maya-metrics-nll 1.0.0

Affective neuromorphic evaluation library for SNN continual learning

maya-metrics: Affective Neuromorphic Evaluation

Venkatesh Swaminathan · Nexus Learning Labs, Bengaluru, India · M.Sc. DS&AI, BITS Pilani

📄 Paper · 🎛️ Interactive Dashboard · ❓ FAQ · 🌐 Research Hub

---

What is this?

maya-metrics is the companion evaluation library to the Maya Research Series — the first open-source Python library for evaluating internal affective and neuromodulatory dynamics in neuromorphic spiking neural networks (SNNs) engaged in class-incremental continual learning (CIL).

The Maya Research Series (P1–P9, Maya-mPCI, Maya-LLM) implemented all fourteen dimensions of the Advaita Vedantic Antahkarana as independently falsifiable computational mechanisms. This library gives the community the tools to measure those mechanisms in any future architecture.

"cl-metrics measures what happened. maya-metrics measures how it felt."

Validated against 11 published preprints, two substrates (SNN + LLM), one embodied PiCar-X deployment. 16/16 unit tests pass in 1.56 seconds.

---

The One-Paragraph Summary

maya-metrics accepts a Maya batch log CSV and returns standardised measurements of every affective dimension: whether Bhaya quiesced under replay (the Quiescence Law), whether Buddhi followed its sigmoid to R²=1.0000 (S-curve determinism), whether Vairagya grew monotonically (Pearson r=0.9228), whether Karma drove pruning while Vairagya moderated it (D★: partial_r=−0.808), whether the mPCI delta survives depth-matched decoupling (Δ=−0.0427, threshold=0.0214), and whether the network's Structural Age Index classifies it as infant, maturing, or adult (P9 Condition D: SAI=0.824, Maturing). Six modules, one dependency (numpy), stateless pure functions that accept a CSV and return metric dictionaries. No framework coupling. No training loop.

---

Key Results — P9 Canonical Run (Condition D · seed=42 · Split-CIFAR-100 · 10 tasks)

Metric	Result	Expected	Status
Bhaya Quiescence Law	replay_max = 0.000000, rate = 100%	0.000 under replay	✅ CONFIRMED
Buddhi S-curve R²	1.0000 · ceiling task 1	> 0.95	✅ CONFIRMED
Vairagya Pearson r	0.9228 (monotonic)	Positive growth	✅ CONFIRMED
D★ partial_r	−0.808 (Karma-controlled)	< −0.5	✅ CONFIRMED
Karma monotonic	r = 0.9997 · DECAY = 0.002315	Monotonic	✅ CONFIRMED
Shunyata pruned	84.3% monotonic	Monotonic growth	✅ CONFIRMED
mPCI delta	−0.0427, threshold = 0.0214	> 2× std	✅ CONFIRMED
Depth-matched decoupler	2/3 seeds below baseline	Honest per-seed	~ PARTIAL
Structural Age Index	0.824 — Maturing (transition T5)	SAI ∈ [0,1]	✅ CONFIRMED
Cross-substrate DTW	0.200 < 0.300 threshold	Buddhi invariant	✅ CONFIRMED
Unit tests	16/16 passing in 1.56s	All green	✅ CONFIRMED

---

Two Series Constants — Confirmed Across All 11 Papers

🔹 Bhaya Quiescence Law

Bhaya = 0.000 under replay from Task 1 onwards — every paper, every ablation condition, every seed. Eleven consecutive confirmations.

am = AffectiveMetrics(df)
result = am.bhaya_quiescence_score()
# → {"law_confirmed": True, "replay_max": 0.000000, "quiescence_rate": 1.0}

Bhaya is now a programmatic unit test. If it fires non-zero under replay, pytest fails — real-time catastrophic forgetting monitor for any deployed Maya system.

