ATLAS model capture
Instrumented model-level capture, a common execution index, routing/expert trajectories and experiment-specific forward/backward and memory analysis. Data coverage varies by publication.
A clear line between
now and next.
There is a difference between observing a computation and understanding it. Between identifying an outlier and preventing harm. Our research starts by keeping those distinctions intact.
Established internally.
Scoped publicly.
The descriptions on this site are based on the founder’s supplied ATLAS experiment records and the subsequent agent-capture inspection. They are not independently audited, peer-reviewed performance claims. Private traces and implementation details are not published on this website.
R1 / CROSS-DOMAIN MODEL CAPTUREINTERNAL RESEARCH RECORD
- Domains
- Python; JavaScript / Node; React; ASP.NET / C#; Rust / C; CUDA / Triton.
- Prompt series
- 42 prompts per domain; 252 prompts in total.
- Model depth
- The recorded K3 capture architecture spans 93 layers.
- Output scope
- 32 generated tokens per prompt. This is short-prompt capture coverage, not completed coding tasks or agent episodes.
- Evidence scope
- Indexed routing and configured model observations. Forward/backward, memory and replay work have their own capture configurations and experiment scopes.
R2 / AGENT ARCHIVE INSPECTIONPARTIAL CAPTURE
12 episodes, 72 turns, six domains. Scripted tool/peer prompts, truncated turns and reconstructed history. Raw vector planes are absent from the inspected anchor. No complete, successful live tool lifecycle with continuous state was established by the reviewed archive.
Read the agent boundaryR3 / FORWARD-BACKWARD & MEMORY RESEARCHEXPERIMENT-SPECIFIC
Separate research includes objective-linked paired capture, selective sensitivity analysis, cache/state experiments and retained evidence/replay work. We do not publish private codecs, state-construction methods, extraction recipes or runtime layouts.
The foundation exists.
The runtime layer is next.
ATLAS Shadow State
A one-way, external runtime representation fed by selected model hooks. Intended to stream compact observations through bounded buffers into archive and analysis consumers. Non-interference and overhead require measurement.
Complete live agent capture
Connect full generations, actual tool calls/results and subsequent model execution through the ATLAS index. Preserve session mode and distinguish reconstructed history from persistent model state.
Reliable safety monitoring & control
Pre-action warning, unseen-threat detection, false-positive rates, multi-agent intervention and alignment gains have not been demonstrated by the evidence on this site.
Observe carefully.
Validate adversarially.
Does internal evidence add information?
Compare internal observations with transcript and action baselines. Use held-out tasks and strategies; do not confuse task recognition with detection of unsafe behaviour.
How early is a useful signal available?
Evaluate detection relative to an actual action boundary, with monitoring lag and incomplete streams accounted for. Retrospective separation is not the same as a timely control.
What changes when capture is enabled?
Measure token, routing and state equivalence alongside throughput, memory and storage cost. One-way data flow is a design constraint, not proof of zero overhead.
How does the system fail?
Test benign novelty, missing capture ranges, stream pressure, compromised collectors and model-adapter changes. An outlier alone should not be called a threat.
Can investigated traces improve the next iteration?
Study offline adjudication and held-out evaluation before feeding results into model or monitor updates. A feedback loop can also reinforce errors or learn spurious shortcuts.
A shared problem.
Independent work.
AISI’s published white-box control work discusses observing or modifying internal activations as a complement to external behavioural controls. That is relevant research context—not evidence that AISI has evaluated, partnered with or endorsed ATLAS.
UK AI SECURITY INSTITUTEWhy we’re working on white box controlRead the primary source ↗UK AI SECURITY INSTITUTEResearch agendaRead the primary source ↗
These organisations are sources, not customers or partners. No affiliation, validation or endorsement is implied.
Explain the capability.
Protect the implementation.
The website describes observations, evidence boundaries and research direction. The animated cosmos, example matrices and trace explorer are original, synthetic explanatory visuals. They contain no live model telemetry.
Private capture data, model weights, codecs, matrix construction, internal index layouts, extraction kernels, thresholds and security controls are not distributed here. This site is not an operational monitoring console.