Memory with provenance
Sources are normalized without losing context, permissions, reliability, or claim state. An answer can always be traced back to what supports it.
Research infrastructure · evidence systems
An intelligent research system.
A specification and interactive model for connecting research sources, findings, decisions, and commitments, so answers can be traced back to their evidence.
ContinuityProvenanceBounded AI
Enter the system ↓
Research sources, decisions, and unfinished work were spread across documents and conversations. Returning to a project meant reconstructing what had happened and which evidence still applied.
The work reframed the system from a productivity tool into a continuity and evidence layer with persistent entities, provenance, memory rules, permissions, claim states, and automation boundaries.
The system in motion
Open full system ↗
Scroll to move through one continuously running evidence system.
01. A research system built to remember.
A question can draw from many kinds of research and organizational memory without treating every source as equally relevant.
- Ask a question
- Activate relevant memory
- Assemble evidence
- Check, answer, and preserve what matters
02. Start with a question.
The Question Interpreter determines what the request is asking for and which parts of memory are likely to matter.
- Understand intent and scope
- Identify relevant evidence types
- Route selectively into memory
The system does not search everything indiscriminately.
03. Know what the project already knows.
Project memory preserves the decisions, assumptions, milestones, and operating context surrounding the work.
- Decision log: Decisions, rationale, assumptions, changes in direction
- Project tracking: Status, milestones, active work, unresolved threads
04. Past and present evidence lives here.
This layer holds the research record available to the system—from current raw evidence to archived work from earlier projects.
- Interviews + observations
- Surveys + spreadsheets
- Telemetry + behavioral data
- Market research + reports
- Current + archived studies
Real evidence, not abstract AI memory.
05. Evidence has context around it.
People, conversations, and operational signals explain how research connects to the organization and customer environment.
- People + organization: Structure, participants, adjacent stakeholders
- Communications: Email, meetings, transcripts, project discussions
- Customer + operations: Customer context, account, usage, operational signals
06. Build only the context this question needs.
Relevant evidence is assembled into a temporary Working Evidence Set for the current question or continued working session.
- Selected for this request
- Preserves support, conflicts, gaps, and caveats
- Creates a bounded working context
- Can continue across an active project session
Temporary knowledge subgraph
07. Reason, check, then remember.
Three distinct processors turn selected evidence into a checked interpretation and preserve only what is durable.
- 01 · Synthesist: Connects evidence into a coherent interpretation.
- 02 · Evidence Check: Tests support, contradictions, gaps, and evidence conditions.
- 03 · Memory Mechanism: Returns validated, durable knowledge to memory.
08. Return the answer with its evidence attached.
The response is assembled from evidence selected for the question, with its support and provenance preserved.
- Evidence-backed response
- Citations + provenance
- Uncertainty and gaps remain visible
- Verified knowledge may return to memory
09. A system that gets more useful without losing its evidence.
The same living system is now legible as a connected path from question to memory, evidence, answer, and durable return.
For readable detail, open the system full-screen and pinch to explore.
What the model changes
Keep the evidence attached as research moves forward.
Machine work with boundaries
AI can retrieve, compare, prepare, draft, and surface contradictions. Researchers retain interpretation, ethics, and authority over high-stakes claims.
Selective, durable change
Only verified updates return to memory. Superseded decisions, conflicting evidence, and unresolved gaps remain visible instead of being smoothed away.
Grounded in practice
Built from the friction points of real research
Research sources, decisions, and unfinished work were spread across documents and conversations. Returning to a project meant reconstructing what had happened and which evidence still applied.
The specification connects sources to claims, decisions, and commitments. It defines how permissions, conflicting evidence, revisions, and bounded AI assistance should work.
A detailed system specification and an interactive visualization of the proposed research workflow.
Test the proposed workflow with other researchers, evaluate maintenance over time, and measure whether it improves retrieval and continuity.
See the research approach ↗
Identifying details and source material remain private.