Definitions
#8am-ai#index#glossary#definitions
David Olsson
Every term used across the 8am AI corpus and pipeline, defined once. This page updates as the vocabulary grows.
the substrate
Corpus. The full record: 109 meeting transcripts and everything derived from them. Rebuildable โ delete every derived file and the pipeline reconstructs it from the transcripts.
Meeting. One Wednesday session. Stored as transcripts/MM-DD-YY/, holding the raw transcript and its derived records.
Idea. One distinct claim or topic mined from a meeting. Durable, identified IDEA-NNNN, graded, and tagged to a theme. Cites the meeting it came from.
Theme. A cluster of ideas the group returned to across multiple meetings. Identified THEME-NN. Each has a synthesis.md and, for the major ones, a published chapter.
Directive. An action item from a meeting โ "I'll send X," "we should build Y." Identified DIR-NNNN. Triaged to open, done, or dropped.
Mention. A named tool, product, or entity referenced in a meeting โ Claude, MCP, OpenClaw. Tracked per meeting to build the arrival and departure timeline.
the records
normalized.yaml. The canonical shape of a meeting after intake: participants, utterances with timestamps, and 120-second windows. Downstream steps read this, never the raw transcript, so they don't branch on format.
Utterance. One speaker turn โ who said it, when, and the text.
Window. A 120-second slice of the meeting with 30-second overlap. The unit the classifier reads.
Grade. An idea's score, 0 to 100, across six dimensions: evidence, theme strength, recency, concreteness, uniqueness, and carryover. The top 30% by grade form the queue.
the pipeline
Harvest. Pull new transcripts from Drive and the local Fireflies archive into transcripts/.
Intake. Normalize any source format โ Gemini markdown, Fireflies JSON, summary text โ into normalized.yaml.
Classify. Mine each meeting for ideas, directives, and mentions.
Synthesize. Cluster ideas across meetings into themes. Assign each idea a track.
Grade. Score every idea and rank the queue.
Draft. Compose a chapter from the highest-signal ideas in a theme. Light track.
Experiment. Build a runnable artifact from a buildable idea. Heavy track.
Publish. Push a chapter or experiment to /8am-ai.
phases and tracks
An idea moves through phases. It is never deleted.
captured. In a transcript, not yet extracted. classified. An idea record exists, tagged to a theme. researched-queued. Graded and in the working set. drafted. Written into a chapter draft. experimented. Built into a runnable artifact. published. Live on /8am-ai. backlog. Graded but below the queue cut. The reserve.
Light track. Idea to research to draft to published. Produces chapters.
Heavy track. Idea to experiment to published. Produces runnable artifacts.
terms from the conversation
Scrub. The group's 2024 word for an agent: a basic set of instructions that makes an LLM do one job reliably. The seed of everything agentic.
The four layers. The project-state model that makes harvesting repeatable: agents (read across surfaces), schema (structure the records), data (the durable corpus), time (turn records into trajectory). Harvesting becomes intelligence only when all four are present.
Project-state. A structured substrate that gives a project memory: a corpus that is also a map, so a claim can be traced to where it came from.
Signal harvesting. Reading across everything that's been said to surface the connective tissue no single person can track.
Atomic to architecture. The shift from doing tasks to designing the systems that do tasks. The spine the corpus hangs on.
How do you know. The corpus's central question. Not "is the output good" โ what in your process tells you it's good, independent of the thing that produced it.
Agentic layer. Treating after-the-fact agent work as a default property of how work gets done, not a one-off.
Slop code. Disposable, fast-shipped AI-generated code. Embraced in 2026 as a competitive necessity, on the condition that the control plane and documentation are kept clean.
OpenClaw. The autonomous-agent pattern that, in early 2026, marked the swarm inflection โ directing fleets of agents instead of doing the work.
The index pages, state of the corpus and theme engagement over time, use these terms. The chapters put them to work.