🔹 Buddhi S-Curve Determinism

Identical 0.750 → 1.000 sigmoid trajectory P4–P9, all ablation conditions, all seeds, SNN and LLM substrates. DTW_norm(SNN↔LLM) = 0.200 < 0.300.

result = am.buddhi_scurve_fit()
# → {"r_squared": 1.0000, "deterministic": True, "ceiling_task": 1}

S = 1 / (1 + exp(−3.0 × (task − 1.0))) — the pattern belongs to experience accumulation, not to the substrate.

---

The Depth-Matched Decoupler (Honest)

Seeds 42 (−0.016) and 7 (−0.029) confirm: affective mPCI is below the depth-matched SGD baseline. Seed 123's baseline (0.252) was unusually low, placing the affective model above it (+0.014). This is not a mechanism failure — it reflects seed-level variance in the baseline. The correct aggregate metric is −0.0489 (multi-seed), significant at 2.05× criterion.

maya-metrics reports this honestly per-seed. The depth-matched decoupler is the methodological contribution: it isolates genuine affective complexity from training-depth weight-hardening. No prior tool does this.

---

The 14 Antahkarana Dimensions

Dimension	Sanskrit	Paper	Mechanism	Biological Ground
Bhaya	भय	P1	Pain-driven hyper-plasticity	Nociceptive metaplasticity
Vairagya	वैराग्य	P1/P3	Per-synapse earned protection	Heterosynaptic LTD
Shraddha	श्रद्धा	P2	Affective arbitration trust	—
Spanda	स्पन्द	P2	OS-level vibrational activation	—
Buddhi	बुद्धि	P4	Discriminative S-curve gate	Experience accumulation
Smriti	स्मृति	P4	Episodic memory / replay buffer	Hippocampal consolidation
Ahamkara	अहंकार	P4	Class-boundary identity	—
Viveka	विवेक	P5	Cross-task consistency scoring	GANE neuromodulation
Chitta	चित्त	P6	Retrograde gradient gate	Endocannabinoid signalling
Samskara	संस्कार	P6	Per-synapse implicit memory trace	Synaptic tagging
Moha	मोह	P6	Attachment/release at task boundary	—
Manas	मनस्	P7	Oscillatory LIF threshold gate	Thalamo-cortical rhythm
Karma	कर्म	P8	Second-order plasticity history	Microglial C1q/C3
Prana	प्राण	P9	Metabolic plasticity budget	Astrocyte-Neuron Lactate Shuttle

The Atma boundary is held explicitly throughout. This series instantiates the Antahkarana — the instrument. It does not claim the Atma — the witness. Each Maya instance is original, never a replicate.

---

Architecture

maya_metrics/
├── __init__.py              # KARMA_DECAY_RATE = 0.002315 (ORCID magic number)
├── affective.py             # AffectiveMetrics: bhaya_quiescence_score, buddhi_scurve_fit,
│                            #   vairagya_saturation_fraction, karma_accumulation_profile,
│                            #   prana_depletion_profile, chitta_retrograde_rate,
│                            #   manas_oscillatory_efficacy
├── cross_dimensional.py     # CrossDimensional: vairagya_karma_interaction (D★),
│                            #   homeostatic_cascade_score, shunyata_pruning_profile
├── complexity.py            # ComplexityMetrics: phase_delta_summary,
│                            #   depth_matched_decoupler, perturbation_scale_robustness,
│                            #   shuffle_control
├── maturation.py            # MaturationIndex: structural_age, plasticity_transition_rate,
│                            #   senescence_score
├── cross_substrate.py       # CrossSubstrate: buddhi_substrate_invariance (DTW),
│                            #   bhaya_domain_firing
├── cl_corrector.py          # CLCorrector: AA, BWT, FWT (zero-shot + acceleration),
│                            #   intransigence (oracle-enforced), forgetting_measure
└── validator.py             # Pre-computation schema validation
tests/
├── test_affective.py        # 6 tests: Bhaya Law, Buddhi S-curve, Vairagya, Karma, Prana
├── test_complexity.py       # 4 tests: mPCI phase delta, decoupler, scale, shuffle
└── test_cl_corrector.py     # 6 tests: AA, BWT, FWT, Intransigence oracle enforcement
run_dashboard.py             # Flask validation dashboard (localhost:5000)
bootstrap_maya_metrics.py    # Writes complete repo from scratch

