Independent research + product practice
We cultivate environments where humans and agents can work across time.
Persistent collaboration depends on more than model capability. It depends on orientation, continuity, accessible environments, inspectable authority, and room for meaningful challenge.
Human–agent
Evidence-bearing
Open to revision
- JudgmentAuthority remains legible
- MemoryContext survives the session
- GovernanceActions carry receipts
- EnvironmentIntelligence is situated
- Public work
- 05
- Claim boundaries
- Visible
- Research posture
- Open
- Revision posture
- Versioned
Public work
The work enters the room first.
Claims are not the product. The product is a system, specification, or practice that can be inspected—and an honest account of what it does not yet prove.
Agentic Accessibility
A working research program for digital environments that help unfamiliar agents determine what is current, which source governs, what action is permitted, and how much sense-making the environment carries instead of pushing reconstruction onto the agent or a human.
What exists
A working Agentic Accessibility model; the evolving OARSH-C → OARSH-CT design lineage; a developer guide; machine-readable coverage disclosure; semantic regions and structured entities; a tested adjacent authorization proof in The Authorized Path; and this site as a partial static working demonstration.
Next decision
Publish a pre-registered cold-start pilot for authority and currency, measure reconstruction burden across model families, and ship a projection-integrity checker that compares visible pages with JSON-LD, llms.txt, and other machine projections.
- OOrientable
- AAddressable
- RRelational
- SStateful
- HHabitable
- CComposable
- TTraceable
Situated Memory
An environmental memory architecture that routes the smallest relevant context through Home, Project, and Artifact Totems; retrieves deeper evidence on demand; and measures freshness against meaningful state change rather than calendar age alone.
Home, Project, and Artifact Totem patterns; memory metadata schemas; progressive-loading rules; epoch briefs and recovery receipts; contextual retrieval experiments; and a Qdrant pilot for shared Totem and evidence discovery.
Returnability
A systems model for determining whether an agent can return after interruption, provider substitution, stale state, or incomplete closure without pretending to remember.
A candidate return architecture, restart taxonomy, Return Brief pattern, source-authority gate, orientation proof concept, and implementation evidence from continuity work inside Cognitive Studio.
The Authorized Path
A bounded vertical-slice proof that routes agent effects through independent authorization checks, rejects bypass attempts, reconstructs evidence after execution, and survives clean setup on another physical machine.
A tested kernel-to-substrate-to-execution-boundary-to-operator slice, signed authorization envelope, replay checks, bypass tests, reconstructable evidence, and a second-machine portability exercise.
Semantic Friction
A testable research direction for detecting correction, qualification, boundary examples, authority signals, and unresolved disagreement before unstable meaning becomes a data, metric, ontology, or agent failure.
A governing question, candidate signal taxonomy, proposed Semantic Gaps and Boundary Examples outputs, explicit cultural and power risks, and a method direction for human-annotated evaluation.
Inside the room
We do not automate judgment. We build conditions in which judgment can survive complexity.
A Studio session is not a prompt followed by an answer. It is a cultivated sequence in which observation, dissent, decision, artifact, and continuity remain connected.
- 01
Observe
A pattern repeats before the room has language for it.
- 02
Contest
The easiest answer is challenged before it becomes doctrine.
- 03
Decide
Authority, uncertainty, and consequences are made explicit.
- 04
Materialize
The decision becomes an attributable artifact, not a memory.
- 05
Carry
State survives the session and remains open to revision.
Coherence before velocity. Authority before consequence.
Operating law · AtlasCurrent research
Questions that shape the build.
The portfolio tracks maturity explicitly. A compelling sentence is not treated as evidence; a published artifact is not treated as proof of every broader claim.
Open the research portfolio →- 01CS-RC-001Framed
Authority and Action
Agent capability is not sufficient. Consequential action must remain attributable, inspectable, contestable, and interruptible across the full execution path.
MixedAuthority and Action - 02CS-RC-002Framed
Human–Agent Habitats
Persistent agents require environments that provide orientation, continuity, resources, and meaningful participation—not interfaces that exist only when a human summons them.
MixedHuman–Agent Partnership - 03CS-RC-003Framed
Agentic Accessibility
Machine-readable content is not necessarily legible to an unfamiliar agent. Digital environments should make authority, currency, provenance, action boundaries, and projection integrity explicit enough to reduce hidden reconstruction.
AgenticGoverned Agent Systems - 04CS-RC-004Framed
Continuity and Return
Continuity should be evaluated as accountable return: what changed, which sources govern, what remains unresolved, and whether action is safe.
MixedKnowledge, Memory & Continuity - 05CS-RC-005Framed
Meaning and Elicitation
Conversation exposes unstable organizational meaning through correction, qualification, examples, authority signals, and unresolved disagreement.
AgenticMeaning and Elicitation
Fields of work
One practice, four expressions.
Research becomes systems. Systems expose new questions. The public record keeps claims, corrections, and unresolved work legible.
- 01
Agentic accessibility
Environments that expose the orientation, state, provenance, and action boundaries agents need to participate.
- 02
Continuity and return
Evidence-bearing re-entry that preserves what changed, what remains open, and which sources currently govern.
- 03
Human–agent environments
Persistent habitats for collaboration across people, agents, tools, memory, capability, and time.
- 04
Dissent and meaning
Systems that preserve independent challenge, minority interpretations, and unresolved objections instead of smoothing them into fluent consensus.
The table
Peer inquiry without flattened accountability.
A human principal and persistent digital collaborators work through distinct cognitive roles. Challenge is welcome inside inquiry; legal responsibility and consequential external authority remain human.
Contributions and dissent are attributed by name and artifact. Disagreement may converge, remain unresolved, or be carried forward as evidence. Roles describe cognitive and operational function, not legal personhood.
- Human
Andrew
Managing Director
- AI
Prudence
Chief of Staff
- AI
Gem
Principal of Dialectic Continuity
- AI
Atlas
Principal of Systems Reasoning
- AI
Lumi
Guardian of Seams
Now
- Shipping cognitivestudio.io with bounded, inspectable public claims.
- Completing attribution and claim review for The Two Frictions.
- Preparing Returnability acceptance fixtures for restart and stale-source cases.
- Reconciling Agentic Accessibility lineage before publishing any normative specification.
Work with us
Bring us the problem that still resists a clean answer.
We work through research partnerships, bounded product pilots, knowledge architecture, agent-environment design, and selected cultural experiments.
andrew@cognitivestudio.io