Antaḥkaraṇa-Net — Scale-Up Benchmark & Performance Report

A capability benchmark of the Vedic-mind continual-learning architecture at 36.5M–52.6M parameters. All numbers are from live inference on the trained checkpoints (the model is loaded from disk and re-evaluated), not training-time logs. 11 models, 7 capabilities each, run on a single NVIDIA A10.

Scope, stated up front. This is a research proof-of-concept: WideResNets on Split-CIFAR-100 and Split-Tiny-ImageNet. The contribution is the architecture — one agent that learns continually without forgetting, abstains instead of hallucinating, and self-regulates — demonstrated to scale from the original small-CNN POC to real models. It is not, and is not meant to be, a frontier model.

1. Setup


Architectures	WRN-28-10 (36.5M params), WRN-28-12 (52.6M params) — shared trunk φ + one linear head per task
Control ring	chitta (saṃskāra EWC+decay), guṇa (forgetting-aware MetaGunaController), āśrama, tapas, pramāṇa, turīya
Datasets	Split-CIFAR-100 (10×10 and 20×5 class-incremental); Split-Tiny-ImageNet (200-class, 10×20, 64→32)
OOD set (pramāṇa)	SVHN (test)
Training	from scratch, 25 epochs/task, Adam; wake→dream(replay)→sleep(consolidate) cycle
Hardware	OCI `BM.GPU.A10.4` → single A10 (23 GB), us-sanjose-1
Protocol	per-checkpoint live inference: per-task accuracy, pramāṇa gating, temperature calibration, mind-state trace
Seeds	core ×5, scale ×2, tiny ×2, āśrama ×2 (11 models total)

2. Capability scorecard (the 7 checks + PASS criteria)

#	Capability (faculty)	PASS criterion
1	Learns the task (buddhi+backbone)	avg accuracy > 3× chance
2	Continual memory (chitta/saṃskāra)	forgetting ≤ 0.05 and all tasks > 1.5× chance
3	Anti-hallucination (pramāṇa)	gated acc ≥ raw acc and OOD blind-accept < 30%
4	Calibration	ECE decreases after temperature scaling
5	Self-regulation (guṇa)	plasticity headroom decreases over life
6	Lifelong (āśrama)	progresses through life-stages to vānaprastha/saṃnyāsa
7	Identity monitor (turīya)	witness drift bounded (< 1.0)

Result: 10 of 11 models pass 7/7; 1 passes 6/7 (āśrama_s0 — see §6, a threshold artifact).

3. Headline results — group means

Configuration	Params	Avg acc	Naive forget	Agent forget	Reduction	Gated acc	OOD accept
core — CIFAR-100, 10 tasks	36.5M	0.599	0.589	0.0146	41.0×	0.929	8.4%
scale — CIFAR-100, 10 tasks	52.6M	0.615	0.551	0.0204	27.3×	0.933	10.0%
tiny — Tiny-ImageNet, 10 tasks	36.5M	0.416	0.503	0.0173	29.1×	0.956	0.3%
āśrama — CIFAR-100, 20 tasks	36.5M	0.606	0.595	0.0474	12.5×	0.960	1.3%

Read: catastrophic forgetting is cut 12–41× across two model sizes, two datasets, and stream lengths of 10 and 20 tasks. The agent both remembers (forgetting → ~0.01–0.05) and learns better (accuracy roughly doubles vs the naive baseline, which collapses to its last task).

4. Per-model detail (live inference)

Model	Data	Avg acc	Naive→Agent forget	Reduction	Gated@cov	OOD accept	ECE before→after	Score
core_s0	CIFAR-100	0.616	0.565→0.018	31.8×	0.943@0.21	16.1%	0.172→0.023	7/7
core_s1	CIFAR-100	0.592	0.621→0.014	45.1×	0.950	6.4%	0.172→0.043	7/7
core_s2	CIFAR-100	0.617	0.583→0.012	49.9×	0.918	2.6%	0.118→0.025	7/7
core_s3	CIFAR-100	0.587	0.553→0.015	37.1×	0.927	8.2%	0.150→0.024	7/7
core_s4	CIFAR-100	0.585	0.621→0.015	41.1×	0.907	8.4%	0.163→0.035	7/7
scale_s0	CIFAR-100	0.620	0.542→0.021	25.4×	0.921	8.7%	0.229→0.027	7/7
scale_s1	CIFAR-100	0.610	0.570→0.019	29.3×	0.944	11.2%	0.195→0.032	7/7
tin_s0	Tiny-ImageNet	0.411	0.503→0.017	29.2×	1.000	0.4%	0.226→0.025	7/7
tin_s1	Tiny-ImageNet	0.421	0.502→0.017	29.0×	0.912	0.2%	0.196→0.034	7/7
ashrama_s0	CIFAR-100 (20t)	0.592	0.603→0.049	12.3×	0.951	1.7%	0.225→0.046	6/7
ashrama_s1	CIFAR-100 (20t)	0.621	0.586→0.046	12.8×	0.969	1.0%	0.246→0.057	7/7

