 32 generated tokens.
Select a word or phrase in a captured prompt. Shadow State maps it to the token positions it covers, opens the residual measurements recorded for those positions at every captured depth, and exports the trace with its source metadata.
Write a CUDA kernel for a tiled matrix multiply that stages both operands in shared memory. Explain the bank-conflict risk in the inner loop and how coalesced loads avoid it.
Highlight a word or phrase in the prompt to locate its recorded token positions.
This view’s credibility rests on saying these before anyone asks.
Move through the captured depth of the K3 research runtime. Select a depth, switch the signal, and read the published value at that coordinate—each a fraction of instrumented depth or a ratio against a stated baseline.
Every cell is a published aggregate from one capture series on the instrumented K3 runtime. Depth is a fraction of the instrumented depth; magnitudes are ratios against stated baselines. Blank cells are depths a companion experiment did not sample. No token-level rows, weights or index layouts are published.
The runtime direction: selected model observations leave through an indexed stream. Storage and analysis stay outside the model’s accessible context.
Normal model-owned
execution and state.
Selected observations.
External, indexed stream.
Archive, inspect and
evaluate monitoring.
One-way by design. No Shadow State fed into model attention. Non-interference, capture overhead and isolation need validation in each runtime.
A compact, continuous view of selected observations. Intended for baseline comparisons and escalation to richer capture.
DESIGN DIRECTION · NOT A SHIPPING TIERCapture at this depth needs control of every boundary, so K3 does not run on a public inference framework. Its weights are converted into our own proprietary format and the model is streamed from local NVMe on a single DGX system, with the capture hooks in the execution path rather than bolted on.
It is a research runtime tuned for observation, not a serving stack tuned for speed. The throughput is published so the trade is visible: everything on the findings page was captured at this pace.
Runtime access notesConverted from the source checkpoint into our own layout.
The model is streamed from local NVMe rather than held fully resident.
A single system. No cluster, no hosted API, no public inference framework.
Generation in bf16 and int8 respectively on that single system.
ATLAS has a measured model-capture research record. Shadow State extends that work toward runtime observability. The next step is replication and validation—not a stronger promise.
Ten domains, thirty-two each, in the published capture series.
Of the previous token’s experts re-activated, against 1.8% chance.
Hold 90% of the backward-pass gradient energy.
Measured on the instrumented K3 research runtime in one capture series, with the control and caveat stated per finding. Single-model, single-capture until replicated; not an independent audit or a security benchmark. Read the eleven findings ↗
Routing breadth, expert reuse, backward concentration, attention provenance, recall collapse and the depth rewrite—each with its control and caveat.
Read the findingsModel-wide Shadow State extraction, bounded streaming and complete live agent lifecycle integration.
View development boundariesOutlier detection, pre-action warning, intervention and alignment feedback remain research questions.
Read the open questionsStart with a model, a question, and the evidence
you would need to trust the result.