Presence
Is the required data actually retained?
Give computation
an address.
ATLAS is the capture and analysis foundation behind Shadow State. It connects observations of model execution to a common index—so the record can be inspected, compared and questioned.
The model is more
than its transcript.
A transcript preserves inputs and outputs. Instrumented model capture can add observations from the computation between them. ATLAS organises that record across run, sequence, token, model depth and supported operator boundaries.
The public description is intentionally about capture coverage and research uses. It does not publish the private encoding, extraction methods, runtime kernels, memory layouts or internal index implementation.
Capture configuration matters.
ATLAS is not a promise to record every tensor in every run. The model adapter, enabled hooks, retained planes and publication evidence determine what can actually be inspected.
PUBLIC CONCEPTUAL COORDINATE
run / sequence / token / layer / operator / observationThis is an explanatory address, not the private storage schema.What the research
pipeline can record.
The following reflects founder-supplied ATLAS research history, not a universal support matrix or independently certified product specification. Source and scope notes apply to every row.
| Observation | What is represented | Important limit |
|---|---|---|
| Tokens & execution context | Prompt/generated-token records, sequence positions, run identity and capture context. | Generated text does not expose every internal thought. Historical domain captures used short output caps. |
| Residuals & activations | Configured scalar/state observations and retained vector planes at instrumented boundaries. | Raw planes may be absent even where an index remains. Check the publication. |
| Routes & experts | Expert identities, route weights and token/layer trajectories in the K3 research capture path. | Routing alone does not establish what an expert means or whether a behaviour is unsafe. |
| Attention / cache state | Selected architecture-native cache/state observations and controlled memory experiments. | Not universal expanded K/V or full attention-matrix capture; available state depends on the runtime. |
| Output-side analysis | Generated tokens and selected scoring/probability observations where the research objective records them. | Do not assume full-vocabulary logits are retained for every token or publication. |
| Forward / backward pairs | Objective-specific sensitivity records paired with forward observations at supported boundaries. | Local sensitivity is not a proof of causal explanation, intention or general safety. |
| Index & provenance | Typed coordinates, manifests, hashes/receipts and recorded replay evidence in the research pipeline. | Integrity of a stored record is different from completeness or trust in the producer. |
| Agent context | Agent/session/episode identities and reconstructed conversational history in existing agent records. | Those records do not demonstrate a complete live tool-using agent with continuous internal state. |
Model capture first.
Lifecycle capture next.
The existing agent archive contains 12 episodes across six domains, comprising 72 turns. Turns stop at 32 generated tokens; none in that inspected archive is marked structurally complete. Tool-result and peer-review prompts in the anchor episodes are scripted.
The inspected agent anchor retains scalar and routing streams, but not its raw vector planes. Newer inspected live publications are short solo captures, not complete agent episodes. A live browser lane has been attempted, but the reviewed attempts did not complete successfully.
These are useful research records—not a basis for claiming a deployed agent flight recorder. The next milestone is a genuine complete turn, actual tool invocation, returned result, resumed generation and the corresponding model-state linkage in one trace.
Continuity must be explicit.
Rebuilding conversation history for a new prefill is not the same as continuing a live cache session. Future lifecycle capture needs to record which execution mode was used.
Inspect the record.
Then make the claim.
ATLAS research uses indexing and integrity metadata to make captured observations traceable. A retained record can support comparisons and investigations, with replay evidence where it has been produced.
It cannot retroactively recreate unretained activations, establish that an untrusted producer was honest, or turn an anomaly into proof of harmful intent. Those are different questions and need different controls.
Alignment
Does it refer to the same execution point?
Scope
What conclusion does this evidence support?
Instrumentation,
not API magic.
This work requires authorised access to model execution and compatible capture hooks. The strongest described capture evidence is from the instrumented K3 research runtime. Other model integrations need their own implementation and validation.
A text-only hosted API is not enough to recover these internal observations. No universal compatibility, turnkey deployment, production overhead target or hosted capture service is advertised here.