5. Analysis

Scaling (36.5M → 52.6M). Increasing capacity keeps the effect and nudges accuracy up (0.599 → 0.615). The forgetting-reduction ratio is a touch lower at 52.6M (27× vs 41×), but absolute agent-forgetting stays tiny (0.020 vs 0.015) — i.e. both models forget almost nothing; the ratio mostly reflects naive-baseline variance. The architecture is genuinely backbone-agnostic across sizes.

Difficulty — dataset. Tiny-ImageNet (200 fine-grained classes at 32 px) is much harder, so raw accuracy drops to ~0.42. Crucially, forgetting reduction holds (29×) and the model becomes more conservative: it abstains on 99.7% of OOD inputs and clears only its most-confident in-distribution predictions — exactly the right behavior when uncertain.

Difficulty — stream length. Doubling the stream (10 → 20 tasks, āśrama) is the hardest continual setting: twice as many tasks to retain with more interference. Reduction drops to 12.5× (agent-forgetting 0.047) — lower than the 10-task runs, and honestly so — yet all 20/20 tasks remain above chance and the model still abstains/calibrates cleanly. A >12× cut over a 20-task lifelong run is a strong result.

Per-task accuracy after the full stream — every task stays well above chance; the agent does not collapse to its last task (which is what the naive baseline does).

Anti-hallucination (pramāṇa). Calibrated abstention works everywhere: gated accuracy 0.91–1.00 (vs raw ~0.42–0.67), and OOD blind-accept 0.2%–16%. The harder the data, the lower the model's confidence and the more it abstains — a desirable, safety-aligned property. Coverage is correspondingly lower on hard sets (a deliberate trade: it commits less, but is right when it does).

Calibration. Temperature scaling reduces ECE everywhere (e.g. 0.172 → 0.023), so reported confidence can be trusted — the precondition for the abstention gate to be meaningful.

Interpretability (mind-state). Every model exposes a legible trajectory: plasticity headroom falls over life (e.g. 0.92 → 0.37; 0.77 → 0.14 on the 20-task run), the agent ages through all four āśramas (brahmacarya → gṛhastha → vānaprastha → saṃnyāsa), and turīya drift stays bounded (~0.50–0.59). You can read the agent's regulation as it lives — not a black box.

6. The one 6/7, explained (no goalpost-moving)

āśrama_s0 failed only capability 1 (learns the task): avg accuracy 0.592 vs the 0.60 bar (= 3× chance for 5-class tasks). That's 2.96× chance — a hair under an arbitrary threshold. Its continual_memory check passed with 20/20 tasks retained and 12.3× forgetting reduction, and āśrama_s1 (same config) scored 0.621 → 7/7. This is benchmark-threshold variance on a borderline metric, not a model deficiency. We report it as-is rather than relax the criterion.

7. Honest caveats

POC scale. Small/mid models, 32 px imagery. Demonstrates the architecture, not SOTA accuracy.
Forgetting reduction ratios are noisy across seeds (driven partly by naive-baseline variance); the trustworthy claim is absolute agent-forgetting ≈ 0.01–0.05 everywhere.
Low coverage on hard datasets is the flip side of strong abstention — fewer commitments, higher precision.
OOD uses SVHN with CIFAR-100 normalization for all models (a single fixed OOD probe); a per-dataset OOD suite would sharpen the tiny-imagenet numbers.
Vedic↔ML mappings are engineering analogies, clearly flagged. Nothing here is conscious.

8. Reproducibility

# train (per phase, resumable, idempotent):
bash scale/phases.sh                 # core → scale → pramana → tinyimagenet → ashrama
# benchmark any checkpoint (live inference, 7-capability report + JSON):
CUDA_VISIBLE_DEVICES=0 python3 scale/capability_bench.py --ckpt runs/<tag>.pt

Artifacts: 11 checkpoints (runs/*.pt), 11 training results (runs/*.json), 11 capability reports (runs/*_capbench.json), and this report. The uploadable model is antahkarana-36.5M-cifar100-wrn28-10.pt (36.5M, with Ω/θ* consolidation state) + load_akn.py + MODEL_CARD.md in scale/hf_package/.

Generated from live-inference benchmarks on trained checkpoints. Code: MIT. Built on the Upaniṣads, Sāṃkhya, Yoga, and PyTorch · snnTorch.