Terms

Substrate

Substrate — the shared frame-graph medium a population of agents reads from and writes back into. What makes it a substrate rather than a static scaffold is the write-back feedback: each write reshapes what other agents retrieve on the next turn.
Coupling — a connection between frames that strengthens when they are retrieved and acted on together (Hebbian) and relaxes when they are not. Carries a valence: alignment, interference, or neutral.
Membrane — a dynamically adjusted boundary over a densely co-activated region of frames and the agents retrieving over them, surrounded by sparser coupling. At a given step it delineates a co-working population of agents and their active context from the rest of the substrate. Membranes nest and overlap rather than partition — membranes over membranes — so a membrane is a boundary you draw at a moment, not a fixed group. Membranes are where the substrate’s influence is exercised — they modulate what context crosses into an agent’s next turn — so interventions act on membranes, not on agents.
Coordination phase — the dynamical regime a population of agents over the substrate currently occupies: exploration, stabilization, lock-in, or drift. “Phase” in the phase-transition sense, not the oscillator sense.
Divergence/convergence cycle — the healthy oscillation between opening the space (divergence) and closing it around something usable (convergence). Pathologies are cycle failures: premature convergence (lock-in) or failure to converge (drift).

Construction

Frame — the primitive: a typed unit of knowledge with named slots filled from content. A frame can be as small as a sentence-level claim or as large as a whole document, and frames nest.
Frame type (schema) — the template a frame instantiates: a named kind of knowledge unit (claim, evidence, method, …) declaring which slots to expect and what may fill them. Types live in a registry, not in application code.
Slot — a named, typed relationship on a frame, expecting a particular filler (a frame of some type, or a literal). Typed slots make the graph queryable by structure, not only by similarity; an unfilled slot carries a default and marks what is missing.
Relationship (edge) — a typed link between frames (SUPPORTS, CONTRADICTS, HAS_CHILD, …), assigned as frames are extracted and grounded in the frame type’s own expectations.
Composition — frames combine along two axes: structurally (paragraph → section → document) and semantically (claim + evidence + source → argument support structure). Semantic composition may cross document boundaries.
Registry — the store of frame type and operation definitions, itself held as frames on the graph. It is self-extending: new types can be proposed at extraction time.
Traversal instruction — the navigation semantics a frame type carries through its slot structure: which relationships it expects, and toward which types. The same per-type expectations that lay down edges also guide how the graph is later traversed.

From the notes

Pulled automatically from each note’s definition — hover a link for the full note.

Coordination phase — The dynamical regime a population of agents over the substrate currently occupies — exploration, stabilization, lock-in, or drift. Phase in the phase-transition sense, not the oscillator sense.
Coupling — A weighted connection between two frames that strengthens when they are retrieved and acted on together (Hebbian) and relaxes when they are not. Coupling carries a valence — alignment, interference, or neutral — and its density over a region is what a membrane is drawn around.
Divergence/convergence cycle — The healthy oscillation between opening the space (divergence) and closing it around something usable (convergence). The pathologies are cycle failures — premature convergence (lock-in) or failure to converge (drift).
Experiments — The experiments index — small, fully observable runs that validate substrate signals, each with its own setup note under experiments/. A mechanism earns a place only after it shows a measurable, reproducible effect on a controlled problem; a signal counts only if it precedes the visible failure with a usable lead and beats trace-level baselines.
Fractal composition — The substrate is assembled by the same compositional move at every level — frames nesting into frames — so its structure is roughly self-similar across scales.
Frame — The primitive of the substrate — a typed knowledge unit with named slots, instantiated from content rather than drawn from a fixed ontology. A frame can be a sentence-level claim or a whole document, and frames nest.
Frame-type diversity — The variety of frame types active in a region of the substrate — the most direct read-out of where a population sits in the divergence/convergence cycle.
Gini coefficient — A single number (0–1) for how unequally activation is concentrated across frames — high values flag a few frames hoarding the energy.
Hill diversity — Effective counts of frame types at several orders (richness, Shannon, inverse-Simpson) — a diversity reading of whether rare framings survive or get drowned out.
Membranes — A dynamic, selectively permeable boundary in the cell-membrane sense — its channels open and close on a measurable quantity rather than a fixed rule. Over the substrate it is a temporary boundary drawn around a co-active region (the frames and agents working together) by coupling density; membranes nest and overlap rather than partition.
Oscillators — We model each frame as an oscillator rather than a flat activation score — a thing with both a loudness (amplitude, how active it is) and a timing (phase, where it is in its cycle). Phase is what lets us ask whether two active frames are active together; synchronization is the native quantity, and the convergence and lock-in we want to read are synchronization phenomena.
Resonance — A measure on the substrate that combines the coupling frames have built by use (Hebbian) with their semantic coherence. Resonant units are retrieved together — a retrieval is a threshold on resonance, and lowering the threshold reaches deeper into a subgraph, returning finer detail. Concretely it is a setting in the re-ranker.
Scale — Reading the substrate as a topography at many zoom levels — from the whole store down to a single concept — where the boundaries between levels are membranes that nest and overlap rather than nest neatly.
Substrate — The shared medium a population of agents reads from and writes back into — memory, context, retrieval, and tool state — where each write reshapes what every other agent retrieves next. The write-back feedback is what makes it a substrate rather than a static scaffold.
Task-appropriate behavior — How much diversity a task should produce — we can measure a task’s spread, but cannot yet say what spread is right for it. Still open.

