Four readings of one dynamical object.

A samskara is not the memory of an event — it is the geometry that an event leaves behind. Each pattern stored in the network changes the recurrent matrix, deepening the basin around that pattern. Future trajectories now bend toward it without anyone consciously remembering why. The math is identical to a recurrent network's learned-weight bias: an LSTM that has seen 10,000 customer-service transcripts will now resolve ambiguous tokens toward customer-service answers, even when the user asks about cooking. Same equation, same effect.

• math · Recurrent weight matrix W = Σₖ pₖpₖᵀ; encodes the learned attractor landscape.
• lct · Semantic Density (SD+) — accumulated condensation of meaning into a symbol that resists unpacking.
• eng · Persistent learned bias / pre-trained weights; learned priors that resist update.

Sattva is low β with low α — the landscape is shallow and forces are weak; the system drifts to its quiescent attractor. Rajas is high α — strong forcing pushes the trajectory across basin boundaries; energetic but coherent. Tamas is high β with low α — the landscape is sharp, basins are deep, no force is large enough to escape. These are not moral categories. They are distinct dynamical regimes with measurable signatures: time-to-convergence, basin width, escape probability under fixed perturbation.

• math · Control parameters (α, β): forcing magnitude α and recurrence gain β tune basin geometry.
• lct · Semantic Gravity (SG±) — the three modes parameterize landscape rigidity vs plasticity.
• eng · System tuning knobs / gain scheduling; control-loop hyperparameters.

Karma is the integral, not the integrand. It is the actual path the trajectory traced — every basin visited, every transition taken, every force absorbed. In the engine, this is the time-series of Q(t) you can replay. In an LLM, it is the chain-of-thought before the final token. In a distributed system, it is the event log. Two systems with identical W can have arbitrarily different karma if their forcing histories differ: same potential, different paths.

• math · Trajectory x(t) under dQ/dt = -Q + tanh(βWQ) + αF; integrated path through state-space.
• lct · Semantic profile / semantic wave — the temporal sequence of moves along the SG-SD axes.
• eng · Execution trace / event log; trajectory replay; the path the system actually took.

The paper's central empirical claim — out-degree > 0 ⇔ escape — is a statement about moksha. A basin from which no edge departs in the directed transition graph is, definitionally, inescapable under the available forcing budget. Adding more α (rest, more rest, even more rest) does not produce moksha; it only changes the cost of edges that already exist. Liberation requires a graph rewrite — a force pattern not yet in the basis, an introduction of new pattern p that adds an edge. This is the surprising part: the topology, not the energy, is what binds.

• math · deg⁺(v) > 0 in the directed transition graph; basin escape under unforced dynamics (α = 0).
• lct · Semantic-gravity flattening — uncoupling from a fixed code; recovering profiling capacity.
• eng · System reset / escape from deadlock; releasing a held lock; cold-start of a stuck process.

Buddhi is the level above karma. Karma is the trajectory; buddhi is the policy that picks which force to apply at each step. In control theory, this is the controller. In ML, this is the agent's policy network. In LCT, this is the rare meta-skill of consciously moving from a high-SG ground (concrete examples) to a high-SD abstraction (general principles) and back — what Maton calls a semantic wave. Buddhi is what you train. Karma is what comes out.

• math · Policy / objective function over forces F; the meta-controller that selects α and direction.
• lct · Semantic profiling — the meta-skill of moving up and down the SG axis on demand.
• eng · Loss / objective function; meta-controller; the policy that decides which intervention to apply.

Ahamkara is the fixed point. Once Q is inside this basin and β·W is large, every direction looks the same: dQ/dt is a vanishing return-to-self. The trajectory stays. The classical claim is that ahamkara is the source of suffering; the topological reading is that ahamkara is just an attractor with deg⁺ = 0 in the reachable graph. The same shape appears in an LLM whose system prompt has fixed it into a persona that cannot be argued out of, in a corporate culture that interprets every external signal as confirmation of itself, and in a person who has made their job their identity. The shape is the diagnosis. The intervention is the same: build an edge.

• math · Fixed-point attractor under high β·W with dQ/dt → 0; deg⁺(v) = 0 — no out-edges at accessible α.
• lct · Strong Semantic Gravity (SG++) — a tightly coded position resistant to re-codification.
• eng · Hard-coded identity loop / system-prompt-induced fixed point; a node refusing to release a distributed lock.