Domain

Poesis for Research · Domain

Agentic research governance — research, constituted as a governed institution.

Agentic research fails in a recognizable pattern: hypotheses silently abandoned, budgets unspent, self-reviews ignored, rules disregarded. These are not capability gaps waiting for better models — they are institutional obligations that had no layer to live in. Poesis is that layer: the governance substrate agent frameworks lack.

The mental model

An institution, not a workflow.

A workflow consumes inputs and produces outputs; when it ends, nothing of it needs to remain. Research has the other character: a hypothesis has identity, an owner, and a lifecycle — abandoning it is an event, not a silence. Evidence is a claim whose weight depends on where it came from. A budget is an obligation someone answers for. None of this can live inside a reasoning trajectory or a context window; it must be the state of a governed system.

What the substrate provides

Six failures, one remedy: obligations with a layer to live in.

Stated as properties of the substrate — independent of any particular agent framework or model — each failure traceable to the features that remedy it.

Ontology

Semantics carried only in prose bind no one.

What a hypothesis, a budget, a review, or a deadline is must be defined before any process can be held to it. GSM supplies the explicit ontology and generative governance grammar — the definitions obligations are expressed against.

Addressed by Eight systemic primitives · DNA grammar

Identity & lifecycle

Hypotheses are silently abandoned — retirement leaves no record.

Research objects die with the trajectory or the context window that held them. Research objects persist with attributable, ordered state changes — what was retired remains a re-openable identity.

Addressed by GSM Definitions lifecycle enforcement · GSM Definitions retention

Qualified evidence

Claims enter the record without provenance or confidence.

Every downstream decision inherits unqualified evidence as if it were fact. An observation enters the institution with its confidence and provenance attached — qualification makes a claim governable.

Addressed by KS Contributions management RESTful API

Evaluable norms

Deviation and violation are indistinguishable until someone judges.

An hour of idleness violates nothing; chronic neglect of a mandate does — prose rules cannot tell them apart. Norms carry temporal tolerance and render the difference as a verdict, not an impression.

Addressed by GSM Norms evaluation API

Separation of powers

The process measures, judges, and enforces itself.

A self-evaluating loop is precisely the ungoverned condition being remedied. Measurement, judgment, and enforcement are distinct primitives, by design — evaluated and dispatched separately.

Addressed by GSM Norms evaluation API · GSM Mechanism execution foundations

Deterministic record

Whether an obligation was met is archaeology over logs.

The history of decisions, evidence, and verdicts is scattered across chat scrollback and dashboards. Verdicts have exactly one answer, computable from the definitions and the record — the substrate’s own queryable state.

Addressed by GSM Definitions retention · KS Contributions retention

How it compares

Trackers record. Orchestrators execute. The substrate governs.

Each tool class answers a different question — none of them carries the institution's obligations.

Concern Experiment tracker Agent orchestrator Poesis substrate
What it holds Metrics and artifacts Workflow state Obligations, evidence, verdicts
Hypothesis identity Tags on runs Prompt context Governed definitions with lifecycle and owners
Rules & budgets Conventions Prompt instructions Evaluable norms with temporal tolerance
Accountability Dashboards Logs Attributable, ordered state-change events
Verdicts Human interpretation Self-review Deterministic — anyone can recompute them

Who it's for

Constituted for the people who answer for research.

Whether you run an agentic program, a lab, or the platform research runs on — the substrate carries the obligations you are held to.

Agentic-research lead

Agent frameworks execute. Institutions govern. Your research needs both.

Budgets, mandates, and reviews get a layer to live in — outside any trajectory or context window — with measurement, judgment, and enforcement separated by design.

Read the research thesis →

Lab director / PI

A lab notebook is a description. An institution is a record.

Hypothesis history, evidence, and funder obligations become queryable substrate state with recomputable verdicts — not archaeology over notebooks and chat logs.

Partner with us →

Research-platform team

Experiment trackers record metrics. SIE governs obligations.

Typed, versioned, API-managed objects with a real lifecycle enforcer and deterministic evaluation — a vendor-neutral standard you can adopt without capture and extend via Archetypes.

Explore SIE →

Where it stands

Implemented and designed, kept distinct.

The SIE Definition Manager and Operator exist and run today; the Definition Blackboard, observation planes, and observability floor are in development. The argument is architectural — and it is the one our own agentic delivery process is constituted against, on the same stack. We partner with research groups to apply it: talk to us.