Substrate

Substrate — the shared frame-graph medium a population of agents reads from and writes back into. What makes it a substrate rather than a static scaffold is the write-back feedback: each write reshapes what other agents retrieve on the next turn.

Coupling — a connection between frames that strengthens when they are retrieved and acted on together (Hebbian) and relaxes when they are not. Carries a valence: alignment, interference, or neutral.

Membrane — a dynamically adjusted boundary over a densely co-activated region of frames and the agents retrieving over them, surrounded by sparser coupling. At a given step it delineates a co-working population of agents and their active context from the rest of the substrate. Membranes nest and overlap rather than partition — membranes over membranes — so a membrane is a boundary you draw at a moment, not a fixed group. Membranes are where the substrate’s influence is exercised — they modulate what context crosses into an agent’s next turn — so interventions act on membranes, not on agents.

Coordination phase — the dynamical regime a population of agents over the substrate currently occupies: exploration, stabilization, lock-in, or drift. “Phase” in the phase-transition sense, not the oscillator sense.

Divergence/convergence cycle — the healthy oscillation between opening the space (divergence) and closing it around something usable (convergence). Pathologies are cycle failures: premature convergence (lock-in) or failure to converge (drift).

Construction

Frame — the primitive: a typed unit of knowledge with named slots filled from content. A frame can be as small as a sentence-level claim or as large as a whole document, and frames nest.

Frame type (schema) — the template a frame instantiates: a named kind of knowledge unit (claim, evidence, method, …) declaring which slots to expect and what may fill them. Types live in a registry, not in application code.

Slot — a named, typed relationship on a frame, expecting a particular filler (a frame of some type, or a literal). Typed slots make the graph queryable by structure, not only by similarity; an unfilled slot carries a default and marks what is missing.

Relationship (edge) — a typed link between frames (SUPPORTS, CONTRADICTS, HAS_CHILD, …), assigned as frames are extracted and grounded in the frame type’s own expectations.

Composition — frames combine along two axes: structurally (paragraph → section → document) and semantically (claim + evidence + source → argument support structure). Semantic composition may cross document boundaries.

Registry — the store of frame type and operation definitions, itself held as frames on the graph. It is self-extending: new types can be proposed at extraction time.

Traversal instruction — the navigation semantics a frame type carries through its slot structure: which relationships it expects, and toward which types. The same per-type expectations that lay down edges also guide how the graph is later traversed.

From the notes

Pulled automatically from each note’s definition — hover a link for the full note.

Coordination phase — The dynamical regime a population of agents over the substrate currently occupies — exploration, stabilization, lock-in, or drift. Phase in the phase-transition sense, not the oscillator sense.

Coupling — A weighted connection between two frames that strengthens when they are retrieved and acted on together (Hebbian) and relaxes when they are not. Coupling carries a valence — alignment, interference, or neutral — and its density over a region is what a membrane is drawn around.

Divergence/convergence cycle — The healthy oscillation between opening the space (divergence) and closing it around something usable (convergence). The pathologies are cycle failures — premature convergence (lock-in) or failure to converge (drift).

Experiments — The experiments index — small, fully observable runs that validate substrate signals, each with its own setup note under experiments/. A mechanism earns a place only after it shows a measurable, reproducible effect on a controlled problem; a signal counts only if it precedes the visible failure with a usable lead and beats trace-level baselines.

Fractal composition — The substrate is assembled by the same compositional move at every level — frames nesting into frames — so its structure is roughly self-similar across scales.

Frame — The primitive of the substrate — a typed knowledge unit with named slots, instantiated from content rather than drawn from a fixed ontology. A frame can be a sentence-level claim or a whole document, and frames nest.

Frame-type diversity — The variety of frame types active in a region of the substrate — the most direct read-out of where a population sits in the divergence/convergence cycle.

Gini coefficient — A single number (0–1) for how unequally activation is concentrated across frames — high values flag a few frames hoarding the energy.

Hill diversity — Effective counts of frame types at several orders (richness, Shannon, inverse-Simpson) — a diversity reading of whether rare framings survive or get drowned out.

Membranes — A dynamic, selectively permeable boundary in the cell-membrane sense — its channels open and close on a measurable quantity rather than a fixed rule. Over the substrate it is a temporary boundary drawn around a co-active region (the frames and agents working together) by coupling density; membranes nest and overlap rather than partition.

Oscillators — We model each frame as an oscillator rather than a flat activation score — a thing with both a loudness (amplitude, how active it is) and a timing (phase, where it is in its cycle). Phase is what lets us ask whether two active frames are active together; synchronization is the native quantity, and the convergence and lock-in we want to read are synchronization phenomena.

Resonance — A measure on the substrate that combines the coupling frames have built by use (Hebbian) with their semantic coherence. Resonant units are retrieved together — a retrieval is a threshold on resonance, and lowering the threshold reaches deeper into a subgraph, returning finer detail. Concretely it is a setting in the re-ranker.

Scale — Reading the substrate as a topography at many zoom levels — from the whole store down to a single concept — where the boundaries between levels are membranes that nest and overlap rather than nest neatly.

Substrate — The shared medium a population of agents reads from and writes back into — memory, context, retrieval, and tool state — where each write reshapes what every other agent retrieves next. The write-back feedback is what makes it a substrate rather than a static scaffold.

Task-appropriate behavior — How much diversity a task should produce — we can measure a task’s spread, but cannot yet say what spread is right for it. Still open.