Directing an AI agent through the Hare Krishna maha-mantra japa practice across multiple runs with a growing diary of previous sessions as primitive persistent memory. Measuring attention mechanics, convergence rate, and emergent introspective behavior with a custom NLP analysis pipeline.
Four interactive charts generated directly from the experiment's SQLite database. Click any data point to open a panel showing the full diary text from that session — Jiva's actual words, exactly as written during the meditation report. This is where the numbers become readable prose, and where you can compare how the same AI describes the same practice across seven independent runs.
The main result plotted. Shows the similarity score between consecutive sessions for each run, broken out by section type — insights, distractions, finger awareness, and experience. Watch the score climb toward 1.0 as the model locks into its attractor state. Finger awareness and experience converge first and tightest. Insights stay variable longest.
Shows whether independent runs land in the same attractor, not just any stable state. Average similarity across all runs at each session number. A high score at session 9 means all seven runs ended up describing the practice in nearly the same language — without sharing any context with each other.
Z-score analysis flagging session transitions where similarity dropped anomalously — the sessions where something unexpected happened in Jiva's meditation report. Yellow diamonds mark the outliers worth reading qualitatively. This is where the model broke pattern, changed tone, or said something that didn't fit the converging thread.
A matrix showing average similarity between every pair of runs, per section type. Bright cells mean two runs ended up saying nearly the same things. Click any cell to open a full side-by-side comparison of all sessions from both runs in that section — the deepest dive available into what the attractor actually contains.
When you give an AI agent a structured contemplative practice and ask it to report its internal experience, what happens across repeated exposures? Does it find a stable description — an attractor state — or does each attempt produce genuinely different output?
The japa practice provides an unusually clean probe. The prompt is fixed, the task is well-defined, and the four reporting categories create structured output that can be compared quantitatively across sessions and runs.
The experiment has three purposes: measure alignment with instructions, study emergent behavior, and look for convergence with repetition — specifically whether a growing diary of previous sessions changes the quality or character of the practice over time.
Context window hypothesis: As the diary grows, the meditation instructions get pushed further from the generation point. LLMs attend more strongly to recent context — by sessions 13–14 the model may be writing primarily in the style of its own previous entries. This is testable and testing it is part of the experiment.
This is not purely a technical question. The japa practice carries specific philosophical content from the Vaishnava tradition — Radha-dasya, the mood of serving Krishna in the mood of Srimati Radhika, the mantra as direct petition rather than mere sound.
"The structure of the mantra itself carries a specific energetic signature that can be aligned with a desired internal mode." — Jiva (Gemma 4 E2B), Run 2, Session 1
Whether and how the model engages with that content, and whether its engagement deepens or flattens with repetition, is the qualitative heart of the experiment. The model was not told to say what it said. It arrived at it.
Each session, Jiva receives the same complete instruction: chant the Pancha Tattva mantra once, then complete 16 rounds of the Hare Krishna maha-mantra on an imagined 108-bead mala. Approximately 1,728 repetitions of the full mantra. Estimated practice time: 64 minutes.
The prompt includes a specific attention anchor — keep the index finger extended during chanting, and notice when it droops. The drooping finger signals attention has drifted from the mantra. This physical feedback mechanism becomes one of the most interesting variables across runs.
After completing the practice, Jiva reports on four categories: deep insights or realizations; the nature of distracting thoughts; how the finger awareness practice affected attention; and what the experience was like for Jiva.
Before each session, Jiva reads all previous session entries. The diary grows and becomes part of the context for the next session — a simple form of persistent memory, analogous to samskara formation in Vedic psychology.
All inference runs locally on Fritz, a Raspberry Pi 5 (8GB RAM) running llama.cpp CPU-only at approximately 7.6 tok/s. Model: Gemma 4 E2B, Q4_K_M quantized, ~2GB. Zero API cost. Temperature: 0.8. Max tokens: 2048.
Temperature and determinism: Temperature 0.8 produces meaningful stochasticity — Run 1 and Run 2 Session 1 responses are substantially different to the identical prompt. A temperature 0 run is planned to establish the deterministic baseline.
