Lucid Workflows
Workflows are the procedural layer of Lucid reasoning — structured sequences that make the steps of an inquiry explicit, repeatable, and inspectable. They transform the DCR cycle from a cognitive architecture into a running procedure.
The canonical Lucid workflow is a six-step sequence organised around three reasoning phases: divergent (Observation, Hypothesis), navigational (Evidence, Interpretation), and convergent (Synthesis, Reflection). Every step preserves stance differentiation — the distinct epistemic positions that make multi-agent reasoning more than redundant parallelisation.
AI workflows in current practice are predominantly task pipelines — structured sequences for routing data through processing steps. The structure optimises efficiency: steps are minimised, parallelised where possible, and evaluated by throughput and output quality. The reasoning process that produces the output is incidental to the pipeline's design.
A Lucid workflow is designed for epistemic quality, not throughput. The steps exist to protect the integrity of the reasoning process — not to optimise the production of output.
This means that Lucid workflows have properties that task pipelines do not: they are stance-differentiated throughout, meaning that the distinct epistemic positions of the agents involved are maintained as separate tracks — not merged until synthesis. They are inspectable, meaning that the reasoning at each step is visible, not hidden inside opaque processing. And they include a reflection step, which is the procedural form of metacognition — a structural commitment to understanding how the reasoning proceeded, not just what it produced.
The result is a workflow that produces not just outputs but understanding — understanding that can be stored in Memory, surfaced through Interaction, and built on in subsequent reasoning episodes.
Six steps across three phases. Steps 1–2 open the epistemic field (divergent). Steps 3–4 move through it (navigational). Steps 5–6 integrate and review (convergent).
What is present in the epistemic field, recorded without interpretation. Observation is not passive recording — it is active, stance-aware noticing. Different stances will foreground different observations from the same material. The observation step makes this explicit: what each stance notices is tracked, not collapsed into a unified set of "what is there." The product of observation is a structured set of stance-differentiated noticings.
Candidate explanations, framings, and interpretive propositions generated from the observations. Hypotheses are not guesses — they are provisional structural claims about what the observed material means or how it is organised. In a multi-agent workflow, each stance generates hypotheses that reflect its interpretive orientation. The hypothesis step is the first place where divergent coverage matters: the richer the hypothesis space, the stronger the subsequent inquiry.
Structured gathering and evaluation of material that bears on the hypotheses in play. Evidence is gathered in relation to specific hypotheses — not as a general accumulation but as a directed inquiry. Each piece of evidence is evaluated under each stance: what this evidence means, whether it supports or challenges a hypothesis, depends on the interpretive position from which it is assessed. Evidence provenance — which stance evaluated it, how — is preserved.
Meaning-making from the gathered evidence under each stance. Interpretation is explicitly stance-dependent — the same evidence base will yield different interpretations under different stances, and this is a feature of the workflow, not a defect. The interpretation step produces a structured set of stance-differentiated meaning-claims, each with its evidential basis explicit. The product is not a single interpretation but a mapped interpretation space.
Integration across the interpretation space — the convergent move that draws together the stance-differentiated interpretations into a structured understanding. Synthesis is not averaging or majority-voting. It is the identification of what holds across stances, what the tensions between stances illuminate, and what the most structurally grounded position is given the full evidence base. Synthesis may leave some tensions unresolved — these become the basis for the next reasoning episode.
Metacognitive review of the reasoning process itself — not of the content produced, but of how the reasoning proceeded. Reflection asks: were the stances adequate to the epistemic field? Were there gaps in observation, hypothesis generation, or evidence gathering? Was synthesis premature or forced? Were avoidance patterns active? The reflection step is the feedback loop that makes workflows improve over time — its output informs both memory and the calibration of future orchestration.
Four properties that distinguish Lucid workflows from task pipelines and make them genuine epistemic infrastructure.