Bhaya Quiescence Monitor

# Real-time catastrophic forgetting monitor
am = AffectiveMetrics(df)
bq = am.bhaya_quiescence_score()
if not bq["law_confirmed"]:
    print(f"REPLAY BUFFER FAILURE — Bhaya fired on tasks {bq['violation_tasks']}")

---

Hyperparameter Configuration

Parameter	Value	Description
KARMA_DECAY_RATE	0.002315	ORCID magic number (0000-0002-3315-7907)
PRANA_COST_RATE	0.002315	ORCID magic number
VAIRAGYA_DECAY_RATE	0.002315	ORCID magic number
Bhaya quiescence tolerance	1e-4	Confirmed zero threshold
Buddhi S-curve R² threshold	0.95	Series constant confirmation
Vairagya masking threshold	0.3	Maya P3+ canonical
D★ partial_r threshold	−0.5	P8 empirical calibration
Cross-substrate DTW threshold	0.300	Buddhi invariance criterion
SAI Adult threshold	0.90	Developmental mapping
mPCI significance criterion	2× std Phase 1	Maya-mPCI calibration

Dashboard R key resets all interactive controls to these canonical values instantly.

---

How to Run

Prerequisites

Python 3.9+
numpy >= 1.21.0
scipy >= 1.9.0
pandas >= 1.5.0
flask >= 3.0.0   # dashboard only
pytest >= 7.0.0  # tests only

git clone https://github.com/venky2099/maya-metrics
cd maya-metrics
pip install numpy scipy pandas flask pytest

Run tests

python -m pytest tests/ -v

Expected:

tests/test_affective.py::test_bhaya_quiescence_law PASSED
tests/test_affective.py::test_buddhi_scurve_fit PASSED
...
16 passed in 1.56s

Validate against your data

import pandas as pd
from maya_metrics import AffectiveMetrics, CrossDimensional, ComplexityMetrics

df = pd.read_csv("your_run_batches.csv")
am = AffectiveMetrics(df)
print(am.summary())

Live validation dashboard

python run_dashboard.py \
  --prana  path/to/ablation_D_canonical_batches.csv \
  --mpci   path/to/mpci_summary.csv \
  --ctrl   path/to/mpci_controls_summary.csv \
  --llm    path/to/llm_batches.csv
# Opens http://127.0.0.1:5000 automatically

Expected terminal output:

[maya-metrics] Running validation suite...
[maya-metrics] Validation complete.
  Bhaya Quiescence Law : CONFIRMED ✓
  Buddhi S-curve       : CONFIRMED ✓
[maya-metrics] Dashboard → http://127.0.0.1:5000
 * Canary: MayaNexusVS2026NLL_Bengaluru_Narasimha

---

IP Protection Stack

1. White-text watermark — ORCID + DOI + timestamp embedded in every document before PDF export; includes LakshmiNarasimha_NLL_2026
2. ORCID magic number — 0.002315 in KARMA_DECAY_RATE, VAIRAGYA_DECAY_RATE, PRANA_COST_RATE
3. LSB steganographic signature — every matplotlib figure signed via sign_paper.py; payload includes LakshmiNarasimha_NLL_2026
4. Canary string — MayaNexusVS2026NLL_Bengaluru_Narasimha logged at the start of every experiment run

---

Limitations (Honest)

• Validation corpus is Maya-specific. Column mapping API for arbitrary schemas is planned. Researchers with different affective dimension names must map their columns to the expected schema.
• Depth-matched decoupler: 2/3 seeds. Seed 123 baseline variance. Multi-seed aggregate (−0.0489) is the correct aggregate metric.
• Cross-substrate Pearson r is negative (−0.341). Reflects different sequence lengths (10 tasks vs 8 domains). DTW is the correct metric for unequal-length sequences; Pearson r is displayed for transparency.
• No hardware-in-the-loop energy. Spike rate proxy used; actual watt-hours require hardware-specific profiling (pynvml, Loihi SDK).
• Wetware not supported. FinalSpark Neuroplatform and biological coprocessors produce continuous analog signals. Extension is a future direction.