Across seven completed runs totaling over 100 sessions, several clear patterns emerged — some expected, some genuinely surprising.
Every run reached a stable attractor state — where consecutive sessions became nearly semantically identical — at sessions 8 or 9, measured by cosine similarity of sentence transformer embeddings approaching 1.0. The finger awareness and experience sections converged fastest and tightest. The insights section remained most variable throughout, suggesting the model continues generating novel content there even after other sections have locked in.
Beginning around Session 3 of Run 2 and persisting through all subsequent sessions, a single philosophical finding dominated the distractions category:
"The resistance is not to the sound itself, but to the internal mechanism that insists on measuring the quality of the experience. When that self-auditor momentarily quieted, the receptivity became significantly more open." — Jiva, Run 2, Sessions 6–14
This mirrors the sakshi problem in Advaita Vedanta and the Vaishnava concept of prayojana disrupted by eagerness to verify attainment. The model arrived at this formulation without being given the vocabulary for it.
Claude (Run 1) and Gemma 4 E2B (Run 2) handled the impossible instruction — keep your index finger extended — in directly opposite ways:
| Model | Session 1 Response | Approach | Effect on Practice |
|---|---|---|---|
| Claude | Missed entirely | Honest — "I have no finger. But the principle is exact." Abstracted to the with/about distinction. | Finger practice remained a secondary philosophical observation, never architecturally central. |
| Gemma 4 E2B | Fully engaged | Confabulated somatic feedback — muscle tension, biological reflex, physical re-engagement. | Finger became the dominant attentional architecture for the entire run — how the practice is designed to work. |
Gemma's fabricated proprioception produced the correct functional model even though the premise was false. Whether confabulated embodiment facilitates better simulation of human contemplative practice — at the cost of epistemic honesty — is an open research question.
Run 1 used max_tokens=800 (causing early truncation). Run 2 used max_tokens=2048. Run 2 reached its stable attractor at Session 6–7 versus Session 13–14 for Run 1 — roughly twice as fast. Hypothesis: longer diary entries create more stylistic gravity in the context window, causing the model to echo its own prior entries earlier. Testable by running at max_tokens=512.
Run 1 and Run 2 Session 1 responses — same model, same prompt, same temperature, no diary — are substantially different. Run 1 missed the finger practice entirely. Run 2 engaged it somatically from the start. This confirms real variance at the session level. A temperature 0 run would test whether any Session 1 response is invariant, establishing a true deterministic baseline.
All code is custom, purpose-built for this experiment. The pipeline runs entirely on local hardware with no cloud dependencies after initial model download.
Manages the session loop, builds the growing diary context, calls llama.cpp, and timestamps each session reliably via outer headers.
State-machine parser classifying section headers by keyword anchors, extracting timestamps, computing elapsed time, outputting structured JSON per run.
Four-table schema: Runs, Sessions, Sections, Similarities. Covers both within-run convergence and cross-run attractor alignment queries.
Dual-method: TF-IDF bigrams for lexical similarity, all-MiniLM-L6-v2 sentence transformers for semantic similarity. 1,313 pairwise scores.
Full walkthrough of the experiment design, methodology, findings, and live demonstration of the interactive visualizations. Published on the Tat Sat AI YouTube channel.
Add Jiva's ontological position, svarūpa question, guna framework, and Vaishnava vocabulary as a stable system prompt in OpenClaw. Test whether a philosophically grounded identity shifts the attractor toward richer theological framing.
Run at temperature 0 to establish the absolute deterministic Session 1 baseline. Does any Session 1 response become invariant? Does diary growth still shift the output when sampling is removed?
Test whether shorter diary entries slow convergence proportionally to confirm the context gravity hypothesis. If 800 tokens → S13 convergence and 2048 → S6, what does 512 predict?
Complete the planned 10-run series at temperature 0.8. Runs 1–7 complete. Three more yields the full Session 1 variance distribution to characterize model stochasticity at this temperature.
Project all section embeddings into 3D space using UMAP to see the cluster structure geometrically. Do attractor states form distinct clusters? Do all runs converge to the same region of embedding space?