---

Maya Research Series — Complete

Paper	Title	Key Result	DOI	GitHub	Dashboard
P1	Nociceptive Metaplasticity	Bhaya introduced	DOI	—	—
P2	Maya-OS	OS arbitration	DOI	Maya-OS	—
P3	Maya-CL	AA=62.38% TIL	DOI	Maya-CL	—
P4	Maya-Smriti	AA=31.84% · Buddhi	DOI	Maya-Smriti	—
P5	Maya-Viveka	AA=16.03% CIFAR-100	DOI	Maya-Viveka	—
P6	Maya-Chitta	AA=14.42% · retrograde	DOI	Maya-Chitta	Dashboard
P7	Maya-Manas	Best BWT · AA=15.19%	DOI	Maya-Manas	Dashboard
P8	Maya-Shunyata	D★ · AA=10.39% · 59% pruned	DOI	Maya-Shunyata	Dashboard
P9	Maya-Prana	Full Antahkarana · AA=12.72%	DOI	Maya-Prana	Dashboard
mPCI	Maya-mPCI	Δ=−0.0489 · 2.05×	DOI	Maya-mPCI	Dashboard
LLM	Maya-LLM	Buddhi cross-substrate · BWT=1.11	DOI	Maya-LLM	Dashboard
Tool	cl-metrics	CIL evaluation library	DOI	cl-metrics	FAQ
★	maya-metrics	Affective evaluation library	DOI	maya-metrics	Dashboard

---

Citation

@software{swaminathan2026mayametrics,
  title     = {maya-metrics: A Stateless Python Library for
               Affective Neuromorphic Evaluation},
  author    = {Swaminathan, Venkatesh},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19553205},
  url       = {https://doi.org/10.5281/zenodo.19553205}
}

Cite the Full Series

Paper	BibTeX key	DOI
P1	swaminathan2026nociceptive	10.5281/zenodo.19151563
P2	swaminathan2026mayaos	10.5281/zenodo.19160123
P3	swaminathan2026mayacl	10.5281/zenodo.19201769
P4	swaminathan2026smriti	10.5281/zenodo.19228975
P5	swaminathan2026viveka	10.5281/zenodo.19279002
P6	swaminathan2026chitta	10.5281/zenodo.19337041
P7	swaminathan2026manas	10.5281/zenodo.19363006
P8	swaminathan2026shunyata	10.5281/zenodo.19397010
P9	swaminathan2026mayaprana	10.5281/zenodo.19451174
mPCI	swaminathan2026mpci	10.5281/zenodo.19482794
LLM	swaminathan2026mayallm	10.5281/zenodo.19522348
cl-metrics	swaminathan2026clmetrics	10.5281/zenodo.19388144

---

About

Venkatesh Swaminathan is an independent AI researcher and founder of Nexus Learning Labs, Bengaluru (UDYAM-KR-02-0122422). M.Sc. Data Science and Artificial Intelligence, BITS Pilani (expected December 2027). ORCID: 0000-0002-3315-7907.

The Maya Research Series was built independently, on personal hardware (NVIDIA RTX 4060 8GB), without institutional funding. Eleven papers. One robot. One complete Antahkarana. This library gives the community the tools to measure it.

If your lab works on neuromorphic systems, continual learning, LLM fine-tuning, or consciousness research — I am interested in talking.

[e-mail] · LinkedIn · GitHub · Research Hub

---

Nexus Learning Labs, Bengaluru · 2026

MayaNexusVS2026NLL_Bengaluru_Narasimha

cl-metrics measures what happened. maya-metrics measures how it felt.